Open Access Review

Artificial intelligence in healthcare: past,

present and future
Fei Jiang,1 Yong Jiang,2 Hui Zhi,3 Yi Dong,4 Hao Li,5 Sufeng Ma,6 Yilong Wang,7
Qiang Dong,4 Haipeng Shen,8 Yongjun Wang9

To cite: Jiang F, Jiang Y, Zhi H, Abstract 3. mechanisms that enable AI systems to gen-
et al. Artificial intelligence in Artificial intelligence (AI) aims to mimic human cognitive erate clinical meaningful results
healthcare: past, present and functions. It is bringing a paradigm shift to healthcare,
future. Stroke and Vascular
4. disease types that the AI communities are
powered by increasing availability of healthcare data and currently tackling.
Neurology 2017;2: e000101.
doi:10.1136/svn-2017-000101 rapid progress of analytics techniques. We survey the
current status of AI applications in healthcare and discuss Motivation
its future. AI can be applied to various types of healthcare The advantages of AI have been extensively
Received 12 June 2017 data (structured and unstructured). Popular AI techniques discussed in the medical literature.3–5 AI
Accepted 14 June 2017 include machine learning methods for structured data,
Published Online First
can use sophisticated algorithms to ‘learn’
such as the classical support vector machine and neural features from a large volume of healthcare
22 June 2017
network, and the modern deep learning, as well as data, and then use the obtained insights to
natural language processing for unstructured data. Major
assist clinical practice. It can also be equipped
disease areas that use AI tools include cancer, neurology
with learning and self-correcting abilities to
and cardiology. We then review in more details the AI
applications in stroke, in the three major areas of early
improve its accuracy based on feedback. An
detection and diagnosis, treatment, as well as outcome AI system can assist physicians by providing
prediction and prognosis evaluation. We conclude with up-to-date medical information from jour-
discussion about pioneer AI systems, such as IBM Watson, nals, textbooks and clinical practices to
and hurdles for real-life deployment of AI. inform proper patient care.6 In addition, an
AI system can help to reduce diagnostic and
1
Department of Statistics and therapeutic errors that are inevitable in the
Actuarial Sciences, University of Overview of the medical artificial human clinical practice.3 4 6–10 Moreover, an
Hong Kong, Hong Kong, China
2
intelligence (AI) research AI system extracts useful information from
Department of Neurology,
Recently AI techniques have sent vast waves a large patient population to assist making
Beijing Tiantan Hospital, Capital
Medical University, Beijing, across healthcare, even fuelling an active real-time inferences for health risk alert and
China discussion of whether AI doctors will eventu- health outcome prediction.11
3
Biostatistics and Clinical ally replace human physicians in the future.
Research Methodology Unit, We believe that human physicians will not
University of Hong Kong Li Ka Healthcare data
be replaced by machines in the foreseeable
Shing Faculty of Medicine, Hong Before AI systems can be deployed in health-
Kong, China future, but AI can definitely assist physicians to care applications, they need to be ‘trained’
4
Department of Neurology, make better clinical decisions or even replace through data that are generated from clin-
Huashan Hospital, Fudan human judgement in certain functional areas
University, Shanghai, China ical activities, such as screening, diagnosis,
5 of healthcare (eg, radiology). The increasing treatment assignment and so on, so that they
China National Clinical
Research Center for availability of healthcare data and rapid devel- can learn similar groups of subjects, associa-
Neurological Diseases, Beijing, opment of big data analytic methods has tions between subject features and outcomes
China made possible the recent successful applica- of interest. These clinical data often exist in
6
DotHealth, Shanghai, China tions of AI in healthcare. Guided by relevant
7
Department of Neurology, but not limited to the form of demographics,
Tiantan Clinical Trial and
clinical questions, powerful AI techniques can medical notes, electronic recordings from
Research Center for Stroke, unlock clinically relevant information hidden medical devices, physical examinations and
Beijing, China in the massive amount of data, which in turn clinical laboratory and images.12
8
Faculty of Business and can assist clinical decision making.1–3 Specifically, in the diagnosis stage, a substan-
Economics, University of Hong
In this article, we survey the current status tial proportion of the AI literature analyses
Kong, Hong Kong, China
9
Department of Neurology, of AI in healthcare, as well as discuss its future. data from diagnosis imaging, genetic testing
Beijing Tiantan Hospital, Beijing, We first briefly review four relevant aspects and electrodiagnosis (figure 1). For example,
China from medical investigators’ perspectives: Jha and Topol urged radiologists to adopt
1. motivations of applying AI in healthcare AI technologies when analysing diagnostic
Correspondence to
Prof Yongjun Wang; 2. data types that have be analysed by AI sys- images that contain vast data information.13
​yongjunwang1962@​gmail.​com tems Li et al studied the uses of abnormal genetic

230 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

Figure 1 The data types considered in the artificial intelligence artificial (AI) literature. The comparison is obtained through
searching the diagnosis techniques in the AI literature on the PubMed database.

