diagnostics

Review
Artificial Intelligence in Neurosurgery: A State-of-the-Art
Review from Past to Future
Jonathan A. Tangsrivimol 1,2 , Ethan Schonfeld 3 , Michael Zhang 4 , Anand Veeravagu 5 , Timothy R. Smith 6 ,
Roger Härtl 7 , Michael T. Lawton 8 , Adham H. El-Sherbini 9 , Daniel M. Prevedello 2 ,
Benjamin S. Glicksberg 10 and Chayakrit Krittanawong 11, *

1 Division of Neurosurgery, Department of Surgery, Chulabhorn Hospital, Chulabhorn Royal Academy,

Bangkok 10210, Thailand; j.tangsrivimol@gmail.com
2 Department of Neurological Surgery, The Ohio State University Wexner Medical Center and Jame Cancer
Institute, Columbus, OH 43210, USA
3 Department Biomedical Informatics, Stanford University School of Medicine, Palo Alto, CA 94305, USA
4 Department of Neurosurgery, Stanford University School of Medicine, Palo Alto, CA 94305, USA
5 Stanford Neurosurgical Artificial Intelligence and Machine Learning Laboratory, Department of
Neurosurgery, Stanford University School of Medicine, Stanford, CA 94305, USA
6 Department of Neurosurgery, Computational Neuroscience Outcomes Center (CNOC), Mass General
Brigham, Harvard Medical School, Boston, MA 02115, USA
7 Weill Cornell Medicine Brain and Spine Center, New York, NY 10022, USA
8 Department of Neurosurgery, Barrow Neurological Institute (BNI), Phoenix, AZ 85013, USA
9 Faculty of Health Sciences, Queen’s University, Kingston, ON K7L 3N6, Canada
10 Hasso Plattner Institute for Digital Health, Icahn School of Medicine at Mount Sinai,
New York, NY 10029, USA
11 Cardiology Division, New York University Langone Health, New York University School of Medicine,
New York, NY 10016, USA
* Correspondence: chayakrit.krittanawong@va.gov

Abstract: In recent years, there has been a significant surge in discussions surrounding artificial
intelligence (AI), along with a corresponding increase in its practical applications in various facets
Citation: Tangsrivimol, J.A.; of everyday life, including the medical industry. Notably, even in the highly specialized realm of
Schonfeld, E.; Zhang, M.; Veeravagu, neurosurgery, AI has been utilized for differential diagnosis, pre-operative evaluation, and improving
A.; Smith, T.R.; Härtl, R.; Lawton, surgical precision. Many of these applications have begun to mitigate risks of intraoperative and
M.T.; El-Sherbini, A.H.; Prevedello, postoperative complications and post-operative care. This article aims to present an overview of the
D.M.; Glicksberg, B.S.; et al. Artificial
principal published papers on the significant themes of tumor, spine, epilepsy, and vascular issues,
Intelligence in Neurosurgery: A
wherein AI has been applied to assess its potential applications within neurosurgery. The method
State-of-the-Art Review from Past to
involved identifying high-cited seminal papers using PubMed and Google Scholar, conducting a
Future. Diagnostics 2023, 13, 2429.
comprehensive review of various study types, and summarizing machine learning applications to
https://doi.org/10.3390/
diagnostics13142429 enhance understanding among clinicians for future utilization. Recent studies demonstrate that
machine learning (ML) holds significant potential in neuro-oncological care, spine surgery, epilepsy
Academic Editor: Andreas Kjaer
management, and other neurosurgical applications. ML techniques have proven effective in tumor
Received: 31 May 2023 identification, surgical outcomes prediction, seizure outcome prediction, aneurysm prediction, and
Revised: 6 July 2023 more, highlighting its broad impact and potential in improving patient management and outcomes
Accepted: 10 July 2023 in neurosurgery. This review will encompass the current state of research, as well as predictions for
Published: 20 July 2023 the future of AI within neurosurgery.

Keywords: artificial intelligence (AI); machine learning (ML); deep learning (DL); artificial Neural
Networks (ANN); Convolutional Neural Networks (CNN); Recurrent Neural Networks (RNN);
Copyright: © 2023 by the authors.
neurosurgery
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
1. Introduction
creativecommons.org/licenses/by/ In the past half-decade, there has been a considerable amount of discourse surround-
4.0/). ing the subject of Artificial Intelligence (AI). AI encompasses the utilization of computer

Diagnostics 2023, 13, 2429. https://doi.org/10.3390/diagnostics13142429 https://www.mdpi.com/journal/diagnostics

 capabilities [1,2]. Upon conducting a comprehensive review of previously published
articles, it is evident that the utilization of AI is progressively becoming more widespread
in various biomedical fields. Numerous systematic reviews have been published [3–6],
with the majority attempting to consolidate a vast number of research papers. However,
Diagnostics 2023, 13, 2429 2 of 33
a significant gap exists in the integration of AI with clinical expertise, which is a crucial
aspect that needs to be addressed. While some clinicians can comprehend, create, and
apply AI techniques, systems a tosignificant proportion,
accomplish objectives throughincluding neurosurgeons,
the simulation need more
of cognitive capabilities [1,2].
familiarity with thisUpon domain.
conducting Most clinicians review
a comprehensive have limited exposure
of previously published to articles,
AI due it istoevident
their
primary concentrationthat theonutilization
surgical of practice
AI is progressively
and thebecoming more widespread
era of their in variousthis
training. Hence, biomedical
paper
fields. Numerous systematic reviews have been published [3–6], with the majority attempt-
aims to identify high-cited seminal papers, as higher citation rates indicate clinician
ing to consolidate a vast number of research papers. However, a significant gap exists in the
engagement, andintegration
to summarizeof AI with and
clinicalhighlight the isvarious
expertise, which applications
a crucial aspect that needs toofbemachine
addressed.
learning. Such an While
endeavor aims to can
some clinicians enhance understanding
comprehend, among
create, and apply clinicians,
AI techniques, paving the
a significant pro-
portion, including neurosurgeons, need more familiarity
way for future utilization as familiarity grows over time. The literature search method with this domain. Most clinicians
have limited exposure to AI due to their primary concentration on surgical practice and
involved searching PubMed and Google Scholar using relevant keywords to identify
the era of their training. Hence, this paper aims to identify high-cited seminal papers, as
English literature publications fromindicate
higher citation rates its inception
clinician until May 2023.
engagement, and toAsummarize
comprehensive
and highlightreview the
was performed on various studies, including observational studies, case–control studies,
various applications of machine learning. Such an endeavor aims to enhance understanding
among clinicians, paving the way for future utilization as familiarity grows over time. The
cohort studies, clinical trials, meta-analyses, reviews, and guidelines, as depicted in Figure
literature search method involved searching PubMed and Google Scholar using relevant
1. For ease of navigation
keywordsand clarity,
to identify the information
English will befrom
literature publications presented in four
its inception untilcategories:
May 2023. A
tumor, spine, functional,
comprehensive and review
vascular. This wide
was performed rangestudies,
on various of topics willobservational
including be examined to
studies,
case–control studies, cohort studies, clinical trials,
obtain a comprehensive understanding of the advancements and challenges associated meta-analyses, reviews, and guidelines,
as depicted in Figure 1. For ease of navigation and clarity, the information will be presented
with integrating AI in these domains.
in four categories: tumor, spine, functional, and vascular. This wide range of topics will be
examined to obtain a comprehensive understanding of the advancements and challenges
associated with integrating AI in these domains.

Figure 1. Literature Search Method. Figure Description: PubMed and Google Scholar were searched
using AI-related keywords for English literature published from inception to May 2023. Observational
studies, case–control studies, cohort studies, clinical trials, meta-analyses, reviews, and guidelines
were reviewed.
Figure 1. Literature Search Method. Figure Description: PubMed and Google Scholar were searched
using AI-related keywords fora English
To obtain literature
comprehensive published
understanding from
of the inception
integration of AItoin May 2023.
medicine, it
Observational studies, case–control studies, cohort studies, clinical trials, meta-analyses, reviews,
is essential to first differentiate between AI, machine learning (ML), and deep learning
and guidelines were(DL).
reviewed.
While the idea of AI has existed for a considerable length of time, ML is a subfield
of AI which seeks to learn patterns from data [7–9] that can be broadly divided into two
categories: supervised learning and unsupervised learning [10]. However, a third category
To obtain a comprehensive understanding of the integration of AI in medicine, it is
of weakly supervised learning and/or reinforcement learning may be considered given
essential to first differentiate
its importancebetween
in modernAI, machine
real-world MLlearning
(e.g., Chat(ML), and deep learning
GPT). Supervised learning (DL).
involves
While the idea of creating
AI has predictions
existed forbased
a considerable
on preliminarylength
data orof time,
data MLwith
groups is alabeled
subfield of AI
outcomes,
whereas unsupervised learning does not learn from or have
which seeks to learn patterns from data [7–9] that can be broadly divided into twoaccess to labeled outcomes [10].
While both types of learning can be employed to create quantitative predictions, unsuper-
categories: supervised learning
vised learning canand unsupervised
uncover learning.[10]
new classification However,
or patterns. a third
Here, we review allcategory
emerging
of weakly supervised learning
applications and/or
of these reinforcement
machine learning
learning technologies in may be considered
neurosurgery (Figure 2).given its
importance in modern real-world ML (e.g., Chat GPT). Supervised learning involves
 creating predictions based on preliminary data or data groups with labeled outcomes,
whereas unsupervised learning does not learn from or have access to labeled outcomes
[10]. While both types of learning can be employed to create quantitative predictions,
unsupervised learning can uncover new classification or patterns. Here, we review all
Diagnostics 2023, 13, 2429 3 of2).
emerging applications of these machine learning technologies in neurosurgery (Figure 33

Figure 2. Summarizes
Summarizes potentials of AI in neurosurgery.

Machine learning
learningforms
formsthethe foundation
foundation of DL
of [11,12], whichwhich
DL [11,12], employs Artificial
employs Neural
Artificial
Networks
Neural (ANN) designed
Networks to mimicto
(ANN) designed cognitive brain function
mimic cognitive braintofunction
learn complex
to learnpatterns
complex in
data. Computer vision tasks often utilize Convolutional Neural Networks
patterns in data. Computer vision tasks often utilize Convolutional Neural Networks (CNN) which can
learn and
(CNN) identify
which can visual
learn patterns [13] (Figure
and identify 3). Natural
visual patterns Language
[13] (Figure 3). Processing
Natural Tasks have
Language
historically used Recurrent Neural Networks (RNN), which can encode
Processing Tasks have historically used Recurrent Neural Networks (RNN), which can time/sequence-
based information,
encode such as in language,
time/sequence-based information,but is nowas
such primarily transformer-based
in language, but is now (e.g., Chat
primarily
GPT, pretrained medical
transformer-based (e.g., large
Chat language
GPT, models [GatorTron],
pretrained medical etc. . . . )language
large [14,15]. Clinicians
models
Diagnostics 2023, 13, x FOR PEER REVIEW 4 of 40
must educateetc.…)
[GatorTron], themselves
[14,15].on the various
Clinicians mustforms of AI,
educate as their understanding
themselves on the various forms of these of
technologies
AI, as their (and not blind trust)
understanding is essential
of these in ensuring
technologies (andits proper
not blindand safe is
trust) translation
essential forin
patient care.
ensuring its proper and safe translation for patient care.

Figure 3. Convolutional neural network.

Figure 3. Convolutional neural network.
2. Tumor
2. TumorAI has been applied for the brain tumor classification task [16–21]. In a recent
AI has
Buchlak etbeen applied for theAI
al. investigation, brain
wastumor classification
leveraged task [16–21].
for identifying and In a recent Buchlak
categorizing glioma
ettumors
al. investigation, AI was leveraged
through neuroimaging for [22].
analysis identifying and categorizing
The researchers glioma evaluated
systematically tumors
through neuroimaging analysis [22]. The researchers systematically evaluated 153 studies
that employed ML [CNN, Support Vector Machine (SVM),Random Forest (RF)] to
enhance tumor grading, diagnosis, segmentation, non-invasive genetic biomarker
identification, progression monitoring, and patient survival prognosis. In general, the
Diagnostics 2023, 13, 2429 4 of 33

153 studies that employed ML [CNN, Support Vector Machine (SVM),Random Forest
(RF)] to enhance tumor grading, diagnosis, segmentation, non-invasive genetic biomarker
identification, progression monitoring, and patient survival prognosis. In general, the
performance of the ML model was excellent (AUC = 0.87 ± 0.09; sensitivity = 0.87± 0.10;
specificity = 0.86 ± 0.10; precision = 0.88 ± 0.11). Their findings revealed that CNN, SVM,
RF demonstrated the most favorable outcomes. This investigation underscores ML’s critical
role in medical classification and its potential to significantly enhance disease diagnosis
and treatment. This highlights the limitations of a review on machine learning applications
to glioma MRI data, including the influence of a large sample size on NLP classification
models and the exclusion of conference papers, while suggesting the use of optimized deep
language models and referring readers to specific papers for further information.
Distinguishing between primary CNS lymphoma (PCNSL) and glioblastoma multi-
forme (GBM) based on MRI findings can be challenging. McAvoy et al. employed the
EfficientNetB4 architecture within a convolutional neural network (CNN) framework to
analyze contrast-enhanced T1-weighted images from a cohort of 320 patients with sus-
pected GBM or PCNSL [23]. The findings demonstrated that CNN-based analysis could
effectively assist radiologists in achieving accurate differential diagnoses between these two
entities (mean 5-fold cross-validation AUC = 0.71). This research highlights the potential
of CNNs as a valuable tool in aiding the diagnostic process and improving the precision
of differential diagnosis for PCNSL and GBM based on MRI imaging. This paper aims
to assist physicians in formulating a comprehensive and accurate differential diagnosis,
ultimately leading to faster and more appropriate treatment measures. This study has
limitations, including its retrospective design with a small number of patients from two
academic institutions, which may limit the generalizability of the findings to other settings,
the use of PNG exports of DICOM images leading to loss of data, and the absence of a
direct comparison between the classification outcomes of CNNs and radiologists, thus
requiring further research to determine the tool’s clinical value.
Accurate cortical segmentation and volume assessment play a vital role in the con-
tinuous surgical planning of patients and monitoring of treatment response. To address
this task, Boaro et al. employed a 3D convolutional neural network (3D-CNN) to achieve
expert-level automated segmentation and volume estimation of meningiomas from MRI
scans [24]. An initial training phase involved training a 3D-CNN by segmenting complete
brain volumes utilizing a dataset comprising 10,099 MRIs of healthy brains. Subsequently,
through the implementation of transfer learning, the network underwent specific train-
ing for meningioma segmentation, using a dataset consisting of 806 labeled MRIs. Their
approach yielded an impressive accuracy of 88.2%. This highlights the potential of their
method for precise and reliable cortical segmentation, enabling enhanced patient care
and treatment evaluation. The limitations of this study include the inability to evaluate
post-operative residuals, tumor recurrence, or tumor growth, due to the inclusion of single
pre-operative scans, the lack of testing the model on brain MRI scans without meningioma
for assessing detection performance, the absence of integrating the algorithm into the
hospital informatics system, and the retrospective nature of the study requiring prospective
validation for real-world clinical applicability.
ML may also be used to identify and characterize predictive variables for treatment
prognosis. The predictive value of isocitrate dehydrogenase (IDH) mutation status and
1p19q codeletion, as indicators of treatment response in glioma, remains an area of interest.
Zhou et al. focused on the predictive value of isocitrate dehydrogenase (IDH) mutation
status and 1p19q codeletion as indicators of treatment response in glioma from MRI
patterns [25]. The training cohort consisted of preoperative MRIs from 538 glioma patients
spanning three different institutions. They utilized a random forest algorithm to construct
a predictive model that classified gliomas into three categories: IDH-wildtype, IDH-mutant
with 1p19q codeletion, and IDH-mutant without 1p19q codeletion. The study results
showed a successful prediction rate of 78.2%, accurately classifying 155 out of 198 cases.
The performance of IDH was evaluated by calculating the area under the receiver operating
Diagnostics 2023, 13, 2429 5 of 33