expression in long non-coding RNAs to diagnose gastric starts and ends with clinical activities. As powerful as AI
cancer.14 Shin et al developed an electrodiagnosis support techniques can be, they have to be motivated by clinical
system for localising neural injury.15 problems and be applied to assist clinical practice in the
In addition, physical examination notes and clinical end.
laboratory results are the other two major data sources
(figure 1). We distinguish them with image, genetic and Disease focus
electrophysiological (EP) data because they contain large Despite the increasingly rich AI literature in healthcare,
portions of unstructured narrative texts, such as clin- the research mainly concentrates around a few disease
ical notes, that are not directly analysable. As a conse- types: cancer, nervous system disease and cardiovascular
quence, the corresponding AI applications focus on first disease (figure 3). We discuss several examples below.
converting the unstructured text to machine-understand- 1. Cancer: Somashekhar et al demonstrated that the IBM
able electronic medical record (EMR). For example, Watson for oncology would be a reliable AI system for
Karakülah et al used AI technologies to extract pheno- assisting the diagnosis of cancer through a double-
typic features from case reports to enhance the diagnosis blinded validation study.19 Esteva et al analysed clinical
accuracy of the congenital anomalies.16 images to identify skin cancer subtypes.20
2. Neurology: Bouton et al developed an AI system to
AI devices restore the control of movement in patients with
The above discussion suggests that AI devices mainly fall quadriplegia.21 Farina et al tested the power of an of-
into two major categories. The first category includes fline man/machine interface that uses the discharge
machine learning (ML) techniques that analyse struc- timings of spinal motor neurons to control upper-limb
tured data such as imaging, genetic and EP data. In prostheses.22
the medical applications, the ML procedures attempt 3. Cardiology: Dilsizian and Siegel discussed the
to cluster patients’ traits, or infer the probability of the potential application of the AI system to diagnose
disease outcomes.17 The second category includes natural the heart disease through cardiac image.3 Arterys
language processing (NLP) methods that extract infor- recently received clearance from the US Food and
mation from unstructured data such as clinical notes/ Drug Administration (FDA) to market its Arterys
medical journals to supplement and enrich structured Cardio DL application, which uses AI to provide
medical data. The NLP procedures target at turning texts automated, editable ventricle segmentations based on
to machine-readable structured data, which can then be conventional cardiac MRI images.23
analysed by ML techniques.18 The concentration around these three diseases is not
For better presentation, the flow chart in figure 2 completely unexpected. All three diseases are leading
describes the road map from clinical data generation, causes of death; therefore, early diagnoses are crucial
through NLP data enrichment and ML data analysis, to to prevent the deterioration of patients’ health status.
clinical decision making. We comment that the road map Furthermore, early diagnoses can be potentially achieved

231
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101
Open Access

Figure 2 The road map from clinical data generation to natural language processing data enrichment, to machine learning
data analysis, to clinical decision making. EMR, electronic medical record; EP, electrophysiological.

through improving the analysis procedures on imaging, We then conclude in section 4 with some discussion about
genetic, EP or EMR, which is the strength of the AI system. the future of AI in healthcare.
Besides the three major diseases, AI has been applied
in other diseases as well. Two very recent examples were
Long et al, who analysed the ocular image data to diag- The AI devices: ML and NLP
nose congenital cataract disease,24 and Gulshan et al, In this section, we review the AI devices (or techniques)
who detected referable diabetic retinopathy through the that have been found useful in the medial applications.
retinal fundus photographs.25 We categorise them into three groups: the classical
The rest of the paper is organised as follows. In section machine learning techniques,26 the more recent deep
2, we describe popular AI devices in ML and NLP; the learning techniques27 and the NLP methods.28
ML techniques are further grouped into classical tech-
niques and the more recent deep learning. Section 3 Classical ML
focuses on discussing AI applications in neurology, from ML constructs data analytical algorithms to extract
the three aspects of early disease prediction and diagnosis, features from data. Inputs to ML algorithms include
treatment, outcome prediction and prognosis evaluation. patient ‘traits’ and sometimes medical outcomes of

232 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

Figure 3 The leading 10 disease types considered in the artificial intelligence (AI) literature. The first vocabularies in the
disease names are displayed. The comparison is obtained through searching the disease types in the AI literature on PubMed.

interest. A patient’s traits commonly include baseline recently, semisupervised learning has been proposed
data, such as age, gender, disease history and so on, and as a hybrid between unsupervised learning and super-
disease-specific data, such as diagnostic imaging, gene vised learning, which is suitable for scenarios where the
expressions, EP test, physical examination results, clin- outcome is missing for certain subjects. These three types
ical symptoms, medication and so on. Besides the traits, of learning are illustrated in figure 4.
patients’ medical outcomes are often collected in clin- Clustering and principal component analysis (PCA)
ical research. These include disease indicators, patient’s are two major unsupervised learning methods. Clustering
survival times and quantitative disease levels, for example, groups subjects with similar traits together into clusters,
tumour sizes. To fix ideas, we denote the jth trait of the ith without using the outcome information. Clustering algo-
patient by Xij, and the outcome of interest by Yi. rithms output the cluster labels for the patients through
Depending on whether to incorporate the outcomes, maximising and minimising the similarity of the patients
ML algorithms can be divided into two major categories: within and between the clusters. Popular clustering algo-
unsupervised learning and supervised learning. Unsuper- rithms include k-means clustering, hierarchical clustering
vised learning is well known for feature extraction, while and Gaussian mixture clustering. PCA is mainly for dimen-
supervised learning is suitable for predictive modelling sion reduction, especially when the trait is recorded in a
via building some relationships between the patient traits large number of dimensions, such as the number of genes
(as input) and the outcome of interest (as output). More in a genome-wide association study. PCA projects the data

Figure 4 Graphical illustration of unsupervised learning, supervised learning and semisupervised learning.