characteristic curve (AUC), which yielded a value of 0.921. Similarly, in the validation
cohort, IDH achieved an AUC of 0.919. Age provided the greatest predictive value, followed
by shape features. These findings highlight the potential of the developed model in
accurately determining the IDH mutation status and 1p19q codeletion status from MRI
patterns in glioma patients. The study’s limitations include its retrospective design with
a focus on known gliomas, limiting the model’s applicability to situations with different
tumor types and non-tumor mimickers, and the need for a more general model using
data from other lesion types to improve generalizability. Furthermore, the study did not
incorporate advanced MR modalities, such as perfusion MRI and MR-spectroscopy, which
could enhance the prediction of IDH genotype.
U-Net is an artificial neural network architecture primarily used for image segmenta-
tion in computer vision. Its U-shaped structure consists of an encoding path that captures
features through convolutional layers and downsampling, and a decoding path that re-
covers spatial resolution using upsampling and skip connections. These skip connections
fuse low-level and high-level features to capture fine details and context. The output is a
segmentation mask with class labels for each pixel. U-Net has proven highly effective in
medical imaging, but can also be applied to other segmentation tasks. The BraTS (Brain
Tumor Segmentation) dataset is a widely used benchmark dataset in medical imaging for
brain tumor segmentation. It contains multi-modal MRI scans of patients with annotated
tumor regions. The dataset serves as a standard for developing and evaluating algorithms
for automatic tumor segmentation, enabling researchers to compare their methods and
advance the field. BraTS has been instrumental in the development of accurate and efficient
brain tumor segmentation techniques. Huang et al. and Yousef et al. address the challenges
associated with the segmentation of brain tumors from MRI scans [26,27]. Huang et al.
propose a deep multi-task learning framework incorporating a multi-depth fusion module
and a distance transform decoder to achieve accurate segmentation [26]. On the other
hand, Yousef et al. highlight the prevalence of U-Net-based models in medical imaging [27].
They evaluate various variants of U-Net and emphasize the importance of developing new
architectures to optimize medical image analysis. The aforementioned paper proves highly
valuable in the context of brain tumor classification. However, more emphasis should
be placed on discussing the architectural aspects of the deep machine learning utilized.
Instead, this paper primarily focuses on clinical knowledge.
AI has also been implemented in surgical procedures to assist in the planning phase,
aiding in decision-making during surgery. Its utilization has shown promising results, en-
hancing the accuracy of surgical positioning and mitigating surgical complications [28–32].
In a recent investigation by Tonutti et al., ML was employed to facilitate the diagnosis of
intraoperative tumors [32]. Specifically, the researchers used ML algorithms, such as ANNs
and Support Vector Machine (SVM), to develop personalized anatomical models for intra-
operative use. Integrating augmented reality (AR) with these models further enhanced
surgical precision. The results revealed that SVR yielded more precise outcomes than ANN,
with positional errors of less than 0.2 mm. Furthermore, the model was observed to be
more accurate and personalized than real-time deformation models, thereby illustrating
its potential to revolutionize the field of surgery. The study highlights assumptions and
simplifications in the development of a biomechanical brain model for machine learning,
including the use of generic mechanical parameters, exclusion of certain brain structures,
and limitations in accounting for topological changes during surgery, suggesting the need
for more advanced simulations and real-time imaging to make the method applicable in
clinical settings.
Shen et al. have developed a pioneering approach for the intraoperative diagno-
sis of glioma utilizing deep CNNs and fluorescence imaging (FL-CNN) [33]. A total of
23 patients diagnosed with glioma participated in the study, wherein they underwent fluo-
rescence image-guided surgery after receiving injections of indocyanine green. Following
the surgical procedures conducted on these patients, 1874 tissue samples were carefully
collected. Additionally, fluorescence images in the second near-infrared window (NIR-II,
Diagnostics 2023, 13, 2429 6 of 33

1000–1700 nm) were acquired to provide detailed visual information. A FL-CNN was
utilized for automated glioma diagnosis and compared to the gold-standard pathology for
intraoperative diagnosis. The study revealed that FL-CNN exhibited superior sensitivity
(93.8% vs. 82.0%, p < 0.001) and specificity of over 80%, without any additional time,
outperforming neurosurgeons. Moreover, the FL-CNN effectively corrected nearly 70% of
neurosurgeon errors. Additionally, the FL-CNN could predict tumor grade and Ki67 with
AUCs of 0.81 and 0.625, respectively. These findings demonstrate that FL-CNNs are more
effective than neurosurgeons, making them suitable for intraoperative glioma diagnosis.
The limitation of the FL-CNN approach include its reliance on NIR-II fluorescence imaging,
which offers advantages over NIR-I but may still exhibit lower specificity compared to
clinically available 5-ALA fluorescence. However, the FL-CNN demonstrates comparable
specificity and higher sensitivity than 5-ALA when equipped with deep learning, and the
experiment results were confirmed by pathological examination using a gold standard,
ensuring precise performance measurement and objective comparison with neurosurgeons.
Hollon et al. conducted a similar investigation to assess the efficacy of Raman-based
imaging, coupled with CNNs, compared to board-certified neuropathologists for diag-
nosing glioma molecular class identification [34]. Inputs to the model included: Raman
spectroscopy derived imaging, coherent anti-Stokes Raman scattering (CARS) microscopy,
and stimulated Raman histology. By employing a boosted tree algorithm to classify intra-
operative Raman spectra, they could discern normal brain tissue from areas invaded by
tumors, specifically those with a tumor cell invasion exceeding 15%. This classification ap-
proach yielded a remarkable accuracy rate of 92% (with a sensitivity of 93% and specificity
of 91%). This imaging technique enables the analysis of specimens down to the molecular
level and has a sub-micron resolution, providing highly detailed information. This advance
can enable subsequent CNN to be applied beyond the details and patterns discernible to
the human eye.
AI has been developed to predict outcomes following brain tumor surgery [35–39].
Given the increasing lifespan of patients with brain metastasis and the rising incidence of
leptomeningeal disease (LMD), studies on LMD as a risk factor are limited. Tewarie et al.
examined leptomeningeal disease (LMD) as a risk factor in brain metastasis patients’ rising
lifespans [40]. They used the conditional survival forest, Cox proportional hazards model,
XGBoost classifier, extra trees classifier, logistic regression, and SMOTE to overcome class
imbalance. In 168 (15.9%) of 1054 brain metastasis patients who had surgery, LMD occurred
at a median time of 7.05 months following diagnosis. For the optimal Leptomeningeal
Disease (LMD) occurrence discrimination, utilizing an XGboost algorithm proved highly
effective, resulting in an impressive AUC of 0.83. Furthermore, when it came to prog-
nosticating the time until LMD development, the random forest algorithm and the Cox
proportional hazards model exhibited comparable performance, with a concordance index
(C-index) of 0.76. Notably, proximity of brain metastasis to the cerebrospinal fluid space and
the site of cerebellar brain metastasis were important factors in both LMD classification and
regression. In addition, lymph node metastasis of the primary tumor at the time of brain
metastasis diagnosis emerged as a significant risk factor influencing both the incidence of
LMD and the time to LMD. The limitations of the study include the wide time span during
which patients were included, the classification of radiographic elements based on clinical
relevance, the need for further research on the isolated role of radiographic components
in LMD, the novelty of lymph node metastasis as an LMD risk factor, the exclusion of
patients receiving only radiation therapy, the reduced variability of the data due to the use
of SMOTE, the theoretical nature of LMD prognostication at BM diagnosis in clinical care,
the need for external validation of the models, and the exploration of possible novel LMD
risk factors.
Currently, the prediction of brain metastasis (BM) often based on radiotherapy. [41–43].
Hulsbergen et al. sought to develop a predictive model for estimating 6-month survival
after surgical resection of brain metastasis [44]. The current reliance on radiotherapy-based
approaches for brain metastasis prediction prompted the need for an alternative approach.
Diagnostics 2023, 13, 2429 7 of 33

The study utilized an institutional database of 1062 patients and tested seven distinct ML
models, with model performance assessed by AUC. The results indicated that logistic
regression outperformed other methods, achieving an AUC of 0.71. In comparison, the
diagnosis-specific graded prognostic assessment achieved an AUC of 0.66. These findings
suggest that the developed model holds promise in accurately predicting 6-month survival
(p < 0.0005) following neurological resection of brain metastasis, facilitating meaningful risk
stratification in clinical practice. This study has limitations, including internal validation
using retrospective data, the need for external validation in a prospective setting, the
focus on survival at a 6-month cutoff rather than overall median survival, the influence
of intraoperative and postoperative factors on survival prediction, and the importance of
randomized trials for surgical decision making. The model can estimate risk and outcomes
for patients undergoing surgery but should not be used to determine whether surgery is
appropriate. Further analysis of predictive variables in the model can enable further efforts
to improve upon the achieved performance.
Given the significant hurdles encountered in accurately predicting individual patient
survival, particularly in glioma [45–48], Senders et al. aimed to address the difficulty
in predicting survival in glioblastoma multiforme (GBM) patients by comparing multi-
ple machine and statistical learning algorithms [49]. They developed an online survival
calculator using a training dataset of GBM patients diagnosed between 2005 and 2015.
Using 15 statistical and machine learning models (e.g., AFT, bagged decision trees, boosted
decision trees, boosted decision trees survival,) trained on demographic, socioeconomic,
clinical, and radiologic characteristics, they predicted one-year survival and generated
individualized survival curves. The study included 20,821 patients who met the criteria,
and the AFT model exhibited superior consistency with a concordance index of 0.70. These
findings emphasize the need for analyzing assessment in developing and utilizing sur-
vival strategies, highlighting the potential of advanced analytical approaches for improved
survival predictions in GBM patients. The limitations of many machine learning algo-
rithms include their restriction to continuous and binary models, the inability to compute
subject-level survival curves, lack of interpretability, computational inefficiency, and the
need for evaluating models based on multiple criteria rather than solely prediction perfor-
mance, as factors unrelated to prediction performance can exclude high-performing models
from clinical deployment. Furthermore, the predictive performance can vary depending
on the number and nature of input features, such as the inclusion of multimodal data
like radiogenomics.
The evaluation of treatment response for glioma frequently necessitates the utiliza-
tion of MRI imaging techniques such as MR perfusion and diffusion tensor imaging
(DTI) [43,50,51]. In an effort to enhance this process, Chang et al. aimed to improve
the evaluation of glioma treatment response using machine learning models with MRI
input [52]. They focused on analyzing the hyperintensity of fluid-attenuated inversion
recovery (FLAIR), the contrast-enhancing tumor region, and determining tumor volume
based on the Neuro-Oncology (RANO) response assessment criteria. Two distinct pa-
tient cohorts were employed. The first cohort comprised 843 preoperative MRIs from
843 patients diagnosed with low- or high-grade gliomas originating from four different
institutions. The second cohort encompassed 713 longitudinal postoperative MRI visits
from 54 patients newly diagnosed with glioblastomas. Each patient in the second cohort
had two “baseline” MRIs conducted before the initiation of treatment. It is important
to note that this second cohort was exclusively derived from a single institution. In the
cohort of postoperative GBM patients, the automatically generated FLAIR hyperintensity
volume, contrast-enhancing tumor volume, and AutoRANO were highly repeatable for
double-baseline visits, with ICCs of 0.986, 0.991, and 0.977, respectively. Preoperative
FLAIR hyperintensity, postoperative FLAIR, and postoperative contrast-enhancing tumor
volumes had ICC values of 0.915, 0.924, and 0.965, respectively. Finally, FLAIR hyperinten-
sity volume, contrast-enhancing tumor volume, and RANO measurements had ICCs of
0.917, 0.966, and 0.850 for comparing manually and automatically calculated longitudinal
Diagnostics 2023, 13, 2429 8 of 33

tumor burden changes. AutoRANO, showed promising results in accurately assessing

tumors following treatment. These trials will provide valuable insights into the reliability
and effectiveness of automated methods for assessing glioma treatment response. The
study has several limitations: reliance on segmentations from a single rater, a small patient
cohort from a single institution, lack of comparison with other approaches, exclusion of
smaller tumors, variability in MR imaging availability, absence of confidence assessment in
segmentations, and the need for further validation in larger cohorts and clinical trials.
Applications of AI to the management of neurosurgical care is quickly growing [53–56].
Senders et al. conducted a study using ML techniques such as ANN, SVM, Fuzzy c-Means
(FCM), Bayesian Learning (BL), RF, Logistic regression (LR), Linear Discriminant Analysis
(LDA) and others [53]. As of 1 January 2017, over 200 studies of 6,402 citations were
identified, involving presurgical planning, intraoperative guidance, neurophysiological
monitoring, and neurosurgical outcome prediction. Such works continue to underscore
the breadth to which ML can make an impact within neuro-oncological care. The study
highlights the limitations of the systematic review, including the need for more detailed
analysis of all studies, focus on perioperative care applications, and caution in interpreting
the quantitative performance summary. However, it emphasizes that these limitations are
proportional to the study’s strengths. It concludes that the review provides an overview of
the neurosurgical literature on machine learning and insights into its future direction in
the field.

Current Challenges
Several recent studies have highlighted the potential of machine learning (ML) models
in various aspects of neuro-oncological care. Buchlak et al. demonstrated the excellent
performance of ML, specifically CNN, SVM, and RF models, in identifying and cate-
gorizing glioma tumors through neuroimaging analysis [22]. McAvoy et al. utilized
CNN-based analysis to aid radiologists in accurate differential diagnoses between primary
CNS lymphoma (PCNSL) and glioblastoma multiforme (GBM) [23]. Boaro et al. achieved
high accuracy in cortical segmentation and volume estimation of meningiomas using a
3D-CNN [24]. Zhou et al. developed a predictive model for glioma classification based
on IDH mutation status and 1p19q codeletion using a random forest algorithm [25].
Tonutti et al. employed ML algorithms to develop personalized anatomical models for
intraoperative tumor diagnosis [32]. Shen et al. utilized deep CNNs and fluorescence imag-
ing for intraoperative glioma diagnosis, outperforming neurosurgeons [33]. Hollon et al.
achieved high-glioma molecular class identification accuracy using Raman-based imaging
coupled with CNNs [34]. Using various ML algorithms, Tewarie et al. examined the risk
factors for leptomeningeal disease (LMD) in brain metastasis patients [40]. Hulsbergen et al.
developed a predictive model for estimating 6-month survival after surgical resection of
brain metastasis [44]. Senders et al. developed an online survival calculator and conducted
a comprehensive review of ML techniques in neuro-oncological care, highlighting their
broad impact in different areas of patient management [49,53]. The limitations include
small sample sizes, retrospective designs, lack of external validation, and the need for fur-
ther research and validation in clinical settings. Despite these limitations, machine learning
shows promise in improving neurosurgical care, but practical and ethical considerations
need to be addressed during implementation.
The potential of AI to revolutionize tumor classification beyond glioblastoma (GBM)
and lymphoma is evident, with the ability to discover marker and prognostic genes across
various tumor types. These applications could significantly shape the trajectory of treatment.
Furthermore, integrating pre-surgical modeling into surgical planning promises to reduce
complications due to anatomical variations. AI has shown efficacy in monitoring tumor
recurrence in postoperative care, yet the patient’s observations remain crucial. The prospect
of leveraging ChatGPT to provide patients with knowledge and alleviate anxiety is feasible,
but adequate training and reliable references are vital prerequisites for its application.
Diagnostics 2023, 13, 2429 9 of 33