233
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101
Open Access

Figure 5 The machine learning algorithms used in the medical literature. The data are generated through searching the
machine learning algorithms within healthcare on PubMed.

onto a few principal component (PC) directions, without 1 represents whether the ith patient is in group 1 or 2,
losing too much information about the subjects. Some- respectively. (The method can be extended for scenarios
times, one can first use PCA to reduce the dimension of with more than two groups.) The basic assumption is that
the data, and then use clustering to group the subjects. the subjects can be separated into two groups through a
On the other hand, supervised learning considers the decision boundary defined on the traits Xij, which can be
subjects’ outcomes together with their traits, and goes written as:
through a certain training process to determine the best ∑p
ai = j=1 wj Xij + b,
outputs associated with the inputs that are closest to the
where wj is the weight putting on the jth trait to manifest
outcomes on average. Usually, the output formulations
its relative importance on affecting the outcome among
vary with the outcomes of interest. For example, the
the others. The decision rule then follows that if ai >0,
outcome can be the probability of getting a particular
clinical event, the expected value of a disease level or the the ith patient is classified to group 1, that is, labelling Yi
expected survival time. = −1; if ai <0, the patient is classified to group 2, that is,
Clearly, compared with unsupervised learning, super- labelling Yi=1. The class memberships are indeterminate
vised learning provides more clinically relevant results; for the points with ai=0. See figure 7 for an illustration
hence AI applications in healthcare most often use super- with p = 2, b = 0 , a1=1, and a2=−1.
vised learning. (Note that unsupervised learning can be The training goal is to find the optimal wjs so that
used as part of the preprocessing step to reduce dimen- the resulting classifications agree with the outcomes as
sionality or identify subgroups, which in turn makes much as possible, that is, with the smallest misclassifi-
the follow-up supervised learning step more efficient.) cation error, the error of classifying a patient into the
Relevant techniques include linear regression, logistic wrong group. Intuitively, the best weights must allow (1)
regression, naïve Bayes, decision tree, nearest neighbour, the sign of ai to be the same as Yi so the classification is
random forest, discriminant analysis, support vector correct; and (2) |ai| to be far away from 0 so the ambiguity
machine (SVM) and neural network.27 Figure 5 displays of the classification is minimised. These can be achieved
the popularity of the various supervised learning tech- by selecting wjs that minimise a quadratic loss function.29
niques in medical applications, which clearly shows that Furthermore, assuming that the new patients come from
SVM and neural network are the most popular ones. This the same population, the resulting   wjs can be applied to
remains the case when restricting to the three major data classify these new patients based on their traits.
types (image, genetic and EP), as shown in figure 6. An important property of SVM is that the determination
Below we will provide more details about the mechanisms of the model parameters is a convex optimisation problem
of SVM and neural networks, along with application exam- so the solution is always global optimum. Furthermore,
ples in the cancer, neurological and cardiovascular disease many existing convex optimisation tools are readily appli-
areas. cable for the SVM implementation. As such, SVM has been
extensively used in medical research. For instance, Orrù et
Support vector machine al applied SVM to identify imaging biomarkers of neurolog-
SVM is mainly used for classifying the subjects into two ical and psychiatric disease.30 Sweilam et al reviewed the use
groups, where the outcome Yi is a classifier: Yi = −1 or of SVM in the diagnosis of cancer.31 Khedher et al used the

234 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

Figure 6 The machine learning algorithms used for imaging (upper), genetic (middle) and electrophysiological (bottom) data.
The data are generated through searching the machine learning algorithms for each data type on PubMed.

combination of SVM and other statistical tools to achieve Neural network

early detection of Alzheimer’s disease.32 Farina et al used One can think about neural network as an extension of
SVM to test the power of an offline man/machine interface linear regression to capture complex non-linear relation-
that controls upper-limb prostheses.22 ships between input variables and an outcome. In neural

235
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101
Open Access

data so that the average error between the outcome and

their predictions is minimised. We describe the method
in the following example.
Mirtskhulava et al used neural network in stroke diag-
nosis.33 In their analysis, the input variables Xi1, . . . , Xip
are p=16 stroke-related symptoms, including paraesthesia
of the arm or leg, acute confusion, vision, problems with
mobility and so on. The outcome Yi is binary: Yi=1/0 indi-
cates the ith patient has/does not have stroke. The output
parameter of interest is the probability of stroke, ai, which
carries the form{∑of }
D ∑p
ai = h w
k=1 2l kf ( w X
l=1 1l il + w10 ) + w 20 .
In the above equation, the w10 and w20≠0 guarantee the
above form to be valid even when all Xij, fk are 0; the w1l
and w 2ls are the weights to characterise the relative impor-
tance of the corresponding multiplicands on affecting
the outcome; the fks and h are prespecified functionals to
manifest how the weighted combinations influence the
Figure 7 An illustration of the support vector machine. disease risk as a whole. A stylised illustration is provided
in figure 8.
network, the associations between the outcome and the The training goal is to find∑the(weights) wij, which mini-
input variables are depicted through multiple hidden mise the prediction error ni=1 Yi − ai 2. The minimis-
layer combinations of prespecified functionals. The goal ation can be performed through standard optimisation
is to estimate the weights through input and outcome algorithms, such as local quadratic approximation or

Figure 8 An illustration of neural network.

236 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

Figure 9 An illustration of deep learning with two hidden layers.