3. Spine
ML has brought about significant innovation in spine surgery, as evidenced by the
consistently high accuracy of outcome data. Most studies in this domain employ machine
learning techniques, including ANN, SVM, RF, and others [57–61]. Nida Fatima et al. con-
ducted an ML study to determine the 30-day adverse event rate in patients with Lumbar
Degenerative Spondylolisthesis (LDS) following surgery [62]. The study encompassed
a dataset of 80,610 patients who underwent LDS surgery, of whom 3965 (4.9%) experi-
enced adverse events within the 30-day postoperative period. ML models, specifically
logistic regression and LASSO, were employed to develop 26 prospective models. The
final ML algorithms identified several predictors, including gender, age, American So-
ciety of Anesthesiologist grade, autogenous iliac bone graft, instrument fusion, surgery
levels, surgical approach, functional status, preoperative serum albumin levels (g/dL), and
serum alkaline phosphatase levels (IU/mL). Logistic regression consistently demonstrated
superior performance in AUC compared to LASSO across various models. This study
highlights the potential of utilizing ML techniques in predicting outcomes when dealing
with large datasets. Such predictive capabilities can greatly assist in patient counseling
and surgical risk assessment. The study has several limitations, including the variation in
patient and surgical characteristics within the database used, limited postoperative out-
come data beyond 30 days, suboptimal performance of the prediction model with an AUC
below 0.8, potential missing variables, coding errors in the data, and the LASSO regression
not demonstrating an improved performance compared to logistic regression. Further
investigation is needed using alternative machine learning algorithms and larger datasets
to enhance predictive accuracy for postoperative adverse events after spinal surgery.
The same assessment can be performed for different outcome statistics. In Karhade et al.’s
study, the authors predicted 30-day mortality following spine metastasis surgery from a
cohort of 1790 patients [63]. Using various ML algorithms (e.g., Neural Network, Support
Vector Machine, Bayes Point Machine, and Decision Tree models), a Neural Network with
a c-statistic of 0.769 was the best model for predicting 30-day mortality. They successfully
identified preoperative prognostic markers, such as albumin, functional status, WBC, Hct,
alkaline phosphatase, spinal location, and concomitant systemic disease. The implication is
that these ML algorithms could, preoperatively, accurately predict postoperative outcomes.
The study has several limitations, including the variation in patient and surgical charac-
teristics within the database used, limited postoperative outcome data beyond 30 days,
suboptimal performance of the prediction model with an AUC below 0.8, potential missing
variables, coding errors in the data, and the LASSO regression not demonstrating improved
performance compared to logistic regression. Further investigation is needed using alterna-
tive machine learning algorithms and larger datasets to enhance predictive accuracy for
postoperative adverse events after spinal surgery.
Similarly, Ames et al. conducted an unsupervised AI study to identify surgical factors
that predict surgical outcomes in adult spinal deformity (ASD) [64]. The study involved
570 patients divided into three groups: young patients with coronal deformity (n = 195),
older patients with a history of spinal surgery (n = 157), and older patients without prior
surgeries (n = 218). Hierarchical clustering was employed as the primary methodology to
generate representative clusters of patients, characterized by high within-group similarity
and the greatest dissimilarity when compared to other groups. Patients were also cate-
gorized into 12 groups based on osteotomy type, instrumentation, and interbody fusion.
Ultimately, the factors identified allowed clinical reasoning of which patients would un-
dergo surgery with minimal risk. The study has limitations, including the dependency on
sample size and observation heterogeneity for determining patient and operative clusters,
the potential for further iterative refinements of the model and classification with future
data, and the need for additional research to test hypotheses and compare patient-reported
outcomes with objective measures.
To evaluate the risk of Adjacent Segment Disease (ASD), it is crucial to consider pa-
tients who have undergone previous Anterior Cervical Discectomy and Fusion (ACDF) for
Diagnostics 2023, 13, 2429 10 of 33

cervical reticulopathy, as they are more prone to the occurrence of this condition [65–69].
Goedmakers et al. aimed to predict the development of adjacent segment disease (ASD) in
patients undergoing surgery for cervical radiculopathy using DL techniques and preopera-
tive MRI data [70]. They developed a DL model with 48 convolutional layers trained on
preoperative T2 sagittal cervical MRI images. The study included 344 eligible patients, of
whom 60% (n = 208) were used for training, and 40% for validation (n = 43) and testing
(n = 93). The results demonstrated that the deep learning model outperformed assessments
made by neuroradiologists and neurosurgeons. The DL model achieved an accuracy of
95%, sensitivity of 80%, and specificity of 97%, whereas the other evaluators had lower
accuracy rates of 58%. These findings suggest the potential of DL algorithms to improve the
accuracy of ASD prediction based on preoperative MRI data in patients ACDF surgery for
cervical radiculopathy. The study had limitations, including reliance on the last available
follow-up, lack of consideration for clinical and demographic characteristics, variability in
surgical techniques and outcomes, small number of MRI scans, imbalanced distribution of
ASD cases, and potential limitations of GradCAM saliency maps.
Karhade et al. specifically applied NLP to operative notes to determine if such analy-
ses could effectively process the large and free-text inputs in our medical record systems,
beginning with the operative notes [71,72]. In a cohort of 1000 patients, NLP identified
93 inadvertent durotomies with an AUC of 0.99 [71]. Within the testing set, the NLP algo-
rithm exhibited an impressive sensitivity of 0.89, successfully detecting 16 out of 18 patients
who had incidental durotomy. The study has several limitations, including its retrospective
design within a single healthcare system, the influence of shared surgical practices on docu-
mentation, the lack of prospective and external validation, the potential for unrecognized or
unrecorded incidental durotomies, and the impracticality of multiple reviews by different
researchers or spine surgeons. The same group used NLP to retrospectively understand the
risk factors associated with intraoperative vascular injuries from operative notes [72]. The
study found that body mass index, diabetes, L4-L5 exposure, and infection-related surgery
(discitis, osteomyelitis) were the best predictors. NLP had a sensitivity of 0.92 when identi-
fying VI from operative notes. Moreover, the algorithm successfully identified 18 out of the
21 patients with VI, resulting in a sensitivity of 0.86. Thus, neurosurgical documentation
may be developed as an input to ML models. The study has multiple limitations, including
its retrospective design limited to a single healthcare entity, the necessity for prospective
validation across multiple institutions, the absence of a well-established gold standard for
intraoperative vascular injury, the potential overfitting of the NLP algorithm, the possibility
of enhancing performance through collaborative efforts and alternative machine learning-
based NLP approaches, and the potential influence of changes in coding practices on the
algorithm’s accuracy. Future research could consider comparing institutional records with
national databases to evaluate the algorithm’s ability to capture adverse events.
Postoperative opioid prolonged use is a significant concern following spine surgery. Since
there has been a noted rise in complications associated with opioids, [73–78] Karhade et al.
employed five predictive models, namely elastic-net penalized logistic regression, random
forest, stochastic gradient boosting, neural network, and the support vector machine, to
construct models for predicting prolonged opioid prescriptions [79]. A total of 5413 patients
were identified, among whom 416 individuals (7.7%) maintained a prescription for opioid
medication between 90 and 180 days following their surgical procedures. The elastic-
net penalized logistic regression model had the best discrimination (c-statistic 0.81) and
good calibration and overall performance. The investigation revealed that preoperative
prediction of prolonged postoperative opioid prescription could enhance surveillance and
monitoring post-surgery. Notably, the three most influential predictors in the models
were instrumentation, duration of preoperative opioid prescription, and comorbidity of
depression. These findings underscore the potential of preoperative prediction in improving
the management of postoperative opioid use in patients undergoing spinal surgery. The
study acknowledges several limitations, including the unavailability of opioid dose data,
the exclusion of illicit opioid use, approximation of opioid use based on medical record
Diagnostics 2023, 13, 2429 11 of 33

data, lack of patient-reported outcomes, limited applicability to other spinal conditions,

potential influence of changing surgical techniques over the study period, limited diversity
of institutions, and the need for external validation. The web application provided offers
opportunities for further evaluation of the models.
ML has also become popular in predicting the accuracy of discharge timelines to
improve hospital census. In their study, Stopa et al. investigated the factors influencing
delayed discharge after elective spine surgery [80]. A cohort of 144 patients underwent
elective inpatient surgery to address lumbar disc disorders, and among them, the rate of
nonroutine discharge stood at 6.9% (n = 10). The parameters analyzed included age, sex,
BMI, ASA class, preoperative functional status, number of fusion levels, comorbidities,
preoperative laboratory findings, and discharge disposition. Performance metrics included:
discrimination (c-statistic), calibration, and positive and negative predictive values. When
applied to the institutional data, the neural network algorithm demonstrated excellent
generalization capabilities, as indicated by a c-statistic (AUC) of 0.89. Moreover, the
algorithm exhibited a calibration slope of 1.09 and a calibration intercept of −0.08, further
attesting to its reliability. Regarding predictive accuracy, at a threshold of 0.25, the positive
predictive value (PPV) reached 0.50, while the negative predictive value (NPV) stood
impressively at 0.97. Despite the positive findings of this external validation study, it
is important to consider other predictive algorithms developed for this population, as
they have demonstrated varying levels of performance in terms of discrimination and
calibration, highlighting the need for further research and direct comparisons between
algorithms in prospective samples. This application to the delayed discharge prediction
highlights the applicability of ML to all stages of a patient’s healthcare interaction, where
hospitals can translate ML to an array of stages.
For the continuation or follow-up of the treatment plan, it is crucial to identify the
brand of the instrument fusion specifically. Huang et al. utilized ML techniques to classify
anterior cervical fusion systems, achieving a remarkable accuracy rate of 91.5% ± 3.8%
across at least nine distinct types [81]. When considering the top three classifications, the
accuracy increased to 98.4% ± 1.3%. These findings demonstrate the potential of machine
learning algorithms in conjunction with clinical information to accurately classify anterior
cervical fusion systems. The limitations of the current model include the limited number
of hardware systems available for training, the need for additional datasets to evaluate
visual artifacts and overlapping radiopaque “noise”, the requirement for prospective data
to assess clinical utility, and the potential applications of hardware classification beyond
revision ACDF surgery. This paper proves to be highly beneficial in situations where
patients undergo fixation treatment and experience difficulty recalling the specific brand of
rod and screw. Notably, its accuracy in such cases is notably commendable.

Current Challenges
Using machine learning algorithms, Fatima et al. developed a predictive model
for adverse events after lumbar degenerative spondylolisthesis surgery [62]. Logistic
regression outperformed LASSO methods, and a web application was created for risk
assessment. The study acknowledges limitations such as data variation, limited postop-
erative outcome data, suboptimal model performance, potential missing variables, and
coding errors. Karhade et al. achieved promising results in predicting short-term mortality
in spinal metastatic disease using machine learning algorithms and an open access web
application [63]. Limitations include data veracity and completeness, limited predictors,
inability to capture overall disease trajectory, and need for further evaluation. Ames et al.
applied AI-based clustering to classify ASD surgery, but limitations include reliance on
radiographic parameters, manual segregation challenges, and need for validation [64].
Goedmakers et al. demonstrated that a deep learning algorithm outperformed experts in
predicting adjacent segment disease [70]. Limitations included limited follow-up, demo-
graphic considerations, surgical variability, small sample size, imbalanced distribution,
and limitations of saliency maps. Karhade et al. used NLP algorithms to detect incidental
Diagnostics 2023, 13, 2429 12 of 33

durotomy, but further validation is needed [71]. Another study by Karhade et al. developed
algorithms for detecting intraoperative vascular injury during lumbar spine surgery, but
external and prospective validation is necessary [72]. Karhade et al. also created prediction
algorithms for prolonged opioid prescription after lumbar disc herniation surgery, with
limitations including missing data and limited patient-reported outcomes [79]. Stopa et al.
validated a machine learning algorithm for predicting nonroutine discharge after spinal
surgery, but other algorithms should be considered for direct comparisons [80]. Huang et al.
developed a computer vision algorithm for classifying anterior cervical fusion systems, but
further validation and exploration are needed [81]. The utilization of AI in the classification
of spine diseases has been extensive. Still, the focus lies in using AI to simulate surgical
procedures for enhanced surgical planning and outcome visualization. Key considerations
in spinal surgery involve determining the optimal decompression and fixation extent,
accounting for the dynamic nature of the spine and the potential risks associated with
excessive fixation. Tailoring individualized surgical plans is essential. Retrospective data
collection on AI implementation and examination of durotomy and vascular injury cases
offer valuable insights. At the same time, robotic surgery adoption faces cost challenges
that could be addressed through AI-driven multicenter data analysis. Postoperative care
necessitates optimal analgesic use and effective discharge planning, with the potential for
ChatGPT to bridge the gap in patient self-observation and engagement in this context.

4. Epilepsy
AI has been utilized for predicting the outcome of epilepsy surgery since 1998 [11,82–89].
In one of these early investigations, Grigsby et al. developed a simulated neural network
(SNN) to predict seizure-free outcomes after anterior temporal lobectomy using model data
from 87 patients [82]. They determined that SNN was superior to a discriminant function
in its ability to predict Class 1 (completely seizure-free) and Class 1 or Class 2 (almost or
totally seizure-free). The accuracy of the SNNs was 81.3% vs. 78.5% and 95.4% vs. 72.7%,
respectively. The retrospective design using patient records and the need for prospective
validation with new patients, as well as the potential inclusion of additional input variables
such as SPECT and PET, are acknowledged as limitations; however, the study results
indicate that simulated neural networks have potential as decision-making adjuncts in
epilepsy surgery.
Torlay et al. utilized a language network analysis to classify epilepsy patients based
on pre-operative fMRI data [90]. To address this issue, the Extreme Gradient Boosting
(XGBoost) technique was employed on five language areas (three frontal and two temporal)
activated by fMRI for phonological (PHONO) and semantic (SEM) language tasks. The
study, which included 135 patients, found that the subset of left frontotemporal activation
caused by the SEM task could distinguish two groups (healthy/typical vs. epilepsy/atypical)
with the highest accuracy (AUC of 91 ± 5%).
Meanwhile, Memarian et al. conducted a study on predicting post-surgical outcomes
in complicated cases of mesial temporal lobe epilepsy [89]. This retrospective study em-
ployed supervised ML to predict postsurgical seizure independence in drug-resistant focal
seizures of temporal origin. The study included 20 preoperative patients; these individuals
were diagnosed with mesial temporal lobe epilepsy (MTLE) and subsequently underwent
the standard procedure of anteromesial temporal lobectomy. The results showed that a
combination of maximum relevance minimal redundancy (mRMR) and LA-SVM classifier
predicted surgical outcomes with 95% accuracy in atypical mesial temporal lobe epilepsy.
The limited spatial coverage of depth electrodes in intracranial EEG recordings poses a
constraint, as they are not consistently implanted in all brain areas among patients, but
the study’s findings, regarding a higher number of contacts at seizure onset and greater
seizures in the ipsilateral amygdala, support the efficacy of amygdalohippocampectomy
for achieving seizure freedom in this patient population.
Some groups have leveraged bigger data sets for powerful clinical applications.
Abbasi et al. used ML to improve epilepsy diagnosis and therapy by prediction of pharma-
Diagnostics 2023, 13, 2429 13 of 33

ceutical response, medical and surgical outcomes, and seizure detection from EEG video
and kinetic data [91]. The study highlights the limitations of machine learning techniques in
epilepsy, particularly the lack of external validation studies, and emphasizes the importance
of larger datasets, cloud-based repositories, and robust external validation to improve the
generalizability and interpretability of machine learning models for enhanced clinician
confidence and integration into clinical practice. Hosseini et al. studied epileptogenicity
locations using Multimodal rs-fMRI and EEG [92]. The study was divided into three phases.
First, autonomic edge computing was used to process patient data for determining surgical
candidacy. Next, EEG and rs-MRI were used to predict epileptogenic networks. Finally, an
unsupervised model-based electrocorticography (ECoG) signals were created to separate
interictal epileptic discharge (IED) periods from non-IED periods. The study highlights
the limitation of current computational algorithms in reliably identifying preictal periods
for effective intervention in epilepsy, emphasizing the need for an autonomic method that
accurately detects and localizes epileptogenicity to enhance seizure control and improve
quality of life. Using this information as feedback, the authors aimed to improve upon
responsive neurostimulation (RNS; Neurospace) management of epilepsy patients.
Temporal lobe epilepsy (TLE) is the most prevalent form of drug-resistant epilepsy
in adults [93]. In a noteworthy study by Larivière et al., a multimodal MRI investigation
was conducted on 30 drug-resistant TLE patients, utilizing a supervised machine learning
approach with fivefold cross-validation [94]. When comparing TLE patients to normal
subjects, the findings revealed decreased connectivity distance within the Temporoinsular
and Prefrontal networks. Notably, imaging data from patients who underwent anterior
temporal lobectomy for seizure treatment and were followed up for one year exhibited an
accuracy of 76±4%. While this accuracy may not yet meet translational standards, it marks
a significant step forward. This study presents a captivating narrative concerning epilepsy
surgery and opens up possibilities for leveraging past data to gain novel insights across
multiple centers. However, due to the small training set size, there may be potential bias in
patient selection, which is a factor that should be considered in many ML applications in less
frequent conditions. The study encountered limitations in sample size, but regularization
techniques were used, and the classifier’s performance was compared to a baseline model;
however, variability in follow-up times and lack of generalizability to other types of
drug-resistant focal epilepsies require further investigation, with initiatives like ENIGMA-
Epilepsy being valuable for data coordination, while the openly available surface-based
features used in the study can facilitate validation and dissemination.

Current Challenges
Grisby et al. demonstrate the potential of simulated neural networks (SNN) as decision-
making tools for patient selection in epilepsy surgery, while acknowledging the need for
further validation and prospective studies [82]. Torlay et al. show promising results in
identifying language patterns in epilepsy patients using functional MRI and the Extreme
Gradient Boosting algorithm, but call for further research and discussion on the limita-
tions [90]. Abbasi et al. highlight the progress and potential of machine learning in epilepsy
but note the lack of critical analysis of challenges and limitations [91]. Hosseini et al.
propose autonomic edge computing for epilepsy monitoring but acknowledge the limita-
tions of current computational algorithms in identifying preictal periods accurately [92].
Memarian et al. demonstrate the accuracy of supervised machine learning in predicting
postsurgical outcomes for temporal lobe epilepsy, emphasizing the need for additional
features [89]. Larivière et al. explore functional and structural changes in epilepsy using
machine learning and suggest the role of connectivity distance contractions in personalized
surgical prognostication, while acknowledging sample size limitations and the need for
further investigation [94].
Epilepsy surgery holds significant importance, but its complexity and limited patient
population pose challenges to acquire an accurate diagnosis. Integrating AI as a screening
tool and diagnostic aid in epilepsy shows great potential for improving precision and pa-
Diagnostics 2023, 13, 2429 14 of 33

tient care. Furthermore, virtual reality (VR) technology in epilepsy surgery allows surgeons
to visualize and simulate procedures with enhanced accuracy and safety. While previous
AI research on epilepsy surgery outcomes has provided valuable insights, modern ML
applications with larger sample sizes, multicenter data collection, and modern algorithm
designs, with mitigation of potential biases, will be necessary before translation to the
epilepsy domain.

5. Vascular
Artificial intelligence has been widely used in diagnostic imaging to detect cerebrovas-
cular lesions [95–97]. Park et al. have used DL to diagnose cerebral aneurysms [97]. The
objective of their research was to develop a neural network segmentation model called the
“HeadXnet Model” for the prediction of intracranial aneurysms from computed tomogra-
phy angiography (CTA) data. To achieve this, they utilized a training dataset comprising
611 head CTA data to generate accurate aneurysm segmentation. The model was then
evaluated by radiologists using 115 test cases. The study was conducted at one academic
medical center where the model was trained, validated, and tested on 818 CTA examina-
tions from 662 patients. Among these cases, 328 were diagnosed with cerebral aneurysms
(40.1%), while 490 were negative (59.9%), with the exclusion of cases involving hemorrhage,
ruptured aneurysms, arteriovenous malformations, surgical clips, coils, catheters, or other
surgical devices. This study’s findings revealed noteworthy improvements among clini-
cians in various performance measures. The mean sensitivity demonstrated a significant
increase of 0.059 (95% CI, 0.028–0.091; adjusted p = 0.01), while the mean accuracy exhibited
a notable increase of 0.038 (95% CI, 0.014–0.062; adjusted p = 0.02). Furthermore, the mean
interrater agreement (Fleiss κ) displayed a considerable enhancement, rising from 0.799 to
0.859 with an increase of 0.060 (adjusted p = 0.05). Conversely, there was no statistically
significant change observed in mean specificity, with an increase of 0.016 (95% CI, −0.010
to 0.041; adjusted p = 0.16), or in the meantime to diagnosis, with a difference of 5.71 s
(95% CI, 7.22–18.63 s; adjusted p = 0.19). The study has several limitations, including
the exclusion of ruptured aneurysms and aneurysms associated with other conditions,
uncertainty regarding the model’s performance in the presence of surgical hardware or
devices, potential interpretation bias due to the high prevalence of aneurysms in the test set
and the binary task of clinicians, and limited generalizability of the findings to institutions
with different imaging protocols and equipment, as the study was conducted using data
from a single institution.
Silva et al. conducted a study utilizing ML techniques to investigate clinical features
for detecting aneurysm rupture [98]. They employed three models: RF, linear SVM, and
radial basis function kernel SVM. The analysis encompassed 845 aneurysms in 615 patients,
of which 309 were classified as ruptured aneurysms. Among the ruptured aneurysms,
307 exhibited aneurysm rupture, accounting for approximately 37% of the study population.
The findings revealed that ruptured aneurysms were larger and more commonly located
in the posterior circulation than unruptured aneurysms. The ML models achieved AUC
values of 0.77 for linear SVM, 0.78 for radial basis function kernel SVM, and 0.81 for the
random forest model. The study demonstrated the ability of these ML models to predict
aneurysm rupture based on factors such as size and location, with posterior, anterior, and
posterior inferior cerebellar arteries frequently associated with aneurysm rupture. Con-
versely, Paraclinoid and middle cerebral arteries showed a lower likelihood of rupture.
These findings align with previous research highlighting the strong correlation between
aneurysm location, size, and the risk of rupture. Overall, this study underscores the effec-
tiveness of ML in analyzing complex and extensive data within the field of cerebrovascular
neurosurgery, identifying location and size as significant predictors of aneurysm rupture.
The study has limitations including the single-institution nature of the patient cohort, the
need to assess model performance on external data, the retrospective nature of the data
comparing ruptured and unruptured cases, and the lack of long-term follow-up data on
Diagnostics 2023, 13, 2429 15 of 33

untreated aneurysms, suggesting that prospective validation and inclusion of additional

radiographic characteristics would be valuable.
Predicting the stability of aneurysms is very useful, especially for small aneurysms [99–102].
Liu et al. utilized machine learning and radiomics-derived morphological features to
predict aneurysm stability, specifically focusing on small aneurysms [103]. The LASSO
regression identified several significant morphological features for predicting aneurysm
stability. Flatness emerged as the most crucial feature, followed by spherical disproportion,
maximum 2D diameter slice, and surface area. Notably, surface area (odds ratio [OR], 0.697;
95% CI, 0.476–0.998), spherical disproportion (OR, 1.730; 95% CI, 1.143–2.658), flatness (OR,
0.584; 95% CI, 0.374–0.894), hyperlipemia (OR, 2.410; 95% CI, 1.029–5.721), multiplicity
(OR, 0.182; 95% CI, 0.082–0.380), location of middle cerebral artery (OR, 0.359; 95% CI,
0.134–0.902), and internal carotid artery (OR, 0.087; 95% CI, 0.030–0.211) were included in
the final prediction model. The model exhibited commendable performance, with an AUC
of 0.853 (95% CI, 0.767–0.940).
Regarding unstable aneurysms, patients with hypertension demonstrated lower values
for compactness (p = 0.035), sphericity (p = 0.035), and flatness (p = 0.010), while spherical
disproportion (p = 0.034) was found to be higher. This showcases the potential of ML
in enhancing our understanding and management of aneurysms, especially for smaller
aneurysms. QingLin et al.’s work contributes valuable insights to the field of aneurysm
stability prediction and emphasizes the role of advanced algorithms in improving clinical
decision-making and patient care. This study’s limitations include its single-center nature,
the reliance on post-rupture morphology as a surrogate for rupture risk evaluation, the
potential misclassification of unstable aneurysms in the absence of definite symptoms,
and the limited focus on aneurysms within a specific size range, impeding analysis of
smaller aneurysms.
Aneurysmal subarachnoid hemorrhage (aSAH) is frequently linked to high-mortality
rates and unfavorable neurological outcomes [104–107]. In an effort to shed light on
prognostic indicators, Koch et al. employed the elastic net (EN) regression and orthogonal
partial least squares-discriminant analysis (OPLS-DA) to investigate cerebrospinal fluid
(CSF) metabolites associated with poor outcomes [108]. Through the analysis of CSF
samples from 81 aSAH patients, 138 metabolites were measured and quantified in each
sample. The study revealed the significance of certain vasoactive molecules within the
nitric oxide pathway, namely symmetric dimethylarginine (SDMA), dimethylguanidine
valeric acid (DMGV), and ornithine, which is associated with poor mRS at discharge
(p = 0.0005, 0.002, and 0.0001, respectively) and at 90 d (p = 0.0036, 0.0001, and 0.004,
respectively). Furthermore, in comparison to the nonaneurysmal subarachnoid hemorrhage
controls, SDMA exhibited a notably higher concentration in the cerebrospinal fluid (CSF),
signifying a significant elevation (p = 0.0087). These metabolites were identified as robust
predictors of poor outcomes following severe aSAH. This finding offers valuable insights
into the underlying molecular mechanisms and provides a potential avenue for prognostic
assessment in patients with aSAH. The biased patient cohort limited the retrospective
study, lack of correlation between metabolite levels and vasospasm, moderate effect sizes,
and potential changes in metabolite profiles over time. Nonetheless, the study further
supports the role of symmetric dimethylarginine (SDMA) in subarachnoid hemorrhage
(SAH) pathophysiology, highlighting the importance of targeting secondary injury. The
unique insight provided by cerebrospinal fluid (CSF) analysis near the site of injury, and
whether these metabolites are markers or causative agents, requires further investigation.
Delayed cerebral ischemia (DCI) represents a significant and detrimental complication
associated with aneurysmal subarachnoid hemorrhage (aSAH) [109]. To enhance prediction
accuracy, Ramos et al. employed ML algorithms that incorporated clinical data combina-
tions with imaging data to forecast the occurrence of DCI [110]. The study, conducted on
a cohort of 317 patients, demonstrated that utilizing ML algorithms (e.g., LR, SVM, RFC
and MLP) significantly improved DCI prediction. The best AUC of the logistic regression
models was 0.63 (95% ci 0.62 to 0.63). The ML algorithms with clinical data improved the
Diagnostics 2023, 13, 2429 16 of 33

AUC to 0.68 (95% ci 0.65 to 0.69). ML models incorporating clinical data and image features
achieved the highest AUC values, reaching 0.74 (95% CI 0.72 to 0.75). These findings
indicate the potential of ML in augmenting the accuracy of DCI prediction for patients with
aSAH, highlighting the value of integrating clinical and imaging data in the prediction
process. A limitation of the LR model used in the study is the low number of events per
feature, making it prone to overfitting, while the ML algorithms employed can handle
high-dimensional feature spaces with less risk of overfitting, but still require external
validation; furthermore, determining the best parameter configurations for the ML models
can be computationally expensive, and interpreting the 3D image features is challenging,
indicating the need for future research in alternative feature extraction techniques for better
visualization and interpretation.
In addition to aneurysms, AI has been employed to investigate the factors influenc-
ing brain arteriovenous malformation following endovascular embolization, including
imaging and clinical presentation to predict procedure complication and outcomes [111].
The study comprised a cohort of 199 participants who underwent brain arteriovenous
malformation (BAVM) treatment, with an average follow-up duration of 63 months. The
results demonstrated that the standard regression analysis model demonstrated an accu-
racy of 43% in predicting the outcome (mortality), with the predictor being the overlap of
treatment. In contrast, ML exhibited a remarkable accuracy of 97% in outcome prediction of
mortality, identifying the presence or absence of a nidal fistula as the most significant factor,
irrespective of blinding. Machine learning algorithms have limitations, including their de-
pendency on large training datasets for improved performance and accuracy, the challenge
of uncovering the true underlying relationships between factors, the risk of overfitting
with irrelevant data, and the need for techniques like cross-validation or regularization to
optimize performance and prevent random errors.
Microvascular anastomosis, a surgical procedure that demands exceptional skill, repre-
sents a significant clinical challenge. Mastery of this technique requires extensive training,
dedication, and persistence. In light of this, Gonzalez-Romo et al. conducted a comprehen-
sive investigation into hand motion during microvascular anastomosis, utilizing a CNN to
track 21 hand positions [112]. The study involved six participants, including two experts,
two intermediates, and two novices, with no physical constraints imposed on their hand
movements. During the subsequent 600-s stimulation period, four non-experts performed
26 anastomotic bites, with an average excess motion of 14.3 (15.5) seconds per bite. In
contrast, the expert group completed 33 bites (18 and 15, respectively), exhibiting a mean
(SD) excess motion of 2.8 (2.3) seconds per bite for the dominant hand. Additionally, within
a 180 s timeframe, the experts accomplished 13 bites, with mean (SD) latencies of 22.2 (4.4)
and 23.4 (10.1) seconds, while the intermediate group achieved 9 bites, displaying mean
(SD) latencies of 31.5 (7.1) and 23.4 (22.1) seconds per bite. These findings present an in-
triguing contribution to the field. Although the study featured a relatively small number of
participants, the implications are noteworthy, as the results can serve as a valuable resource
for future endeavors. This research has the potential to significantly benefit aspiring young
neurosurgeons embarking on their journey in microvascular anastomosis, providing them
with an opportunity to assess and enhance their skills by comparing their performance to
that of experts. Limitations of their study include a small sample size, absence of prospec-
tive follow-up, limited assessment of other technique domains, unclear understanding of
the relationship between motion analysis and learning curves using different simulators,
and the need for further validation and application of the hand detector in clinical settings
and with established assessment scales.

Current Challenges
In a series of studies, researchers utilized various machine learning (ML) techniques to
enhance the prediction and understanding of different aspects of aneurysms and cerebral
vascular conditions. Park et al. found that integrating the HeadXNet neural network model
can enhance clinician performance in detecting intracranial aneurysms, but limitations
Diagnostics 2023, 13, 2429 17 of 33

include the exclusion of ruptured aneurysms, uncertainty regarding the model’s perfor-
mance with surgical hardware, potential interpretation bias, and limited generalizability to
other institutions [97]. Silva et al. demonstrated that machine learning models effectively
differentiate between ruptured and unruptured aneurysms based on location and size,
but limitations include the single-institution nature of the study and the need for external
validation [98]. Liu et al. developed a machine learning model to predict aneurysm stability
based on morphological features, but limitations include the single-center nature of the
study and the focus on post-rupture morphology [103]. Koch et al. identified metabolites
associated with poor outcomes in aneurysmal subarachnoid hemorrhage using machine
learning, but limitations include the biased patient cohort and the need for further investi-
gation [108]. Ramos et al. showed improved prediction of delayed cerebral ischemia using
machine learning algorithms, but limitations include the moderate predictive accuracy
and the need for external validation [110]. Asadi et al. demonstrated the superiority of
machine learning in predicting outcomes for brain arteriovenous malformations, but limi-
tations include the dependency on large training datasets and the risk of overfitting [111].
Gonzalez-Romo et al. developed a machine learning-based hand motion detector for mi-
crovascular anastomosis simulation, but limitations include the small sample size and the
need for further validation and clinical application [112].
The application of AI In aneurysm detection and monitoring has shown promise,
although more comparable information is needed for Cavernoma classification. Machine
learning (ML) can potentially classify conditions like arteriovenous malformations (AVMs).
Complex treatment planning for procedures like clip aneurysms with bypass requires
collaboration across departments, and simulating treatment plans using AI for success
assessment could be valuable. Integrating AI to capture and compare hand motions
between experts and beginners can accelerate skill development and potentially lead to
AI-assisted surgical coaching. While molecular-level outcome monitoring has begun, post-
surgical care and surveillance require improvement, with limited information available for
vasospasm prevention and hypotension monitoring.
The applications of ML to neurosurgery in tumor, spine, epilepsy, and vascular sub-
domains that were discussed above are summarized in Table 1. There are also currently
several interesting clinical trials that have been sourced from clinical trials registered in
ClinicalTrials.gov. Most of these trials focus on the diagnostic tests for glioma, followed by
clinical trials on aneurysm, which were observational studies presented in Table 2.
Diagnostics 2023, 13, 2429 18 of 33

Table 1. Studies evaluating machine learning algorithms used for neurosurgical outcome prediction.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
Tumor
- Large sample size influences NLP
classification models.
- Conference papers were excluded
AUC = 0.87 ± 0.09
from the review.
Disease Diagnosis, AUC, Sensitivity, Sensitivity = 0.87 ± 0.10;
Buchlak et al., 2021 [22] Glioma MRI data CNN, SVM, RF 153 - Optimized deep language models
Outcome Specificity, Accuracy Specificity = 0.0.86 ± 0.10;
are suggested for improved
Precision = 0.88 ± 0.11
performance.
- Readers are referred to specific
papers for further information.
- Retrospective design with a small
number of patients from two
academic institutions.
- The findings may have limited
AUC = 0.94 generalizability to other settings.
(95% CI: 0.91–0.97) - The use of PNG exports of DICOM
GBM and PCNSL for GBM images results in
McAvoy et al., 2021 [23] Disease Diagnosis AUC CNN 320
MRI data AUC = 0.95 data loss.
(95% CI: 0.92–0.98) - There is no direct comparison
for PCNL. between the classification outcomes
of CNNs and radiologists.
- Further research is needed to
determine the clinical value of
the tool.
Dice score of 85.2% (mean
Hausdorff = 8.8 mm; mean
- Limited in its ability to evaluate
average Hausdorff
post-operative residuals, tumor
distance = 0.4)
recurrence, or tumor growth due to
Median of 88.2% (median
the inclusion of single
Automatically Dice score, Hausdorff = 5.0 mm;
pre-operative scans.
Boaro et al., 2021 [24] segment meningiomas Meningioma MRI data Hausdorff distance, 3D-CNN 806 median average Hausdorff
- Model’s detection performance was
from MRI scan Inter-expert variability distance = 0.2 mm)
not tested on brain MRI scans
Inter-expert variability in
without meningioma.
segmenting the same
- Algorithm has not been integrated
tumors with means
into the hospital informatics system.
ranging from 80.0
to 90.3%
-Retrospective design and focuses
IDH genotype and IDH AUC training 0.921, specifically on known gliomas.
Preoperative MRI of
Zhou et al., 2019 [25] 1p19q codeletion in AUC, Accuracy ML, RF 538 validate 0.919 - Limiting its applicability to
glioma patients
gliomas Accuracy 78.2% different tumor types and non
- tumor mimickers.
ANN model Predicting the
position of the nodes with - Use of generic mechanical
Load-driven FEM
Tonutti et al., 2017 [32] Tumor deformation Accuracy, Specificity ANN, SVR - errors <0.3 mm parameters and exclusion of certain
simulations of tumor
SVR models positional brain structures
errors < 0.2 mm
Diagnostics 2023, 13, 2429 19 of 33