gradient descent optimisation, that are included in both Different from the classical neural network, deep
MATLAB and R. If the new data come from the same learning uses more hidden layers so that the algorithms
population, the resulting wij can be used to predict the can handle complex data with various structures.27 In the
outcomes based on their specific traits.29 medical applications, the commonly used deep learning
Similar techniques have been used to diagnose cancer algorithms include convolution neural network (CNN),
by Khan et al, where the inputs are the PCs estimated from recurrent neural network, deep belief network and deep
6567 genes and the outcomes are the tumour catego- neural network. Figure 12 depicts their trends and rela-
ries.34 Dheeba et al used neural network to predict breast tive popularities from 2013 to 2016. One can see that the
cancer, with the inputs being the texture information CNN is the most popular one in 2016.
from mammographic images and the outcomes being The CNN is developed in viewing of the incompetence
tumour indicators.35 Hirschauer et al used a more sophis- of the classical ML algorithms when handling high dimen-
ticated neural network model to diagnose Parkinson’s sional data, that is, data with a large number of traits. Tradi-
disease based on the inputs of motor, non-motor symp- tionally, the ML algorithms are designed to analyse data
toms and neuroimages.36 when the number of traits is small. However, the image
data are naturally high-dimensional because each image
Deep learning: a new era of ML
normally contains thousands of pixels as traits. One solu-
Deep learning is a modern extension of the classical
neural network technique. One can view deep learning tion is to perform dimension reduction: first preselect
as a neural network with many layers (as in figure 9). a subset of pixels as features, and then perform the ML
Rapid development of modern computing enables deep algorithms on the resulting lower dimensional features.
learning to build up neural networks with a large number However, heuristic feature selection procedures may lose
of layers, which is infeasible for classical neural networks. information in the images. Unsupervised learning tech-
As such, deep learning can explore more complex niques such as PCA or clustering can be used for data-
non-linear patterns in the data. Another reason for the driven dimension reduction.
recent popularity of deep learning is due to the increase of The CNN was first proposed and advocated for the
the volume and complexity of data.37 Figure 10 shows that high-dimensional image analysis by Lecun et al.38 The
the application of deep learning in the field of medical inputs for CNN are the properly normalised pixel values
research nearly doubled in 2016. In addition, figure 11 on the images. The CNN then transfers the pixel values
shows that a clear majority of deep learning is used in in the image through weighting in the convolution layers
imaging analysis, which makes sense given that images are and sampling in the subsampling layers alternatively.
naturally complex and high volume. The final output is a recursive function of the weighted

237
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101
Open Access

Figure 10 Current trend for deep learning. The data are generated through searching the deep learning in healthcare and
disease category on PubMed.

input values. The weights are trained to minimise the et al performed the CNN to identify skin cancer from
average error between the outcomes and the predic- clinical images.20 The proportions of correctly predicted
tions. The implementation of CNN has been included in malignant lesions (ie, sensitivity) and benign lesions (ie,
popular software packages such as Caffe from Berkeley AI specificity) are both over 90%, which indicates the supe-
Research,39 CNTK from Microsoft40 and TensorFlow from rior performance of the CNN. Gulshan et al applied the
Google.41 CNN to detect referable diabetic retinopathy through
Recently, the CNN has been successfully implemented the retinal fundus photographs.25 The sensitivity and
in the medical area to assist disease diagnosis. Long et al specificity of the algorithm are both over 90%, which
used it to diagnose congenital cataract disease through demonstrates the effectiveness of using the technique
learning the ocular images.24 The CNN yields over 90% on the diagnosis of diabetes. It is worth mentioning that
accuracy on diagnosis and treatment suggestion. Esteva in all these applications, the performance of the CNN is

Figure 11 The data sources for deep learning. The data are generated through searching deep learning in combination with
the diagnosis techniques on PubMed.

238 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

Figure 12 The four main deep learning algorithm and their popularities. The data are generated through searching algorithm
names in healthcare and disease category on PubMed.

competitive against experienced physicians in the accu- disease-associated variables through implementing NLP
racy for classifying both normal and disease cases. on the clinical notes.45 The resulting variables are success-
fully used for classifying the normal patients and the
Natural language processing patients with cerebral, with 95% and 86% accuracy rates
The image, EP and genetic data are machine-understand- on the training and validation samples, respectively. Afzal
able so that the ML algorithms can be directly performed et al implemented the NLP to extract the peripheral arte-
after proper preprocessing or quality control processes. rial disease-related keywords from narrative clinical notes.
However, large proportions of clinical information are in The keywords are then used to classify the normal and the
the form of narrative text, such as physical examination, patients with peripheral arterial disease, which achieves
clinical laboratory reports, operative notes and discharge over 90% accuracy.42
summaries, which are unstructured and incomprehen-
sible for the computer program. Under this context, NLP
targets at extracting useful information from the narra- AI applications in stroke
tive text to assist clinical decision making.28 Stroke is a common and frequently occurring disease that
An NLP pipeline comprises two main components: affects more than 500 million people worldwide. It is the
(1) text processing and (2) classification. Through text leading cause of death in China and the fifth in North
processing, the NLP identifies a series of disease-relevant America. Stroke had cost about US$689 billion in medical
keywords in the clinical notes based on the historical expenses across the world, causing heavy burden to coun-
databases.42 Then a subset of the keywords are selected tries and families.46 47 Therefore, research on prevention
through examining their effects on the classification of and treatment for stroke has great significance. In recent
the normal and abnormal cases. The validated keywords years, AI techniques have been used in more and more
then enter and enrich the structured data to support clin- stroke-related studies. Below we summarise some of the
ical decision making. relevant AI techniques in the three main areas of stroke
The NLP pipelines have been developed to assist clin- care: early disease prediction and diagnosis, treatment,
ical decision making on alerting treatment arrangements, as well as outcome prediction and prognosis evaluation.
monitoring adverse effects and so on. For example,
Fiszman et al showed that introducing NLP for reading Early detection and diagnosis
the chest X-ray reports would assist the antibiotic assistant Stroke, for 85% of the time, is caused by thrombus in
system to alert physicians for the possible need for anti-in- the vessel called cerebral infarction. However, for lack of
fective therapy.43 Miller et al used NLP to automatically judgement of early stroke symptom, only a few patients
monitor the laboratory-based adverse effects.44 Further- could receive timely treatment. Villar et al developed a
more, the NLP pipelines can help with disease diagnosis. movement-detecting device for early stroke prediction.48
For instance, Castro et al identified 14 cerebral aneurysms Two ML algorithms — genetic fuzzy finite state machine

239
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101
Open Access

and PCA — were implemented into the device for the and subgroup analysis to explore appropriate tPA dosage
model building solution. The detection process included based on patient characteristics, taking into account both
a human activity recognition stage and a stroke-onset the risk of bleeding and the treatment efficacy.57
detection stage. Once the movement of the patient is
significantly different from the normal pattern, an alert Outcome prediction and prognosis evaluation
of stroke would be activated and evaluated for treatment Many factors can affect stroke prognosis and disease
as soon as possible. Similarly, Maninini et al proposed a mortality. Compared with conventional methods, ML
wearable device for collecting data about normal/patho- methods have advantages in improving prediction perfor-
logical gaits for stroke prediction.49 The data would be mance. To better support clinical decision-making process,
extracted and modelled by hidden Markov models and Zhang et al proposed a model for predicting 3-month
SVM, and the algorithm could correctly classify 90.5% of treatment outcome by analysing physiological parameters
the subjects to the right group. during 48 hours after stroke using logistic regression.58
For diagnosis of stroke, neuroimaging techniques, Asadi et al compiled a database of clinical information of
including MRI and CT, are important for disease evalu- 107 patients with acute anterior or posterior circulation
ation. Some studies have tried to apply ML methods to stroke who underwent intra-arterial therapy.59 The authors
neuroimaging data to assist with stroke diagnosis. Rehme analysed the data via artificial neural network and SVM, and
et al used SVM in resting-state functional MRI data, by obtained prediction accuracy above 70%. They also used
which endophenotypes of motor disability after stroke ML techniques to identify factors influencing outcome
were identified and classified.50 SVM can correctly clas- in brain arteriovenous malformation treated with endo-
sify patients with stroke with 87.6% accuracy. Griffis vascular embolisation.60 While standard regression anal-
et al tried naïve Bayes classification to identify stroke ysis model could only achieve a 43% accuracy rate, their
lesion in T1-weighted MRI.51 The result is comparable methods worked much better with 97.5% accuracy.
with human expert manual lesion delineation. Kamnitsas Birkner et al used an optimal algorithm to predict 30-day
et al tried three-dimensional CNN (3D CNN) for lesion mortality and obtained more accurate prediction than
segmentation in multimodel brain MRI.52 They also used existing methods.61 Similarly, King et al used SVM to predict
fully connected conditional random field model for final stroke mortality at discharge.62 In addition, they proposed
postprocessing of the CNN’s soft segmentation maps. the use of the synthetic minority oversampling technique
Rondina et al analysed stroke anatomical MRI images to reduce the stroke outcome prediction bias caused by
using Gaussian process regression, and found that the between-class imbalance among multiple data sets.
patterns of voxels performed better than lesion load per Brain images have been analysed to predict the outcome
region as the predicting features.53 of stroke treatment. Chen et al analysed CT scan data via
ML methods have also been applied to analyse CT scans ML for evaluating the cerebral oedema following hemi-
from patients with stroke. Free-floating intraluminal spheric infarction.63 They built random forest to automat-
thrombus may be formed as lesion after stroke, which is ically identify cerebrospinal fluid and analyse the shifts
difficult to be distinguished with carotid plaque on the on CT scan, which is more efficient and accurate than
CT imaging. Thornhill et al used three ML algorithms conventional methods. Siegel et al extracted functional
to classify these two types by quantitative shape analysis, connectivity from MRI and functional MRI data, and used
including linear discriminant analysis, artificial neural ridge regression and multitask learning for cognitive defi-
network and SVM.54 The accuracy for each method varies ciency prediction after stroke.64 Hope et al studied the
between 65.2% and 76.4%. relationship between lesions extracted from MRI images
and the treatment outcome via Gaussian process regres-
Treatment sion model.65 They used the model to predict the severity
ML has also been applied for predicting and analysing of cognitive impairments after stroke and the course of
the performance of stroke treatment. As a critical step recovery over time.
of emergency measure, the outcome of intravenous
thrombolysis (tPA) has strong relationship with the prog-
nosis and survival rate. Bentley et al used SVM to predict Conclusion and discussion
whether patients with tPA treatment would develop symp- We reviewed the motivation of using AI in healthcare,
tomatic intracranial haemorrhage by CT scan.55 They presented the various healthcare data that AI has analysed
used whole-brain images as the input into the SVM, which and surveyed the major disease types that AI has been
performed better than conventional radiology-based deployed. We then discussed in details the two major cate-
methods. To improve the clinical decision-making process gories of AI devices: ML and NLP. For ML, we focused
of tPA treatment, Love et al proposed a stroke treatment on the two most popular classical techniques: SVM and
model by analysing practice guidelines, meta-analyses neural network, as well as the modern deep learning tech-
and clinical trials using Bayesian belief network.56 The nique. We then surveyed the three major categories of AI
model consisted of 56 different variables and three deci- applications in stroke care.
sions for analysing the procedure of diagnosis, treatment A successful AI system must possess the ML component
and outcome prediction. Ye et al used interaction trees for handling structured data (images, EP data, genetic