Table 1. Cont.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
AUC = 0.945 - Reliance on NIR-II
FL-CNN higher Sensitivity fluorescence imaging.
Intraoperative glioma Fluorescence of AUC, Sensitivity, 93.8% vs. 82.0%, p < 0.001) - While NIR-II offers advantages
Shen et al., 2021 [33] FL-CNN 1874
diagnosis glioma tissue Specificity Predict grade and over NIR-I, it may still have lower
Ki-67 level specificity compared to clinically
(AUC 0.810 and 0.625) available methods
Raman spectroscopy,
Diagnose glioma coherent anti-Stokes accuracy of 92%
Hollon et al., 2021 [34] molecular classes Raman scattering (CARS) Accuracy CNN - sensitivity = 93%
intraoperatively microscopy, Stimulated specificity = 91%
Raman histology (SRH)
The study includes limitations such
as a wide time span for
patient inclusion.
Conditional survival - Including lymph node metastasis
forest, a Cox as an LMD risk factor is novel and
proportional hazards requires more investigation.
XGboost AUC = 0.83
Predict outcomes of Clinical Characteristic model, Extreme - Patients receiving only radiation
RFand Cox proportional
Tewarie et al., 2022 [40] LMD patients in patient in Risk ratio, p value gradient boosting 1054 therapy were excluded from
hazards model
Brain Metastasis Brain Metastasis (XGBoost), Extra trees, the study.
C-index = 0.76
LR, Synthetic Minority - Use of SMOTE reduced
Oversampling data variability.
Technique (SMOTE) - LMD prognostication at brain
metastases (BM) diagnosis is
theoretical and not yet widely used
in clinical care.
- Use of retrospective data for
internal validation.
AUC of 0.71
Predicts 6-month - The study focuses on survival at a
Gradient boosting, predicted both 6-month
survival after Data of Brain AUC, Calibration, 6-month cutoff rather than overall
Hulsbergen et al., 2022 [44] K-nearest neighbors, 1062 and longitudinal overall
neurosurgical Metastasis patient Brier score median survival.
LR, NB, RF, SVM survival
resection for BM - Intraoperative and postoperative
(p < 0.0005)
factors can influence
survival prediction.
- Being restricted to continuous and
binary models.
- Unable to compute subject-level
AFT, Boosted decision
Demographic, survival curves and lacks
trees survival, CPHR,
Predict Survival in Socioeconomic, interpretability.
Senders et al., 2018 [49] C-index RF, recursive 20,821 C-index = 0.70
GBM patients Radiographical, - Computational inefficiency
partitioning
Therapeutic Characteristics - Evaluating models based on
algorithms
multiple criteria
- Factors unrelated to
prediction performance.
Diagnostics 2023, 13, 2429 20 of 33

Table 1. Cont.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
- Patient cohort is small and from a
Deep Learning, single institution.
Sørensen–Dice coefficient, Hybrid Watershed 843 preopMRIs from - Lack of comparison with
Preoperative MRI of low- Comparing manually and
Sensitivity, Specificity, Algorithm, Robust 843 patients with gliomas other approaches.
or high-grade gliomas, automatically derived
Evaluation of Dunnet’s test, Spearman’s Learning-Based Brain 713 longitudinal postop - Smaller tumors were excluded
Chang et al., 2019 [52] Postoperative MRI with longitudinal changes in
treatment response rank correlation coefficient, Extraction, Brain MRI from 54 patients with from the study.
newly diagnosed tumor burden were 0.917,
intraclass correlation ExtractionTool, newly diagnosed - Variability in MR imaging
glioblastoma 0.966, and 0.850
coefficient (ICC) 3dSkullStrip, Brain glioblastomas availability.
Surface Extractor - Confidence assessment in
segmentations is absent
Brain tumor
Median Accuracy = 92%
ANN
Dice similarity
SVMFuzzy C-means
coefficient = 88%
Bayesian Learning
Presurgical planning, Radiological of
RFQuadratic - Need for more detailed analysis of
Intraoperative critical/target brain
Median accuracy discriminant analysis all studies and a focus on
guidance, median Accuracy = 94%
Dice similarity LDA perioperative care applications.
Senders et al., 2018 [53] Neurophysiological Neurosurgical treatment 6402 Dice similarity
Median sensitivity Gaussian mixture - Caution is advised when
monitoring, and coefficient = 91%
coefficient models interpreting the quantitative
Neurosurgical Predict epileptogenic focus
LR, performance summary.
outcome prediction Median Accuracy = 86%
K-nearest neighbor,
Detect seizure by iEEG
NLP
Median Sensitivity = 96%
K-means
Intraop tumor demarcation
Median Accuracy = 89%
Spine
Gender, age, American
Society of
-Variation in patient and surgical
Anesthesiologists grade, AUC = 0.7
characteristics within the
Autogenous iliac bone Brier score = 0.08
database used.
Clinical graft, Instrumented fusion, Discrimination, Predicting overall AEs
- Limited postoperative outcome
Fatima et al., 2020 [62] decision-making, Levels of surgery, Surgical Calibration, Brier score, LRand LASSO 3965 Logistic regression = 0.70
data beyond 30 days
Patient outcomes approach, Functional Decision analysis (95% CI, 0.62–0.74)
- Potential missing variables and
status, Preoperative serum LASSO = 0.65 (95% CI,
coding errors in the data are
albumin (g/dL), Serum 0.61–0.69)
additional limitations.
alkaline phosphatase
(IU/mL)
- Variable data veracity.
- Limited availability of
Discrimination (c-statistic), pertinent predictors
Calibration (assessed by - Unable to capture the overall
Postoperative Preoperative SVM, SVM0.760
Karhade et al., 2019 [63] calibration slope and 1790 trajectory of metastatic disease
outcome prognostic factor NeuralNetwork (NN) NNwith c-statistic 0.769.
intercept), Brier score, - lack of explanatory capability.
Decision analysis - No examination of multivariate
logistic regression or proportional
hazard models.
Diagnostics 2023, 13, 2429 21 of 33

Table 1. Cont.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
- Dependency on sample size
Predict surgical Unsupervised - Observation heterogeneity for
Ames et al., 2019 [64] Patient, Surgical factor p-Value 570 overall p-value 0.004
outcome hierarchical clustering determining patient and
operative clusters.
- Reliance on the last available
follow up.
- Clinical and demographic
Accuracy, Sensitivity, characteristics were not considered
Predict ASD
Predicting Adjacent Specificity, PPV, NPV, in the analysis.
Goedmakers et al., 2021 VGGNet19, Resnet18, Accuracy = 95%
Segment Disease Preoperative Cervical MRI F1-score, Matthew 344 - Variability in surgical techniques
[70] Resnet50 Sensitivity = 80%
(ASD) correlation coefficient, and outcomes.
Specificity = 97%
Informedness, Markedness - Small number of MRI scans limited
the study.
- Distribution of ASD cases
were imbalanced.
- Retrospective nature within a
single healthcare system
- Influence of shared surgical
practices on documentation could
AUC-ROC = 0.99
affect the results.
Incidental durotomies operative notes of patients Sensitivity = 0.89
AUC-ROC, Precision-recall - Unrecognized or unrecorded
Karhade et al., 2020 [71] in free-text undergoing lumbar NLP 1000 Specificity = 0.99
curve, Brier score incidental durotomies may have
operative notes spine surgery PPV = 0.89
been overlooked.
NPV = 0.99.
- Impracticality of multiple reviews
by different researchers or spine
surgeons is a limitation of the
current work.
- Retrospective design from a single
healthcare entity.
C-statistic = 0.92 - Prospective and multi-institutional
age, male sex, body mass C-statstic, Sensitivity,
Sensitivity 0.86 validation is needed to confirm the
index, diabetes, L4-L5 Specificity,
Intraoperative Specificity = 0.93 findings.
Karhade et al., 2021 [72] exposure, and PPV, NLP 1035
vascular injury PPV = 0.51 - Lack of a rigorous gold standard
infection-related surgery NPV,
NPV = 0.99 for intraoperative vascular injury is
(discitis, osteomyelitis) F1-score
F1-score of 0.64. a limitation.
- NLP algorithm used in the study
may be prone to overfitting
- Unavailability of opioid dose data
and exclusion of illicit opioid use.
- Opioid use approximation was
C-statistic = 0.81
Prediction of based on medical record data
Elastic-net AUC 0.81
prolonged opioid Chart review of patients - Patient-reported outcomes were
C-statistic or AUC, penalizedLR, RF, calibration
Karhade et al., 2019 [79] prescription after undergoing surgery for 5413 not included in the study.
Calibration, Brier Score Stochastic Gradient (slope = 1.13,intercept = 0.13)
surgery for lumbar lumbar disc herniation - Changing surgical techniques over
Boosting, NN, SVM overall performance
disc herniation the study period could have
(Brier = 0.064)
influenced the results.
- The study included a limited
diversity of institutions.
Diagnostics 2023, 13, 2429 22 of 33

Table 1. Cont.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
Age, Sex, BMI, ASA class,
Preoperative functional AUC, Discrimination Python (version 3.6) AUC 0.89, - Positive findings in terms of
status, Number of fusion (c-statistic), Calibration, and the R calibration slope = 1.09, external validation.
Stopa et al., 2019 [80] Nonroutine discharge levels, Comorbidities, and Positive and Negative programming 144 calibration - Different algorithms have shown
Preoperative laboratory predictive values language intercept = −0.08. varying levels of performance in
findings, Discharge (PPVs and NPVs) (version 3.5.1). PPV = 0.50NPV = 0.97. discrimination and calibration.
disposition
- Limited number of available
hardware systems for training.
KAZE feature detector - Additional datasets are needed to
K-means clustering evaluate visual artifacts and
Identification of MATLAB software Top choice 91.5% ± 3.8% overlapping radiopaque “noise.”
AP film cervical Cross-validation analysis
Huang et al., 2019 [81] implanted spinal Vision System Toolbox 321 2 choice 97.1% ± 2.0% - Prospective data is required to
radiography after ACDF Accurracy
hardware and Statistics and 3 choice 98.4% ± 1.3% assess the clinical utility of
Machine the model.
Learning Toolbox - Potential applications of hardware
classification beyond revision
ACDF surgery.
Epilepsy
History, Demographics,
Clinical examination,
Routine scalp EEG,
- Retrospective design with
Video-scalp EEG
patient records
Predict seizure monitoring, Intracranial Accuracy = 81.3%
Grisby et al., 1998 [82] Accuracy SNN 87 - Prospective validation with new
outcomes EEG monitoring, and 95.4%
patients is needed for
Intracarotid amobarbital
further validation
(Wada) testing, CT, MRI,
Neuropsychological
assessment
Atypical language
patterns Differentiate
Torlay et al., 2017 [90] fMRI AUC ML, XGBOOST 55 AUC = 91 ± 5%
patients with epilepsy
from healthy people
Electroencephalography
- limitations in reliably identifying
(EEG), Resting
p-value preictal periods.
Epilepsy Seizure state-functional Magnetic Multiple t-test, Differential
Hosseini et al., 2017 [92] CNN 9 Normal 1.85 × 10-14 - Need for an autonomic method
Localization Resonance Imaging connectivity graph (DCG)
Seizure 4.64 × 10 -27 that accurately detects and localizes
(rs-fMRI), Diffusion Tensor
epileptogenicity.
Imaging (DTI)
LDA, NB, SVM with
radial basis function - The limited spatial coverage of
Clinical,
kernel (SVM-rbf), depth electrodes in intracranial EEG
Electrophysiological,
Predict surgery SVM with multilayer recordings poses a constraint.
Memarian et al., 2015 [89] Structural magnetic Accuracy 20 Accuracy = 95%
outcome perceptron kernel - Depth electrodes are not
resonance imaging
(SVM-mlp), consistently implanted in all brain
(MRI) features
Least-Square SVM areas among patients.
(LS-SVM).
Diagnostics 2023, 13, 2429 23 of 33

Table 1. Cont.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
- Limitations in sample size.
- Regularization techniques were
Supervised machine
Predict postsurgical used- Variability in follow-up times
Larivière et al., 2020 [94] Multimodal MRI imaging Accuracy learning with fivefold 30 Accuracy = 76± 4%
seizure outcome and lack of generalizability to other
cross-validation
types of drug-resistant
focal epilepsies
Vascular
mean Sensitivity - Exclusion of ruptured aneurysms
increased = 95%, and aneurysms associated with
mean Accuracy other conditions.
increased = 95%, - Performance of the model in the
Clinician performance Sensitivity, Specificity,
mean Interrater agreement presence of surgical hardware or
Park et al., 2019 [97] with and without CTA examinations Accuracy, time, CNN 818
(Fleiss κ) increased = 0.060, devices remains uncertain.
model augmentation interrater agreement
from 0.799 to 0.859 - Potential interpretation bias
(adjusted p = 0.05) may exist
mean Specificity = 95% - Conducted using data from a
Time to Diagnosis 95% single institution.
- Single institution for the
patient cohort
p value, AUC, Sensitivity, AUC - The retrospective nature of the data
Clinical Features, RF, Linear SVM,
Vascular imaging data of Specificity, Linear SVM = 0.77 comparing ruptured and
Silva et al., 2019 [98] Detection of Radial basis function 845
cerebral aneurysms PPV, Radial basis function unruptured cases is a limitation.
Aneurysm Rupture kernel SVM
NPV kernel SVM = 0.78 - Long-term follow-up data on
untreated aneurysms is lacking,
which affects the analysis.
Flatness (OR, 0.584; 95% - Single-center nature
CI, 0.374–0.894) - Reliance on post-rupture
Spherical Disproportion morphology as a surrogate for
(OR, 1.730; 95% CI, rupture risk evaluation
Predicting Morphological p value, Odds ratio, AUC, 1.143–2.658) - Potential misclassification of
Liu et al., 2019 [103] Lasso regression 1139
Aneurysm Stability feature aneurysm chi square test, t test SurfaceArea unstable aneurysms without
(OR) = 0.697 (95% CI, definite symptoms
0.476–0.998) - Limited focus on aneurysms within
AUC = 0.853 (95% CI, a specific size range, hindering
0.767–0.940) analysis of smaller aneurysms.
Diagnostics 2023, 13, 2429 24 of 33

Table 1. Cont.

1st Author Paper, Year Output Input Output Measures ML Model Number of Enrollment Model Performance Limitation
Poor mRS Biased patient cohort.
At Discharge - No correlation found between
Elastic net (EN) ML,
Vasoactive molecule p value (p = 0.0005, 0.002, metabolite levels and vasospasm.
Orthogonal partial
Koch et al., 2021 [108] that predict CSF of aSAH patients 2-tailed student t-test, 138 and 0.0001) - Effect sizes observed were
least squares-
poor outcome Fischer’s exact test At 90 day moderate.
(OPLS-DA)
(p = 0.0036, 0.0001, - Possibility of changes in metabolite
and 0.004) profiles over time.
- LR model used in the study had a
Logistic regression models
limitation of a low number of events
AUC = 0.63 (95% CI 0.62
per feature, making it prone
to 0.63)
to overfitting.
LR, SVM, RFMLP, ML with clinical data
- ML algorithms used in the study
Prediction of Delay Stock Convolutional AUC = 0.68 (95% CI 0.65
Ramos et al., 2019 [110] Clinical and CT image data AUC, 317 can handle high-dimensional feature
Cerebral Ischemia Denoising to 0.69)
spaces with less risk of overfitting
Auto-encoder, PCA ML with clinical data and
but still require external validation.
image feature
- Determining the best parameter
AUC = 0.74 (95% CI 0.72
configurations for ML models can
to 0.75)
be computationally expensive.
- ML algorithms depend on large
training datasets for improved
performance and accuracy.
Study documented Supervised Machine
Outcome variables, - Uncovering the true underlying
imaging, Clinical learning MATLAB
Asadi et al., 2016 [111] Clinical outcome Accuracy 199 Accuracy = 97.5% relationships between factors can be
presentation, Procedure, Neural Network
prediction challenging for ML algorithms.
complications, Outcomes Toolbox
- There is a risk of overfitting when
irrelevant data is included in the
training process.
6oo s
4 nonexpert 26 bites total
- Small sample size.
2 expert 33 bites(18 bites
- Prospective follow up was
and 15 bites)
not conducted.
180 s
Microvascular 21 tracking hand Python programming - Assessment of other technique
Gonzalez-Romo et al., 2023 Mean (SD), One-way Expert, 13 bites with mean
anastomosis landmarks from language and 6 domains was limited.
[112] ANOVA latencies of 22.2(4.4) and
hand motion 6 participant Mediapipe; CNN - The relationship between motion
23.4 (10.1) seconds
analysis and learning curves using
2 intermediate, 9 bites with
different simulators is not
mean latencies of 31.5(7.1)
well understood.
and 34.4 (22.1) seconds
per bites
Diagnostics 2023, 13, 2429 25 of 33

Table 2. Clinical trials on artificial intelligence in neurosurgery.