240 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

data) and the NLP component for mining unstructured clinical platform that can help cardiologists to diagnose
texts. The sophisticated algorithms then need to be cardiac diseases.23
trained through healthcare data before the system can The second hurdle is data exchange. In order to work
assist physicians with disease diagnosis and treatment well, AI systems need to be trained (continuously) by
suggestions. data from clinical studies. However, once an AI system
The IBM Watson system is a pioneer in this field. The gets deployed after initial training with historical data,
system includes both ML and NLP modules, and has continuation of the data supply becomes a crucial issue
made promising progress in oncology. For example, in for further development and improvement of the system.
a cancer research, 99% of the treatment recommenda- Current healthcare environment does not provide incen-
tions from Watson are coherent with the physician deci- tives for sharing data on the system. Nevertheless, a health-
sions.66 Furthermore, Watson collaborated with Quest care revolution is under way to stimulate data sharing in
Diagnostics to offer the AI Genetic Diagnostic Analysis.66 the USA.69 The reform starts with changing the health
In addition, the system started to make impact on actual service payment scheme. Many payers, mostly insurance
clinical practices. For example, through analysing genetic companies, have shifted from rewarding the physicians by
data, Watson successfully identified the rare secondary shifting the treatment volume to the treatment outcome.
leukaemia caused by myelodysplastic syndromes in Furthermore, the payers also reimburse for a medica-
Japan.67 tion or a treatment procedure by its efficiency. Under
The cloud-based CC-Cruiser in24 can be one prototype this new environment, all the parties in the healthcare
to connect an AI system with the front-end data input system, the physicians, the pharmaceutical companies
and the back-end clinical actions. More specifically, when and the patients, have greater incentives to compile and
patients come, with their permission, their demographic exchange information. Similar approaches are being
information and clinical data (images, EP results, genetic explored in China.
results, blood pressure, medical notes and so on) are
collected into the AI system. The AI system then uses the Correction notice This paper has been corrected since it was published Online
First. Owing to a scripting error, some of the publisher names in the references
patients’ data to come up with clinical suggestions. These were replaced with 'BMJ Publishing Group'. This only affected the full text version,
suggestions are sent to physicians to assist with their clin- not the PDF. We have since corrected these errors and the correct publishers have
ical decision making. Feedback about the suggestions been inserted into the references. Figures 6-9 have also been corrected.
(correct or wrong) will also be collected and fed back into Competing interests None declared.
the AI system so that it can keep improving accuracy. Provenance and peer review Commissioned; internally peer reviewed.
Stroke is a chronic disease with acute events. Stroke Data sharing statement No additional data are available.
management is a rather complicated process with a series Open Access This is an Open Access article distributed in accordance with the
of clinical decision points. Traditionally clinical research Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which
solely focused on a single or very limited clinical questions, permits others to distribute, remix, adapt, build upon this work non-commercially,
and license their derivative works on different terms, provided the original work is
while ignoring the continuous nature of stroke manage- properly cited and the use is non-commercial. See: http://​creativecommons.​org/​
ment. Taking the advantage of large amount of data with licenses/​by-​nc/​4.​0/
rich information, AI is expected to help with studying © Article author(s) (or their employer(s) unless otherwise stated in the text of the
much more complicated yet much closer to real-life clin- article) 2017. All rights reserved. No commercial use is permitted unless otherwise
ical questions, which then leads to better decision making expressly granted.
in stroke management. Recently, researchers have started
endeavours along this direction and obtained promising
initial results.57
Although the AI technologies are attracting substantial References
1. Murdoch TB, Detsky AS. The inevitable application of big data to
attentions in medical research, the real-life implementa- health care. JAMA 2013;309:1351–2.
tion is still facing obstacles. The first hurdle comes from 2. Kolker E, Özdemir V, Kolker E. How Healthcare can refocus on its
Super-Customers (Patients, n =1) and Customers (Doctors and
the regulations. Current regulations lack of standards Nurses) by Leveraging Lessons from Amazon, Uber, and Watson.
to assess the safety and efficacy of AI systems. To over- OMICS 2016;20:329–33.
3. Dilsizian SE, Siegel EL. Artificial intelligence in medicine and cardiac
come the difficulty, the US FDA made the first attempt to imaging: harnessing big data and advanced computing to provide
provide guidance for assessing AI systems.68 The first guid- personalized medical diagnosis and treatment. Curr Cardiol Rep
ance classifies AI systems to be the ‘general wellness prod- 2014;16:441.
4. Patel VL, Shortliffe EH, Stefanelli M, et al. The coming of age of
ucts’, which are loosely regulated as long as the devices artificial intelligence in medicine. Artif Intell Med 2009;46:5–17.
intend for only general wellness and present low risk to 5. Jha S, Topol EJ. Adapting to Artificial Intelligence: radiologists and
pathologists as information specialists. JAMA 2016;316:2353–4.
users. The second guidance justifies the use of real-world 6. Pearson T. How to replicate Watson hardware and systems design
evidence to access the performance of AI systems. Lastly, for your own use in your basement. 2011 https://www.​ibm.​com/​
the guidance clarifies the rules for the adaptive design in developerworks/​community/​blogs/​InsideSystemStorage/​entry/​ibm_​
watson_​how_​to_​build_​your_​own_​watson_​jr_​in_​your_​basement7?​
clinical trials, which would be widely used in assessing the lang=​en (accessed 1 Jun 2017).
operating characteristics of AI systems. Not long after the 7. Weingart SN, Wilson RM, Gibberd RW, et al. Epidemiology of medical
error. BMJ 2000;320:774–7.
disclosure of these guidances, Arterys’ medical imaging 8. Graber ML, Franklin N, Gordon R. Diagnostic error in internal
platform became the first FDA-approved deep learning medicine. Arch Intern Med 2005;165:1493–9.