Trial or Registry Number Enroll Condition Interventions Outcome Measure Status

Diagnostic Accuracy of Study Arms
Diagnostic Test: Artificial Intelligence
Central Nervous System
NCT04671368 141 Diagnostic Test: Practicing Pathologists Sensitivity and specificity of Study Arms Unknown status
Neoplasms
Diagnostic Test: Gold Standard
Spearman Coefficient of Study Arms related to Gold Standard
Diagnostic Test: MR and Histopathology
NCT04220424 500 Glioma images based prediction of molecular AUC of Prediction performance Unknown status
pathology and patient survival
Changes of Response to Treatment
Progression-Free Survival (PFS)
NCT04216550 600 Recurrent Glioma Drug: Apatinib Recruiting
Overall Survival (OS)
Incidence of treatment-related adverse events
Diagnostic Test: Survival prediction for
NCT04215211 2500 Glioma AUC of survival prediction performance Recruiting
glioma patients
Diagnostic Test: Prediction of molecular
NCT04217018 3000 Glioma AUC of prediction performance Recruiting
pathology
Diagnostic Test: Histopathology images
NCT04215224 3500 Glioma based survival prediction for glioma AUC of survival prediction performance Recruiting
patients
Diagnostic Test: Histopathology images
NCT04217044 3000 Glioma AUC of prediction performance Recruiting
based prediction of molecular pathology
Aneurysm rupture
NCT04872842 1000 Intracranial Aneurysm Other: Obervation Completed
Aneurysm growth
Aneurysm rupture
NCT05608122 1000 Intracranial Aneurysm Other: Obervation Recruiting
Aneurysm growth
Algorithm Performance
Algorithm Processing Time
Time to Notification
NCT04733638 500 Intracerebral Hemorrhage Device: Viz ICH VOLUME Time to Treatment Enrolling by invitation
Length of Stay
In Hospital Complications
Modified Rankin Scale (mRS) at Discharge and 90 Days
Diagnostics 2023, 13, 2429 26 of 33

Table 2. Cont.

Trial or Registry Number Enroll Condition Interventions Outcome Measure Status

Intracranial aneurysm size
Intracranial aneurysm volume
Intracranial aneurysm height
Intracranial aneurysm neck diameter
NCT05804474 1500 Intracranial Aneurysm Other: Observational study Completed
Parent artery diameter
Intracranial aneurysm width
Aspect ratio
Size ratio
Transfer patients: Time from spoke CT/CTA to door-out
Non-transfer patients: Time from Hub door to groin puncture
Time from Spoke Door-In to Door-Out (DIDO)
Time from Spoke CT/CTA to Specialist Notification
Device: Viz LVO (De Novo Time from Spoke CT/CTA to Groin Puncture
NCT04608617 1000 Stroke, Ischemic Recruiting
Number DEN170073) Time from Spoke Door to Groin Puncture
Length of ICU Stay/Total Length of Stay
Modified Rankin Scale (mRS) at Discharge and 90 Days
National Institutes of Health Stroke Scale (NIHSS) at Discharge
Patient Disposition at Discharge and 90 Days
Other: The questionnaire “Cervical JOACMEQ questionnaire
myelopathy treatment outcome mJOA questionnaire
questionnaire” Cervical myelopathy treatment outcome questionnaire”
NCT05099627 300 Cervical Myelopathy Not yet recruiting
Other: The retrospective questionnaire CSM early diagnosis questionnaire”
“CSM early diagnosis questionnaire” Nurick score
Other: JOACMEQ questionnaire and mJOA DOR (diagnostic odds ratio)
Low Back Pain Visual Analogue Scale for Back Pain
NCT05161130 1115 Disk Herniated Lumbar Procedure: Lumbar Spinal Fusion Visual Analogue Scale for Leg Pain Completed
Spinal Stenosis Lumbar Oswestry Disability Index
Accuracy of the artificial intelligence model
NCT04359745 500 Glioblastoma Recruiting
Failure rate of the artificial intelligence model
Cerebral Cavernous
Circulating Diagnostic and Prognostic Biomarkers of CASH
Malformation
NCT04467489 1040 Other: observational Correlation of Imaging and Plasma Biomarkers of CASH Recruiting
Cavernous Angioma
Confounders of CASH Biomarkers
Hemorrhagic Microangiopathy
modified Rankin Scale
NCT04819074 4000 Aneurysm, Brain Procedure: Microsurgery Sensorimotor neurological deficits Recruiting
Clavien Dindo Complication Grading
Diagnostics 2023, 13, 2429 27 of 33

6. Future Directions
AI has already demonstrated its potential in various aspects of neurosurgery, such as
surgical planning, navigation, and image analysis. Looking into the future, AI is expected
to play an increasingly significant role in neurosurgery, potentially revolutionizing the field.
First, as a precision medicine tool, it can assist neurosurgeons in developing personalized
treatment plans. By analyzing an extensive cohort of patient data, medical records, imaging,
and genomics, ML can identify patterns that predict treatment response for individual
patients. Second, AI supports surgical planning and navigation. Patient imaging can
be processed to enable more accurate surgical guidance and real-time feedback during
procedures, thereby reducing operative errors. Third, AI enhances the efficiency and
accuracy of large data processing, thereby improving diagnoses or uncovering novel
therapies. Finally, AI has many important implications for medical education, providing
new means of accessing data repositories, such as operative videos to personalize learning
and enhanced patient education.
The recent rise of generative AI has the potential to catalyze AI for neurosurgery
in multiple ways. First, generative AI can synthesize new data making training possi-
ble for rare conditions and allow for sharing of such synthetic data across centers for
multiple-center dataset designs. Second, generative foundational models represent a mas-
sive increase in the ability to understand longitudinal multimodal patient data from the
patient record and incorporate this data into current predictive models for outcome pre-
diction, surgical planning, and decision–making support. Lastly, large language models
like GPT-4 may facilitate ease of use for both clinicians and patients. Clinicians may inter-
act with AI research assisted by GPT-4 code generation as prompted in plain English by
clinicians, and automated dataset analysis. Patients can benefit from these models, as such
language generation can support physician communication with patients, where patient
education is tailored to the individual. One particular use of large language models is
that of transparency, where the models can be queried as to what data in their training
set was used to make a certain prediction, which may both solve the current ‘black box’
Diagnostics 2023, 13, x FOR PEER REVIEW 34 of 40
problem as well as allow for active limitation of biases as AI is translated to neurosurgical
practice. Based on the information presented above, Figure 4 succinctly encapsulates the
forthcoming trajectory of AI in the field of neurosurgery.

Figure 4. Future direction AI in neurosurgery.

Figure 4 Future direction AI in neurosurgery.

7. Limitations
Alongside future successes, it is crucial to acknowledge potential challenges that the
field may face. One obstacle is a lack of translation and scalability, as clinicians do not
widely adopt many AI models in neurosurgery and lack external replication and
Diagnostics 2023, 13, 2429 28 of 33

7. Limitations
Alongside future successes, it is crucial to acknowledge potential challenges that the
field may face. One obstacle is a lack of translation and scalability, as clinicians do not
widely adopt many AI models in neurosurgery and lack external replication and validation.
Additionally, the regulations surrounding AI in healthcare are unclear, and the absence of
mandated representation of different backgrounds, such as ethnicities and races in training
sets, may perpetuate biases observed in other fields like drug development. Defining AI as a
software medical tool in patient care is necessary. Another barrier is limited generalizability
of findings, particularly to medically underserved and marginalized groups. This is
partially caused by a limited ability to obtain or share data across institutions, creating non-
representative training sets. These necessitate a conversation on the potential for similar
regulations for AI models as those imposed for clinical trials in drug development (subject
cohort design, adverse reporting, etc. . . . ). Finally, as technology progresses beyond ANNs
to large language models (LLMs), the cost of training these tools increases. While industry
funding may be necessary, the involvement of academia in the development of AI for
neurosurgery tools should be considered, and the implications of proprietary AI tools in
the field should be examined.

8. Conclusions
Recent studies have demonstrated the potential of ML in various aspects of neuro-
oncological care, including tumor identification and classification, differential diagnosis,
segmentation, molecular classification, personalized anatomical models, intraoperative di-
agnosis, and survival prediction. ML techniques have also shown promise in spine surgery,
predicting adverse events, mortality rates, adjacent segment disease, and various surgical
outcomes. In epilepsy, ML has been utilized for predicting seizure outcomes, classifying
patients based on fMRI data, predicting postsurgical seizure independence, determining
surgical candidacy, and investigating drug-resistant cases. Additionally, ML has been
applied to aneurysm prediction, stability assessment, metabolite identification, cerebral
ischemia prediction, brain arteriovenous malformation outcome prediction, and microvas-
cular surgery skill assessment. These studies highlight the broad impact and potential of
ML in improving patient management and outcomes in neurology and neurosurgery.
In light of what has been stated above, AI is becoming an increasingly common tool
in neurosurgery. We provide a synopsis of the primary translational research by dividing
the field into four vital neurosurgical sections: tumor, spine, epilepsy, and vascular. As
the level of complexity in AI continues to rise, it is vital for us to have knowledge of AI
and to be aware of how to maximize its benefits for patient care. Such understanding
is essential in order to comprehend the recommendations for application and to be able
to anticipate the ongoing trend toward employing AI in the future. This technological
understanding will allow us to provide better care for patients, beginning with the diagnosis
and counseling, continuing through the planning and procedure, and extending into the
post-operative period.

Author Contributions: Conceptualization, J.A.T., D.M.P. and C.K.; methodology, J.A.T., D.M.P. and
T.R.S.; software, E.S., M.Z., A.V., T.R.S., A.H.E.-S. and B.S.G.; Validation, A.V.; T.R.S., A.H.E.-S.
and B.S.G.; formal analysis, J.A.T., D.M.P. and C.K.; investigation, A.V., T.R.S., A.H.E.-S. and
B.S.G.; resources, J.A.T., M.T.L. and D.M.P.; data curation, J.A.T., R.H., M.T.L., D.M.P. and C.K.;
writing—original draft preparation, J.A.T., E.S. and M.Z.; writing—review and editing, J.A.T., E.S.,
M.Z., T.R.S. and B.S.G.; visualization, A.V., T.R.S. and B.S.G.; supervision, R.H., M.T.L., D.M.P. and
C.K.; project administration, N/A; funding acquisition, none. All authors have read and agreed to
the published version of the manuscript.
Funding: This paper received no external funding.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Diagnostics 2023, 13, 2429 29 of 33

Data Availability Statement: Not applicable.

Conflicts of Interest: Chayakrit Krittanawong discloses the following relationships – Member of the
American College of Cardiology Solution Set Oversight Committee, the American Heart Association
Committee of the Council on Genomic and Precision Medicine, the American College of Cardiol-
ogy/American Heart Association (ACC/AHA) Joint Committee on Clinical Data Standards, and
the ACC/AHA Task Force on Performance Measures, The Lancet Digital Health (Advisory Board),
European Heart Journal Digital Health (Editorial board), Journal of the American Heart Association
(Editorial board), JACC: Asia (Section Editor), The Journal of Scientific Innovation in Medicine (Asso-
ciate Editor), and Frontiers in Cardiovascular Medicine (Associate Editor). This intellectual property
has not been licensed. All other authors have no disclosure.

References
1. Wang, L.; Delgado-Baquerizo, M.; Wang, D.; Isbell, F.; Liu, J.; Feng, C.; Liu, J.; Zhong, Z.; Zhu, H.; Yuan, X.; et al. Diversifying
Livestock Promotes Multidiversity and Multifunctionality In Managed Grasslands. PNAS 2019, 116, 6187–6192. [CrossRef] [PubMed]
2. Obermeyer, Z.; Emanuel, E.J. Predicting the Future—Big Data, Machine Learning, and Clinical Medicine. N. Engl. J. Med. 2016,
375, 12161219. [CrossRef]
3. Senders, J.T.; Staples, P.C.; Karhade, A.V.; Zaki, M.M.; Gormley, W.B.; Broekman, M.L.; Smith, T.R.; Arnaout, O. Machine Learning
and Neurosurgical Outcome Prediction: A Systematic Review. World Neurosurg. 2018, 109, 476–486.e1. [CrossRef] [PubMed]
4. Senders, J.T.; Arnaout, O.; Karhade, A.V.; Dasenbrock, H.H.; Gormley, W.B.; Broekman, M.L.; Smith, T.R. Natural and Artificial
Intelligence in Neurosurgery: A Systematic Review. Neurosurgery 2017, 83, 181–192. [CrossRef]
5. Buchlak, Q.D.; Esmaili, N.; Leveque, J.-C.; Farrokhi, F.; Bennett, C.; Piccardi, M.; Sethi, R.K. Machine Learning Applications to
Clinical Decision Support in Neurosurgery: An Artificial Intelligence Augmented Systematic Review. Neurosurg. Rev. 2019, 43,
1235–1253. [CrossRef]
6. Elfanagely, O.; Toyoda, Y.; Othman, S.; Mellia, J.A.; Basta, M.; Liu, T.; Kording, K.; Ungar, L.; Fischer, J.P. Machine Learning and
Surgical Outcomes Prediction: A Systematic Review. J. Surg. Res. 2021, 264, 346–361. [CrossRef] [PubMed]
7. Raj, J.D.; Nelson, J.A.; Rao, K.S.P. A Study on the Effects of Some Reinforcers to Improve Performance of Employees in a Retail
Industry. Behav. Modif. 2006, 6, 848–866. [CrossRef] [PubMed]
8. Noble, W.S. What Is a Support Vector Machine? Nat. Biotechnol. 2006, 24, 1565–1567. [CrossRef] [PubMed]
9. Raschka, S.; Mirjalili, V. Python Machine Learning: Machine Learning and Deep. Learning with Python, Scikit-Learn, and TensorFlow,
2nd ed.; Packt Publishing Ltd.: Birmingham, UK, 2017.
10. Deo, R.C. Machine Learning in Medicine. Circulation 2015, 132, 1920–1930. [CrossRef] [PubMed]
11. Munsell, B.C.; Wee, C.-Y.; Keller, S.S.; Weber, B.; Elger, C.; da Silva, L.A.T.; Nesland, T.; Styner, M.; Shen, D.; Bonilha, L. Evaluation
Of Machine Learning Algorithms for Treatment Outcome Prediction in Patients With Epilepsy Based on Structural Connectome
Data. Neuroimage 2015, 118, 219–230. [CrossRef]
12. Staartjes, V.E.; de Wispelaere, M.P.; Vandertop, W.P.; Schröder, M.L. Deep Learning-Based Preoperative Predictive Analytics for
Patient-Reported Outcomes Following Lumbar Discectomy: Feasibility of Center-Specific Modeling. Spine J. 2019, 19, 853–861.
[CrossRef] [PubMed]
13. Izadyyazdanabadi, M.; Belykh, E.; Mooney, M.; Martirosyan, N.; Eschbacher, J.; Nakaji, P.; Preul, M.C.; Yang, Y. Convolutional
Neural Networks: Ensemble Modeling, Fine-Tuning and Unsupervised Semantic Localization for Neurosurgical CLE Images. J.
Vis. Commun. Image Represent. 2018, 54, 10–20. [CrossRef]
14. Chauhan, N.K.; Singh, K. A Review on Conventional Machine Learning vs Deep Learning. In Proceedings of the 2018 International
Conference on Computing, Power and Communication Technologies, GUCON 2018, Greater Noida, India, 28–29 September 2018;
pp. 347–352. [CrossRef]
15. Doppalapudi, S.; Qiu, R.G.; Badr, Y. Lung Cancer Survival Period Prediction and Understanding: Deep Learning Approaches. Int.
J. Med. Informatics 2020, 148, 104371. [CrossRef]
16. Corso, J.J.; Sharon, E.; Dube, S.; El-Saden, S.; Sinha, U.; Yuille, A. Efficient Multilevel Brain Tumor Segmentation With Integrated
Bayesian Model Classification. IEEE Trans. Med. Imaging 2008, 27, 629–640. [CrossRef]
17. Bauer, S.; Nolte, L.-P.; Reyes, M. Fully Automatic Segmentation of Brain Tumor Images Using Support Vector Machine Classifica-
tion in Combination with Hierarchical Conditional Random Field Regularization. Med. Image Comput. Comput. Assist. Interv.
2011, 14, 354–361. [CrossRef] [PubMed]
18. Ismael, S.A.A.; Mohammed, A.; Hefny, H. An Enhanced Deep Learning Approach for Brain Cancer MRI Images Classification
Using Residual Networks. Artif. Intell. Med. 2019, 102, 101779. [CrossRef] [PubMed]
19. Lukas, L.; Devos, A.; Suykens, J.; Vanhamme, L.; Howe, F.; Majós, C.; Moreno-Torres, A.; Van Der Graaf, M.; Tate, A.; Arús, C.; et al.
Brain Tumor Classification Based On Long Echo Proton MRS Signals. Artif. Intell. Med. 2004, 31, 73–89. [CrossRef] [PubMed]
20. Akkus, Z.; Ali, I.; Sedlář, J.; Agrawal, J.P.; Parney, I.F.; Giannini, C.; Erickson, B.J. Predicting Deletion of Chromosomal Arms 1p/19q
in Low-Grade Gliomas from MR Images Using Machine Intelligence. J. Digit. Imaging 2017, 30, 469–476. [CrossRef] [PubMed]
21. Díaz-Pernas, F.; Martínez-Zarzuela, M.; Antón-Rodríguez, M.; González-Ortega, D. A Deep Learning Approach for Brain Tumor
Classification and Segmentation Using a Multiscale Convolutional Neural Network. Healthcare 2021, 9, 153. [CrossRef] [PubMed]
Diagnostics 2023, 13, 2429 30 of 33

22. Buchlak, Q.D.; Esmaili, N.; Leveque, J.C.; Bennett, C.; Farrokhi, F.; Piccardi, M. Machine learning applications to neuroimaging
for glioma detection and classification: An artificial intelligence augmented systematic review. J. Clin. Neurosci. 2021, 89, 177–198.
[CrossRef]
23. McAvoy, M.; Prieto, P.C.; Kaczmarzyk, J.R.; Fernández, I.S.; McNulty, J.; Smith, T.; Yu, K.H.; Gormley, W.B.; Arnaout, O.
Classification of glioblastoma versus primary central nervous system lymphoma using convolutional neural networks. Sci. Rep.
2021, 11, 15219. [CrossRef] [PubMed]
24. Boaro, A.; Kaczmarzyk, J.R.; Kavouridis, V.K.; Harary, M.; Mammi, M.; Dawood, H.; Shea, A.; Cho, E.Y.; Juvekar, P.; Noh, T.; et al.
Deep neural networks allow expert-level brain meningioma segmentation and present potential for improvement of clinical
practice. Sci. Rep. 2022, 12, 15462. [CrossRef] [PubMed]
25. Zhou, H.; Chang, K.; Bai, H.X.; Xiao, B.; Su, C.; Bi, W.L.; Zhang, P.J.; Senders, J.T.; Vallières, M.; Kavouridis, V.K.; et al. Machine
learning reveals multimodal MRI patterns predictive of isocitrate dehydrogenase and 1p/19q status in diffuse low- and high-grade
gliomas. J. Neurooncol. 2019, 142, 299–307. [CrossRef]
26. Huang, H.; Yang, G.; Zhang, W.; Xu, X.; Yang, W.; Jiang, W.; Lai, X. A Deep Multi-Task Learning Framework for Brain Tumor
Segmentation. Front. Oncol. 2021, 11. [CrossRef]
27. Yousef, R.; Khan, S.; Gupta, G.; Siddiqui, T.; Albahlal, B.M.; Alajlan, S.A.; Haq, M.A. U-Net-Based Models towards Optimal MR
Brain Image Segmentation. Diagnostics 2023, 13, 1624. [CrossRef] [PubMed]
28. Juarez-Chambi, R.M.; Kut, C.; Rico-Jimenez, J.J.; Chaichana, K.L.; Xi, J.; Campos-Delgado, D.U.; Rodriguez, F.J.; Quinones-
Hinojosa, A.; Li, X.; Jo, J.A. AI-Assisted In Situ Detection of Human Glioma Infiltration Using a Novel Computational Method for
Optical Coherence Tomography. Clin. Cancer Res. 2019, 25, 6329–6338. [CrossRef] [PubMed]
29. Jermyn, M.; Desroches, J.; Mercier, J.; St-Arnaud, K.; Guiot, M.-C.; Leblond, F.; Petrecca, K. Raman Spectroscopy Detects Distant
Invasive Brain Cancer Cells Centimeters Beyond MRI Capability in Humans. Biomed. Opt. Express 2016, 7, 5129–5137. [CrossRef]
[PubMed]
30. Schucht, P.; Mathis, A.M.; Murek, M.; Zubak, I.; Goldberg, J.; Falk, S.; Raabe, A. Exploring Novel Innovation Strategies to Close
a Technology Gap in Neurosurgery: HORAO Crowdsourcing Campaign. J. Med. Internet Res. 2023, 25, e42723. [CrossRef]
[PubMed]
31. Achkasova, K.A.; Moiseev, A.A.; Yashin, K.S.; Kiseleva, E.B.; Bederina, E.L.; Loginova, M.M.; Medyanik, I.A.; Gelikonov,
G.V.; Zagaynova, E.V.; Gladkova, N.D. Nondestructive Label-Free Detection of Peritumoral White Matter Damage Using
Cross-Polarization Optical Coherence Tomography. Front. Oncol. 2023, 13, 1133074. [CrossRef] [PubMed]
32. Tonutti, M.; Gras, G.; Yang, G.-Z. A Machine Learning Approach For Real-Time Modelling of Tissue Deformation in Image-Guided
Neurosurgery. Artif. Intell. Med. 2017, 80, 39–47. [CrossRef]
33. Shen, B.; Zhang, Z.; Shi, X.; Cao, C.; Zhang, Z.; Hu, Z.; Ji, N.; Tian, J. Real-time intraoperative glioma diagnosis using fluorescence
imaging and deep convolutional neural networks. Eur. J. Nucl. Med. Mol. Imaging. 2021, 48, 3482–3492. [CrossRef] [PubMed]
34. Hollon, T.; Orringer, D.A. Label-Free Brain Tumor Imaging Using Raman-Based Methods. J. Neuro-Oncology 2021, 151, 393–402.
[CrossRef] [PubMed]
35. Emblem, K.E.; Pinho, M.C.; Zöllner, F.G.; Due-Tonnessen, P.; Hald, J.K.; Schad, L.R.; Meling, T.R.; Rapalino, O.; Bjornerud, A. A
Generic Support Vector Machine Model for Preoperative Glioma Survival Associations. Radiology 2015, 275, 228–234. [CrossRef]
36. Akbari, H.; Macyszyn, L.; Da, X.; Bilello, M.; Wolf, R.L.; Martinez-Lage, M.; Biros, G.; Alonso-Basanta, M.; O’Rourke, D.M.;
Davatzikos, C. Imaging Surrogates of Infiltration Obtained Via Multiparametric Imaging Pattern Analysis Predict Subsequent
Location of Recurrence of Glioblastoma. Neurosurgery 2016, 78, 572–580. [CrossRef]
37. Emblem, K.E.; Due-Tonnessen, P.; Hald, J.K.; Bjornerud, A.; Pinho, M.C.; Scheie, D.; Schad, L.R.; Meling, T.R.; Zoellner, F.G.
Machine Learning In Preoperative Glioma MRI: Survival Associations by Perfusion-Based Support Vector Machine Outperforms
Traditional MRI. J. Magn. Reson. Imaging 2013, 40, 47–54. [CrossRef] [PubMed]
38. Knoll, M.A.; Oermann, E.K.; Yang, A.I.; Paydar, I.; Steinberger, J.; Collins, B.; Collins, S.; Ewend, M.; Kondziolka, D. Survival
of Patients With Multiple Intracranial Metastases Treated With Stereotactic Radiosurgery. Am. J. Clin. Oncol. 2018, 41, 425–431.
[CrossRef] [PubMed]
39. Azimi, P.; Shahzadi, S.; Sadeghi, S. Use Of Artificial Neural Networks to Predict the Probability of Developing New Cerebral
Metastases After Radiosurgery Alone. J. Neurosurg. Sci. 2020, 64, 52–57. [CrossRef]
40. Tewarie, I.A.; Senko, A.W.; Jessurun, C.A.C.; Zhang, A.T.; Hulsbergen, A.F.C.; Rendon, L.; McNulty, J.; Broekman, M.L.D.;
Peng, L.C.; Smith, T.R.; et al. Predicting leptomeningeal disease spread after resection of brain metastases using machine learning.
J. Neurosurg. 2022, 1–9. [CrossRef] [PubMed]
41. Blonigen, B.J.; Steinmetz, R.D.; Levin, L.; Lamba, M.A.; Warnick, R.E.; Breneman, J.C. Irradiated Volume as a Predictor of Brain
Radionecrosis After Linear Accelerator Stereotactic Radiosurgery. Int. J. Radiat. Oncol. 2010, 77, 996–1001. [CrossRef]
42. Chang, E.L.; Wefel, J.S.; Hess, K.R.; Allen, P.K.; Lang, F.F.; Kornguth, D.G.; Arbuckle, R.B.; Swint, J.M.; Shiu, A.S.; Maor, M.H.; et al.
Neurocognition in Patients With Brain Metastases Treated With Radiosurgery or Radiosurgery Plus Whole-Brain Irradiation:
A Randomised Controlled Trial. Lancet Oncol. 2009, 10, 1037–1044. [CrossRef]
43. Mardor, Y.; Roth, Y.; Ocherashvilli, A.; Spiegelmann, R.; Tichler, T.; Daniels, D.; Maier, S.E.; Nissim, O.; Ram, Z.; Baram, J.; et al.
Pretreatment Prediction of Brain Tumors Response to Radiation Therapy Using High b-Value Diffusion-Weighted MRI. Neoplasia
2004, 6, 136–142. [CrossRef] [PubMed]
Diagnostics 2023, 13, 2429 31 of 33

44. Hulsbergen, A.F.C.; Lo, Y.T.; Awakimjan, I.; Kavouridis, V.K.; Phillips, J.G.; Smith, T.R.; Verhoeff, J.J.C.; Yu, K.H.; Broekman,
M.L.D.; Arnaout, O. Survival Prediction After Neurosurgical Resection of Brain Metastases: A Machine Learning Approach.
Neurosurgery 2022, 91, 381–388. [CrossRef] [PubMed]
45. Lacroix, M.; Abi-Said, D.; Fourney, D.R.; Gokaslan, Z.L.; Shi, W.; DeMonte, F.; Lang, F.F.; McCutcheon, I.E.; Hassenbusch, S.J.;
Holland, E.; et al. A Multivariate Analysis Of 416 Patients with Glioblastoma Multiforme: Prognosis, Extent of Resection, and
Survival. J. Neurosurg. 2001, 95, 190–198. [CrossRef] [PubMed]
46. Cairncross, J.G.; Ueki, K.; Zlatescu, M.C.; Lisle, D.K.; Finkelstein, D.M.; Hammond, R.R.; Silver, J.S.; Stark, P.C.; Macdon-
ald, D.R.; Ino, Y.; et al. Specific Genetic Predictors of Chemotherapeutic Response and Survival in Patients with Anaplastic
Oligodendrogliomas. Gynecol. Oncol. 1998, 90, 1473–1479. [CrossRef]
47. Eckel-Passow, J.E.; Lachance, D.H.; Molinaro, A.M.; Walsh, K.M.; Decker, P.A.; Sicotte, H.; Pekmezci, M.; Rice, T.W.; Kosel, M.L.;
Smirnov, I.V.; et al. Glioma Groups Based on 1p/19q, IDH, and TERTPromoter Mutations in Tumors. N. Engl. J. Med. 2015, 372,
2499–2508. [CrossRef]
48. Weller, M.; Stupp, R.; Reifenberger, G.; Brandes, A.A.; Bent, M.J.V.D.; Wick, W.; Hegi, M.E. MGMT Promoter Methylation in
Malignant Gliomas: Ready for Personalized Medicine? Nat. Rev. Neurol. 2009, 6, 39–51. [CrossRef]
49. Senders, J.T.; Staples, P.; Mehrtash, A.; Cote, D.J.; Taphoorn, M.J.B.; Reardon, D.A.; Gormley, W.B.; Smith, T.R.; Broekman, M.L.;
Arnaout, O. An Online Calculator for the Prediction of Survival in Glioblastoma Patients Using Classical Statistics and Machine
Learning. Neurosurgery 2020, 86, E184–E192. [CrossRef] [PubMed]
50. Law, M.; Young, R.J.; Babb, J.S.; Peccerelli, N.; Chheang, S.; Gruber, M.L.; Miller, D.C.; Golfinos, J.G.; Zagzag, D.; Johnson, G.
Gliomas: Predicting Time to Progression or Survival with Cerebral Blood Volume Measurements at Dynamic Susceptibility-
weighted Contrast-enhanced Perfusion MR Imaging. Radiology 2008, 247, 490–498. [CrossRef]
51. Price, S.J.; Jena, R.; Burnet, N.G.; Carpenter, T.A.; Pickard, J.D.; Gillard, J.H. Predicting Patterns of Glioma Recurrence Using
Diffusion Tensor Imaging. Eur. Radiol. 2007, 17, 1675–1684. [CrossRef]
52. Chang, K.; Beers, A.L.; Bai, H.X.; Brown, J.M.; Ly, K.I.; Li, X.; Senders, J.T.; Kavouridis, V.K.; Boaro, A.; Su, C.; et al. Automatic
Assessment of Glioma Burden: A Deep Learning Algorithm for Fully Automated Volumetric and Bidimensional Measurement.
Neuro-Oncology 2019, 21, 1412–1422. [CrossRef] [PubMed]
53. Senders, J.T.; Zaki, M.M.; Karhade, A.V.; Chang, B.; Gormley, W.B.; Broekman, M.L.; Smith, T.R.; Arnaout, O. An Introduction and
Overview of Machine Learning in Neurosurgical Care. Acta Neurochir. 2017, 160, 29–38. [CrossRef] [PubMed]
54. Winkler-Schwartz, A.; Bissonnette, V.; Mirchi, N.; Ponnudurai, N.; Yilmaz, R.; Ledwos, N.; Siyar, S.; Azarnoush, H.; Karlik, B.;
Del Maestro, R.F. Artificial Intelligence in Medical Education: Best Practices Using Machine Learning to Assess Surgical Expertise
in Virtual Reality Simulation. J. Surg. Educ. 2019, 76, 1681–1690. [CrossRef]
55. Celtikci, E. A Systematic Review on Machine Learning in Neurosurgery: The Future of Decision Making in Patient Care. Turk.
Neurosurg. 2017, 28, 167–173. [CrossRef] [PubMed]
56. Staartjes, V.E.; Stumpo, V.; Kernbach, J.M.; Klukowska, A.M.; Gadjradj, P.S.; Schröder, M.L.; Veeravagu, A.; Stienen, M.N.;
van Niftrik, C.H.B.; Serra, C.; et al. Machine Learning in Neurosurgery: A Global Survey. Acta Neurochir. 2020, 162, 3081–3091.
[CrossRef] [PubMed]
57. Azimi, P.; Benzel, E.C.; Shahzadi, S.; Azhari, S.; Mohammadi, H.R. Use of Artificial Neural Networks to Predict Surgical
Satisfaction in Patients With Lumbar Spinal Canal Stenosis. J. Neurosurg. Spine 2014, 20, 300–305. [CrossRef] [PubMed]
58. Hoffman, H.; Lee, S.I.; Garst, J.H.; Lu, D.S.; Li, C.H.; Nagasawa, D.T.; Ghalehsari, N.; Jahanforouz, N.; Razaghy, M.;
Espinal, M.; et al. Use of Multivariate Linear Regression and Support Vector Regression to Predict Functional Outcome After
Surgery for Cervical Spondylotic Myelopathy. J. Clin. Neurosci. 2015, 22, 1444–1449. [CrossRef] [PubMed]
59. Shamim, M.S.; Enam, S.A.; Qidwai, U. Fuzzy Logic in Neurosurgery: Predicting Poor Outcomes After Lumbar Disk Surgery in
501 Consecutive Patients. Surg. Neurol. 2009, 72, 565–572. [CrossRef]
60. Azimi, P.; Benzel, E.C.; Shahzadi, S.; Azhari, S.; Zali, A.R. Prediction of Successful Surgery Outcome in Lumbar Disc Herniation
Based on Artificial Neural Networks. Glob. Spine J. 2014, 4. [CrossRef]
61. Azimi, P.; Mohammadi, H.R.; Benzel, E.C.; Shahzadi, S.; Azhari, S. Use of Artificial Neural Networks to Predict Recurrent Lumbar
Disk Herniation. J. Spinal Disord. Tech. 2015, 28, E161–E165. [CrossRef]
62. Fatima, N.; Zheng, H.; Massaad, E.; Hadzipasic, M.; Shankar, G.M.; Shin, J.H. Development and Validation of Machine Learning
Algorithms for Predicting Adverse Events After Surgery for Lumbar Degenerative Spondylolisthesis. World Neurosurg. 2020, 140,
627–641. [CrossRef] [PubMed]
63. Karhade, A.V.; Thio, Q.C.B.S.; Ogink, P.T.; Shah, A.A.; Bono, C.M.; Oh, K.S.; Saylor, P.J.; Schoenfeld, A.J.; Shin, J.H.;
Harris, M.B.; et al. Development of Machine Learning Algorithms for Prediction of 30-Day Mortality After Surgery for Spinal
Metastasis. Neurosurgery 2018, 85, E83–E91. [CrossRef]
64. Ames, C.P.; Smith, J.S.; Pellisé, F.; Kelly, M.; Alanay, A.; Acaroğlu, E.; Pérez-Grueso, F.J.S.; Kleinstück, F.; Obeid, I.;
Vila-Casademunt, A.; et al. Artificial Intelligence Based Hierarchical Clustering of Patient Types and Intervention Categories in
Adult Spinal Deformity Surgery. Spine 2019, 44, 915–926. [CrossRef]
65. Xia, X.-P.; Chen, H.-L.; Cheng, H.-B. Prevalence of Adjacent Segment Degeneration After Spine Surgery. Spine 2013, 38, 597–608.
[CrossRef]
66. Wang, F.; Hou, H.-T.; Wang, P.; Zhang, J.-T.; Shen, Y. Symptomatic Adjacent Segment Disease After Single-Lever Anterior Cervical
Discectomy and Fusion. Medicine 2017, 96, e8663. [CrossRef]
Diagnostics 2023, 13, 2429 32 of 33