241
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101
 Open Access

9. Winters B, Custer J, Galvagno SM, et al. Diagnostic errors in the 36. Hirschauer TJ, Adeli H, Buford JA. Computer-Aided diagnosis of
intensive care unit: a systematic review of autopsy studies. BMJ Qual Parkinson's Disease Using Enhanced Probabilistic Neural Network. J
Saf 2012;21:894–902. Med Syst 2015;39:179.
10. Lee CS, Nagy PG, Weaver SJ, et al. Cognitive and system factors 37. Ravi D, Wong C, Deligianni F, et al. Deep Learning for Health
contributing to diagnostic errors in radiology. AJR Am J Roentgenol Informatics. IEEE J Biomed Health Inform 2017;21:4–21.
2013;201:611–7. 38. Lecun Y, Bottou L, Bengio Y, et al. Gradient-based learning applied
11. Neill DB. Using artificial intelligence to improve hospital inpatient to document recognition. Proc IEEE Inst Electr Electron Eng
care. IEEE Intell Syst 2013;28:92–5. 1998;86:2278–324.
12. Administration UFaD. Guidance for industry: electronic source data 39. Research BA. Caffe. 2017. http://​caffe.​berkeleyvision.​org/ (accessed
in clinical investigations. 2013 https://www.​fda.​gov/​downloads/​ 1 Jun 2017).
drugs/​guidances/​ucm328691.​pdf (accessed 1 Jun 2017). 40. Seide F, Agarwal A, eds. CNTK: microsoft's Open-Source Deep-
13. Gillies RJ, Kinahan PE, Hricak H. Radiomics: images are more than Learning Toolkit. ACM SIGKDD International Conference on
pictures, they are data. Radiology 2016;278:563–77. Knowledge Discovery and Data Mining, 2016.
14. Li CY, Liang GY, Yao WZ, et al. Integrated analysis of long non- 41. Abadi M, Agarwal A, Barham P, et al; TensorFlow: large-scale
coding RNA competing interactions reveals the potential role in Machine Learning on heterogeneous distributed Systems, 2016.
progression of human gastric Cancer. Int J Oncol 2016;48:1965–76. 42. Afzal N, Sohn S, Abram S, et al. Mining peripheral arterial disease
15. Shin H, Kim KH, Song C, et al. Electrodiagnosis support system for cases from narrative clinical notes using natural language
localizing neural injury in an upper limb. J Am Med Inform Assoc processing. J Vasc Surg 2017;65:1753–61.
2010;17:345–7. 43. Fiszman M, Chapman WW, Aronsky D, et al. Automatic detection
16. Karakülah G, Dicle O, Koşaner O, et al. Computer based extraction of of acute bacterial pneumonia from chest X-ray reports. J Am Med
phenoptypic features of human congenital anomalies from the digital Inform Assoc 2000;7:593–604.
literature with natural language processing techniques. Stud Health 44. Miller TP, Li Y, Getz KD, et al. Using electronic medical record data to
Technol Inform 2014;205:570–4. report laboratory adverse events. Br J Haematol 2017;177:283–6.
17. Darcy AM, Louie AK, Roberts LW. Machine Learning and the 45. Castro VM, Dligach D, Finan S, et al. Large-scale identification of
Profession of Medicine. JAMA 2016;315:551–2. patients with cerebral aneurysms using natural language processing.
18. Murff HJ, FitzHenry F, Matheny ME, et al. Automated identification Neurology 2017;88:164–8.
of postoperative complications within an electronic medical record 46. Saenger AK, Christenson RH. Stroke biomarkers: progress and
using natural language processing. JAMA 2011;306:848–55. challenges for diagnosis, prognosis, differentiation, and treatment.
19. Somashekhar SP, Kumarc R, Rauthan A, et al. Abstract S6-07: Clin Chem 2010;56:21–33.
double blinded validation study to assess performance of IBM 47. Heeley E, Anderson CS, Huang Y, et al. Role of health insurance in
artificial intelligence platform, Watson for oncology in comparison averting economic hardship in families after acute stroke in China.
with manipal multidisciplinary tumour board ? first study of 638 Stroke 2009;40:2149–56.
breast Cancer cases. Cancer Res 2017;77(4 Suppl):S6-07. 48. Villar JR, González S, Sedano J, et al. Improving human activity
20. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level recognition and its application in early stroke diagnosis. Int J Neural
classification of skin Cancer with deep neural networks. Nature Syst 2015;25:1450036.
2017;542:115–8. 49. Mannini A, Trojaniello D, Cereatti A, et al. A machine Learning
21. Bouton CE, Shaikhouni A, Annetta NV, et al. Restoring cortical Framework for Gait classification using inertial sensors: application
control of functional movement in a human with quadriplegia. Nature to Elderly, Post-Stroke and Huntington's Disease Patients. Sensors
2016;533:247–50. 2016;16:134.
22. Farina D, Vujaklija I, Sartori M, et al. Man/machine interface based on 50. Rehme AK, Volz LJ, Feis DL, et al. Identifying neuroimaging markers
the discharge timings of spinal motor neurons after targeted muscle of Motor Disability in acute stroke by machine Learning Techniques.
reinnervation. Nat Biomed Eng 2017;1:0025. Cereb Cortex 2015;25:3046–56.
23. Marr B. First FDA approval for clinical Cloud-Based Deep Learning in 51. Griffis JC, Allendorfer JB, Szaflarski JP. Voxel-based gaussian naïve
Healthcare. 2017. https://www.​forbes.​com/​sites/​bernardmarr/​2017/​ Bayes classification of ischemic stroke lesions in individual T1-
01/​20/​first-​fda-​approval-​for-​clinical-​cloud-​based-​deep-​learning-​in-​ weighted MRI scans. J Neurosci Methods 2016;257:97–108.
healthcare/#​7a0ed8dc161c (accessed 1 Jun 2017). 52. Kamnitsas K, Ledig C, Newcombe VF, et al. Efficient multi-scale
24. Long E, Lin H, Liu Z, et al. An artificial intelligence platform for the 3D CNN with fully connected CRF for accurate brain lesion
multihospital collaborative management of congenital cataracts, segmentation. Med Image Anal 2017;36:61–78.
2017. 53. Rondina JM, Filippone M, Girolami M, et al. Decoding post-stroke
25. Gulshan V, Peng L, Coram M, et al. Development and Validation of motor function from structural brain imaging. Neuroimage Clin
a Deep Learning Algorithm for detection of Diabetic Retinopathy in 2016;12:372–80.
retinal fundus photographs. JAMA 2016;316:2402–10. 54. Thornhill RE, Lum C, Jaberi A, et al. Can shape analysis differentiate
26. James G, Witten D, Hastie T, et al. An introduction to Statistical free-floating internal carotid artery Thrombus from atherosclerotic
Learning with applications in R. First Edition: Springer, 2013. plaque in patients evaluated with CTA?for stroke or transient
27. Goodfellow I, Bengio Y, Courville A. Deep Learning. First Edition: The ischemic attack? Acad Radiol 2014;21:345–54.
MIT Press, 2016. 55. Bentley P, Ganesalingam J, Carlton Jones AL, et al. Prediction of
28. Kantor P. Foundations of statistical natural language processing: MIT stroke thrombolysis outcome using CT brain machine learning.
Press, 1999:91–2. Neuroimage Clin 2014;4:635–40.
29. Bishop CM, ed. Pattern recognition and machine Learning 56. Love A, Arnold CW, El-Saden S, et al. Unifying acute stroke
(Information Science and Statistics, 2007. treatment guidelines for a bayesian belief network. Stud Health
30. Orrù G, Pettersson-Yeo W, Marquand AF, et al. Using support Technol Inform 2013;192:1012.
Vector Machine to identify imaging biomarkers of neurological 57. Ye H, Shen H, Dong Y, et al. Using Evidence-Based medicine through
and psychiatric disease: a critical review. Neurosci Biobehav Rev Advanced Data Analytics to work toward a National Standard for
2012;36:1140–52. Hospital-based acute ischemic Stroke treatment. Mainland China,
31. Sweilam NH, Tharwat AA, Abdel Moniem NK, Moniem NKA. Support 2017.
vector machine for diagnosis Cancer disease: a comparative study. 58. Zhang Q, Xie Y, Ye P, et al. Acute ischaemic stroke prediction from
Egyptian Informatics Journal 2010;11:81–92. physiological time series patterns. Australas Med J 2013;6:280–6.
32. Khedher L, Ram?rez J, G?rriz JM, et al. Early diagnosis of 59. Asadi H, Dowling R, Yan B, et al. Machine learning for outcome
Alzheimer?s disease based on partial least squares, principal prediction of acute ischemic stroke post intra-arterial therapy. PLoS
component analysis and support vector machine using segmented One 2014;9:e88225.
MRI images. Neurocomputing 2015;151:139–50. 60. Asadi H, Kok HK, Looby S, et al. Outcomes and complications after
33. In: Mirtskhulava L, Wong J, Al-Majeed S, Pearce G, et al; eds. endovascular treatment of Brain Arteriovenous Malformations: a
Artificial Neural Network Model in Stroke diagnosis. modelling Prognostication Attempt using artificial Intelligence. World Neurosurg
and simulation (UKSim), 2015 17th UKSim-AMSS International 2016;96:562–9.
Conference on: IEEE, 2015. 61. Birkner MD, Kalantri S, Solao V, et al. Creating diagnostic scores
34. Khan J, Wei JS, Ringnér M, et al. Classification and diagnostic using data-adaptive regression: an application to prediction of 30-
prediction of cancers using gene expression profiling and artificial day mortality among stroke victims in a rural hospital in India. Ther
neural networks. Nat Med 2001;7:673–9. Clin Risk Manag 2007;3:475–84.
35. Dheeba J, Albert Singh N, Tamil Selvi S. Computer-aided detection 62. Ho KC, Speier W, El-Saden S, et al. Predicting discharge mortality
of breast Cancer on mammograms: a swarm intelligence optimized after acute ischemic stroke using balanced data. AMIA Annu Symp
wavelet neural network approach. J Biomed Inform 2014;49:45–52. Proc 2014;2014:1787–96.

242 Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101

 Open Access

63. Chen Y, Dhar R, Heitsch L, et al. Automated quantification of cerebral 67. Otake T. IBM Big Data used for rapid diagnosis of rare leukemia
edema following hemispheric infarction: application of a machine- case in Japan. 2016 http://www.​japantimes.​co.​jp/​news/​2016/​08/​11/​
learning algorithm to evaluate CSF shifts on serial head CTs. national/​science-​health/​ibm-​big-​data-​used-​for-​rapid-​diagnosis-​of-​
Neuroimage Clin 2016;12:673–80. rare-​leukemia-​case-​in-​japan (accessed 1 Jun 2017).
64. Siegel JS, Ramsey LE, Snyder AZ, et al. Disruptions of network 68. Graham J. Artificial Intelligence, Machine Learning, and the FDA.
connectivity predict impairment in multiple behavioral domains after 2016 https://www.​forbes.​com/​sites/​theapothecary/​2016/​08/​19/​
stroke. Proc Natl Acad Sci USA 2016;113:E4367–E4376. artificial-​intelligence-​machine-​learning-​and-​the-​fda/#​4aca26121aa1
65. Hope TM, Seghier ML, Leff AP, et al. Predicting outcome and (accessed 1 Jun 2017).
recovery after stroke with lesions extracted from MRI images. 69. Kayyali B, Knott D, Kuiken SV. The big-data revolution in US health
Neuroimage Clin 2013;2:424–33. care: accelerating value and innovation. 2013 http://www.​mckinsey.​
66. LOHR S. IBM is counting on its bet on Watson, and Paying Big com/​industries/​healthcare-​systems-​and-​services/​our-​insights/​the-​
Money for It. 2016 https://www.​nytimes.​com/​2016/​10/​17/​technology/​ big-​data-​revolution-​in-​us-​health-​care (accessed 1 Jun 2017).
ibm-​is-​counting-​on-​its-​bet-​on-​watson-​and-​paying-​big-​money-​for-​it.​
html (accessed 1 Jun 2017).

243
Jiang F, et al. Stroke and Vascular Neurology 2017;2:e000101. doi:10.1136/svn-2017-000101