67. Zhang, J.T.; Cao, J.M.; Meng, F.T.; Shen, Y. Cervical Canal Stenosis and Adjacent Segment Degeneration After Anterior Cervical
Arthrodesis. Eur. Spine J. 2015, 24, 1590–1596. [CrossRef] [PubMed]
68. Kong, L.; Cao, J.; Wang, L.; Shen, Y. Prevalence of Adjacent Segment Disease Following Cervical Spine Surgery. Medicine 2016, 95,
e4171. [CrossRef]
69. Yang, X.; Bartels, R.H.M.A.; Donk, R.; Arts, M.P.; Goedmakers, C.M.W.; Vleggeert-Lankamp, C.L.A. The Association of Cervical
Sagittal Alignment With Adjacent Segment Degeneration. Eur. Spine J. 2019, 29, 2655–2664. [CrossRef] [PubMed]
70. Goedmakers, C.M.W.; Lak, A.M.; Duey, A.H.; Senko, A.W.; Arnaout, O.; Groff, M.W.; Smith, T.R.; Vleggeert-Lankamp, C.L.A.;
Zaidi, H.A.; Rana, A.; et al. Deep Learning for Adjacent Segment Disease at Preoperative MRI for Cervical Radiculopathy.
Radiology 2021, 301, 664–671. [CrossRef] [PubMed]
71. Karhade, A.V.; Bongers, M.E.; Groot, O.Q.; Kazarian, E.R.; Cha, T.D.; Fogel, H.A.; Hershman, S.H.; Tobert, D.G.; Schoenfeld, A.J.;
Bono, C.M.; et al. Natural Language Processing for Automated Detection of Incidental Durotomy. Spine J. 2020, 20, 695–700.
[CrossRef] [PubMed]
72. Karhade, A.V.; Bongers, M.E.; Groot, O.Q.; Cha, T.D.; Doorly, T.P.; Fogel, H.A.; Hershman, S.H.; Tobert, D.G.; Srivastava, S.D.;
Bono, C.M.; et al. Development of Machine Learning and Natural Language Processing Algorithms for Preoperative Prediction
and Automated Identification of Intraoperative Vascular Injury in Anterior Lumbar Spine Surgery. Spine J. 2021, 21, 1635–1642.
[CrossRef] [PubMed]
73. Benyamin, R.; Trescot, A.M.; Datta, S.; Buenaventura, R.; Adlaka, R.; Sehgal, N.; Glaser, S.E.; Vallejo, R. Opioid Complications and
Side Effects. Pain. Physician 2008, 11, S105–S120. [CrossRef]
74. Schofferman, J. Long-Term Use of Opioid Analgesics for the Treatment of Chronic Pain of Nonmalignant Origin. J. Pain. Symptom
Manag. 1993, 8, 279–288. [CrossRef] [PubMed]
75. Schofferman, J. Long-Term Opioid Analgesic Therapy for Severe Refractory Lumbar Spine Pain. Clin. J. Pain. 1999, 15, 136–140.
[CrossRef]
76. Bartleson, J.D. Evidence For and Against the Use of Opioid Analgesics for Chronic Nonmalignant Low Back Pain: A Review:
Table 1. Pain. Med. 2002, 3, 260–271. [CrossRef] [PubMed]
77. Jamison, R.N.; Raymond, S.A.; Slawsby, E.A.; Nedeljkovic, S.S.; Katz, N.P. Opioid Therapy for Chronic Noncancer Back Pain.
Spine 1998, 23, 2591–2600. [CrossRef]
78. Paulozzi, L.J.; Budnitz, D.S.; Xi, Y. Increasing Deaths from Opioid Analgesics in the United States. Pharmacoepidemiol. Drug. Saf.
2006, 15, 618–627. [CrossRef]
79. Karhade, A.V.; Ogink, P.T.; Thio, Q.C.; Cha, T.D.; Gormley, W.B.; Hershman, S.H.; Smith, T.R.; Mao, J.; Schoenfeld, A.J.;
Bono, C.M.; et al. Development of Machine Learning Algorithms for Prediction of Prolonged Opioid Prescription After Surgery
for Lumbar Disc Herniation. Spine J. 2019, 19, 1764–1771. [CrossRef] [PubMed]
80. Stopa, B.M.; Robertson, F.C.; Karhade, A.V.; Chua, M.; Broekman, M.L.D.; Schwab, J.H.; Smith, T.R.; Gormley, W.B. Predicting
Nonroutine Discharge After Elective Spine Surgery: External Validation of Machine Learning Algorithms. J. Neurosurg. Spine
2019, 31, 742–747. [CrossRef]
81. Huang, K.T.; Silva, M.A.; See, A.P.; Wu, K.C.; Gallerani, T.; Zaidi, H.A.; Lu, Y.; Chi, J.H.; Groff, M.W.; Arnaout, O.M. A computer
Vision Approach to Identifying the Manufacturer and Model of Anterior Cervical Spinal Hardware. J. Neurosurg. Spine 2019, 31,
844–850. [CrossRef]
82. Grigsby, J.; Kramer, R.E.; Schneiders, J.L.; Gates, J.R.; Smith, W.B. Predicting Outcome of Anterior Temporal Lobectomy Using
Simulated Neural Networks. Epilepsia 1998, 39, 61–66. [CrossRef] [PubMed]
83. Antony, A.R.; Alexopoulos, A.V.; González-Martínez, J.A.; Mosher, J.C.; Jehi, L.; Burgess, R.C.; So, N.K.; Galán, R.F. Functional
Connectivity Estimated from Intracranial EEG Predicts Surgical Outcome in Intractable Temporal Lobe Epilepsy. PLoS ONE 2013,
8, e77916. [CrossRef] [PubMed]
84. Arle, J.E.; Perrine, K.; Devinsky, O.; Doyle, W.K. Neural Network Analysis of Preoperative Variables and Outcome in Epilepsy
Surgery. J. Neurosurg. 1999, 90, 998–1004. [CrossRef]
85. Armañanzas, R.; Alonso-Nanclares, L.; DeFelipe-Oroquieta, J.; Kastanauskaite, A.; de Sola, R.G.; DeFelipe, J.; Bielza, C.;
Larrañaga, P. Machine Learning Approach for the Outcome Prediction of Temporal Lobe Epilepsy Surgery. PLoS ONE 2013, 8,
e62819. [CrossRef] [PubMed]
86. Bernhardt, B.C.; Hong, S.-J.; Bernasconi, A.; Bernasconi, N. Magnetic Resonance Imaging Pattern Learning in Temporal Lobe
Epilepsy: Classification and Prognostics. Ann. Neurol. 2015, 77, 436–446. [CrossRef] [PubMed]
87. Feis, D.-L.; Schoene-Bake, J.-C.; Elger, C.; Wagner, J.; Tittgemeyer, M.; Weber, B. Prediction of Post-Surgical Seizure Outcome in
Left Mesial Temporal Lobe Epilepsy. NeuroImage Clin. 2013, 2, 903–911. [CrossRef]
88. Njiwa, J.Y.; Gray, K.; Costes, N.; Mauguiere, F.; Ryvlin, P.; Hammers, A. Advanced [18F]FDG and [11C]flumazenil PET Analysis
For Individual Outcome Prediction After Temporal Lobe Epilepsy Surgery for Hippocampal Sclerosis. NeuroImage Clin. 2014, 7,
122–131. [CrossRef] [PubMed]
89. Memarian, N.; Kim, S.; Dewar, S.; Engel, J.; Staba, R.J. Multimodal Data and Machine Learning for Surgery Outcome Prediction
In Complicated Cases of Mesial Temporal Lobe Epilepsy. Comput. Biol. Med. 2015, 64, 67–78. [CrossRef]
90. Torlay, L.; Perrone-Bertolotti, M.; Thomas, E.; Baciu, M. Machine Learning–Xgboost Analysis of Language Networks to Classify
Patients with Epilepsy. Brain Inform. 2017, 4, 159–169. [CrossRef]
91. Abbasi, B.; Goldenholz, D.M. Machine Learning Applications in Epilepsy. Epilepsia 2019, 60, 2037–2047. [CrossRef]
Diagnostics 2023, 13, 2429 33 of 33

92. A Reinforcement Learning-Based Framework for the Generation and Evolution of Adaptation Rules. 2017 IEEE International
Conference on Autonomic Computing (ICAC), Columbus, OH, USA, 17–21 July 2017. [CrossRef]
93. Wiebe, S.; Blume, W.T.; Girvin, J.P.; Eliasziw, M.; Effectiveness and Efficiency of Surgery for Temporal Lobe Epilepsy Study Group.
A Randomized, Controlled Trial of Surgery for Temporal-Lobe Epilepsy. N. Engl. J. Med. 2001, 345, 311–318. [CrossRef]
94. Larivière, S.; Weng, Y.; De Wael, R.V.; Royer, J.; Frauscher, B.; Wang, Z.; Bernasconi, A.; Bernasconi, N.; Schrader, D.; Zhang, Z.; et al.
Functional Connectome Contractions in Temporal Lobe Epilepsy: Microstructural Underpinnings and Predictors Of Surgical
Outcome. Epilepsia 2020, 61, 1221–1233. [CrossRef] [PubMed]
95. Faron, A.; Sichtermann, T.; Teichert, N.; Luetkens, J.A.; Keulers, A.; Nikoubashman, O.; Freiherr, J.; Mpotsaris, A.; Wiesmann, M.
Performance of a Deep-Learning Neural Network to Detect Intracranial Aneurysms from 3D TOF-MRA Compared to Human
Readers. Clin. Neuroradiol. 2019, 30, 591–598. [CrossRef] [PubMed]
96. Zhu, W.; Li, W.; Tian, Z.; Zhang, Y.; Wang, K.; Zhang, Y.; Liu, J.; Yang, X. Stability Assessment of Intracranial Aneurysms Using
Machine Learning Based on Clinical and Morphological Features. Transl. Stroke Res. 2020, 11, 1287–1295. [CrossRef] [PubMed]
97. Park, A.; Chute, C.; Rajpurkar, P.; Lou, J.; Ball, R.L.; Shpanskaya, K.; Jabarkheel, R.; Kim, L.H.; McKenna, E.; Tseng, J.; et al. Deep
Learning–Assisted Diagnosis of Cerebral Aneurysms Using the HeadXNet Model. JAMA Netw. Open. 2019, 2, e195600. [CrossRef]
98. Silva, M.A.; Patel, J.; Kavouridis, V.; Gallerani, T.; Beers, A.; Chang, K.; Hoebel, K.V.; Brown, J.; See, A.P.; Gormley, W.B.; et al.
Machine Learning Models can Detect Aneurysm Rupture and Identify Clinical Features Associated with Rupture. World Neurosurg.
2019, 131, e46–e51. [CrossRef]
99. Sahlein, D.H.; Gibson, D.; Scott, J.A.; De Nardo, A.; Amuluru, K.; Payner, T.; Rosenbaum-Halevi, D.; Kulwin, C. Artificial
Intelligence Aneurysm Measurement Tool Finds Growth in all Aneurysms that Ruptured During Conservative Management. J.
NeuroInterventional Surg. 2022. [CrossRef] [PubMed]
100. Wiebers, D.O. Unruptured intracranial aneurysms: Natural History, Clinical Outcome, And Risks of Surgical and Endovascular
Treatment. Lancet 2003, 362, 103–110. [CrossRef]
101. Brunozzi, D.; Theiss, P.; Andrews, A.; Amin-Hanjani, S.; Charbel, F.T.; Alaraj, A. Correlation Between Laminar Wall Shear Stress
and Growth of Unruptured Cerebral Aneurysms: In Vivo Assessment. World Neurosurg. 2019, 131, e599–e605. [CrossRef]
102. Nomura, S.; Kunitsugu, I.; Ishihara, H.; Koizumi, H.; Yoneda, H.; Shirao, S.; Oka, F.; Suzuki, M. Relationship between Aging and
Enlargement of Intracranial Aneurysms. J. Stroke Cerebrovasc. Dis. 2015, 24, 2049–2053. [CrossRef] [PubMed]
103. Liu, Q.; Jiang, P.; Jiang, Y.; Ge, H.; Li, S.; Jin, H.; Li, Y. Prediction of Aneurysm Stability Using a Machine Learning Model Based on
PyRadiomics-Derived Morphological Features. Stroke 2019, 50, 2314–2321. [CrossRef]
104. Hop, J.W.; Rinkel, G.J.; Algra, A.; van Gijn, J. Case-Fatality Rates and Functional Outcome After Subarachnoid Hemorrhage.
Stroke 1997, 28, 660–664. [CrossRef] [PubMed]
105. Nieuwkamp, D.J.; Setz, L.E.; Algra, A.; Linn, F.H.; de Rooij, N.K.; Rinkel, G.J. Changes in Case Fatality of Aneurysmal
Subarachnoid Haemorrhage over Time, According to Age, Sex, and Region: A Meta-Analysis. Lancet Neurol. 2009, 8, 635–642.
[CrossRef] [PubMed]
106. Al-Khindi, T.; Macdonald, R.L.; Schweizer, T.A. Cognitive and Functional Outcome After Aneurysmal Subarachnoid Hemorrhage.
Stroke 2010, 41, e519–36. [CrossRef]
107. Roos, Y.B.W.E.M.; Zarranz, J.J.; Rouco, I.; Gómez-Esteban, J.C.; Corral, J. Complications and Outcome in Patients With Aneurysmal
Subarachnoid Haemorrhage: A Prospective Hospital Based Cohort Study in the Netherlands. J. Neurol. Neurosurg. Psychiatry
2000, 68, 337–341. [CrossRef]
108. Koch, M.; Acharjee, A.; Ament, Z.; Schleicher, R.; Bevers, M.; Stapleton, C.; Patel, A.; Kimberly, W.T. Machine Learning-Driven
Metabolomic Evaluation of Cerebrospinal Fluid: Insights Into Poor Outcomes After Aneurysmal Subarachnoid Hemorrhage.
Neurosurgery 2021, 88, 1003–1011. [CrossRef] [PubMed]
109. Vergouwen, M.D.; Vermeulen, M.; van Gijn, J.; Rinkel, G.J.; Wijdicks, E.F.; Muizelaar, J.P.; Mendelow, A.D.; Juvela, S.; Yonas, H.;
Terbrugge, K.G.; et al. Definition of Delayed Cerebral Ischemia After Aneurysmal Subarachnoid Hemorrhage as an Outcome
Event in Clinical Trials and Observational Studies. Stroke 2010, 41, 2391–2395. [CrossRef]
110. Ramos, L.A.; Van Der Steen, W.E.; Barros, R.S.; Majoie, C.B.L.M.; Berg, R.V.D.; Verbaan, D.; Vandertop, W.P.; Zijlstra, I.J.A.J.;
Zwinderman, A.H.; Strijkers, G.; et al. Machine Learning Improves Prediction of Delayed Cerebral Ischemia in Patients With
Subarachnoid Hemorrhage. J. NeuroInterventional Surg. 2018, 11, 497–502. [CrossRef]
111. Asadi, H.; Kok, H.K.; Looby, S.; Brennan, P.; O’Hare, A.; Thornton, J. Outcomes and Complications After Endovascular Treatment
of Brain Arteriovenous Malformations: A Prognostication Attempt Using Artificial Intelligence. World Neurosurg. 2016, 96,
562–569.e1. [CrossRef]
112. Gonzalez-Romo, N.I.; Hanalioglu, S.; Mignucci-Jiménez, G.; Koskay, G.; Abramov, I.; Xu, Y.; Park, W.; Lawton, M.T.; Preul, M.C.
Quantification of Motion During Microvascular Anastomosis Simulation Using Machine Learning Hand Detection. Neurosurg.
Focus. 2023, 54, E2. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.