TELKOMNIKA Telecommunication Computing Electronics and Control

Vol. 21, No. 4, August 2023, pp. 825~835

ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v21i4.24065  825

Biomedical-named entity recognition using CUDA accelerated

KNN algorithm

Manish Bali1, Anandaraj Shanthi Pichandi2, Jude Hemanth Duraisamy3

1
Department of Computer Science and Engineering, Presidency University, Bengaluru-560064, Karnataka, India
2
Department of Computer Science and Engineering, Presidency University, Bengaluru-560064, Karnataka, India
3
Department of Electronics and Communication Engineering, Karunya Institute of Technology and Sciences, Coimbatore, India

Article Info ABSTRACT

Article history: Biomedical named entity recognition (Bio-NER) is a highly complex and
time-consuming research domain using natural language processing (NLP).
Received May 28, 2022 It’s widely used in information retrieval, knowledge summarization,
Revised Dec 29, 2022 biomolecular event extraction, and discovery applications. This paper
Accepted Feb 16, 2023 proposes a method for the recognition and classification of named entities in
the biomedical domain using machine learning (ML) techniques. Support
vector machine (SVM), decision trees (DT), K-nearest neighbor (KNN), and
Keywords: its kernel versions are used. However, recent advancements in
programmable, massively parallel graphics processing units (GPU) hold
BioNLP promise in terms of increased computational capacity at a lower cost to
Graphics processing unit address multi-dimensional data and time complexity. We implement a novel
Machine learning parallel version of KNN by porting the distance computation step on GPU
Named entity recognition using the compute unified device architecture (CUDA) and compare the
Natural language processing performance of all the algorithms using the BioNLP/NLPBA 2004 corpus.
Results demonstrate that CUDA-KNN takes full advantage of the GPU’s
computational capacity and multi-leveled memory architecture, resulting in a
35× performance enhancement over the central processing unit (CPU). In a
comparative study with existing research, the proposed model provides an
option for a faster NER system for higher dimensionality and larger datasets
as it offers balanced performance in terms of accuracy and speed-up, thus
providing critical design insights into developing a robust BioNLP system.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Manish Bali
Department of Computer Science and Engineering, Presidency University
Itgalpur, Rajankunte, Yelahanka, Bengaluru-560064, Karnataka, India
Email: balimanish0@gmail.com

1. INTRODUCTION
Biomedical-named entity recognition (Bio-NER) is a sub-task of information extraction that classifies
all unregistered terms found in biomedical texts. Location, person, organization, date, time, percentage,
monetary value, and none of the above are the eight categories used by named entity recognition [1], [2].
It locates named entities in sentences and brackets them under one of the categories. For example, in the sentence:
Apple (organization) CEO Tim Cook (person) introduces two new, larger iPhones, smartwatch at Cupertino
(location), and Flint center (organization) event [3]. As captured against each of the words keeping the context of
the sentence in view, instead of a fruit name apple has been recognized as an organization. A person’s name is
arrived at by combining Tim and Cook representing their chief executive officer (CEO) and so forth. Bio-NER
assigns a named entity tag to a given word using rules and heuristics. To categorize a word into one of the
following three categories: person, organization, or date, it is essential to identify the specified entity and to extract

Journal homepage: http://telkomnika.uad.ac.id

826  ISSN: 1693-6930

symbolic and discernible information from a text [4], [5]. By using current categories and none of the above,
named entity recognition categorizes every term in the document. This includes recognizing named entities such
as proteins, deoxyribonucleic acid (DNA), ribonucleic acid (RNA), cells, and other key entities in biomedical
literature. It is a hot topic of research, but the literature review shows a clear gap of at least 10% in accuracy
when comparing the best-performing algorithms for Bio-NER with generic news or text [6]-[9].
Researchers studying natural language processing are particularly interested in the information extraction
of genes, cancer, and proteins from biomedical literature. Although it is essential to extracting biological
information, Bio-NER has only been treated as the first stage of text mining in biomedical literature [10]-[14].
Recognizing technical terms in the biomedical area has long been one of the most challenging problems in
natural language processing related to biomedical research [15]. In contrast to the categories used in the
conventional named entity recognition technique, in this paper, we use five categories to recognize a protein,
DNA, RNA, cell line, and cell type. An example of a Bio-NER tagged sentence is: IL-2 (B-protein)
responsiveness requires three distinct elements (B-DNA) within the enhancer (B-DNA). As seen from the above
simple sentence, finding biomedical named entity recognition is difficult and complex from the text on account
of the following reasons [16], [17]:
− Because of how quickly contemporary research is progressing, there are an increasing number of new
technical terms. Thus, it is challenging to create a gazetteer that includes all the new jargon.
− The same words or expressions may be classified as different named entities depending on the context.
− An entity’s length is quite long, and it may contain control characters like hyphens (for example,
12-O-Tetradecanoylphorbol-13-acetate).
− In the biomedical domain, abbreviated expressions are widely utilized, and they have a sense of ambiguity.
Tissue culture fluid (TCF) might, for example, stand for T cell factor or tissue culture fluid (TCF) [18].
− In biomedical language, normal and functional meanings are combined or incorporated, which is why a
phrase may become very long. Human T-lymphotropic virus I (HTLVI)-infected and HTLV-I-transformed,
for example, comprise the phrases I, infected, and transformed.
− Biomedical NER has a hard time segmenting sentences with named entities. Changes in spelling can
also be problematic.
− A named entity from one category may be subsumed by a named entity from another category.
As a result, experts continue to find this to be a complicated area. Three types of methodologies have
been studied for identifying biological named entities: rule-based heuristics [19], [20], dictionary-based [21]
approaches, and statistical machine learning techniques [22]. As many named entities as possible are stored in
a gazette using a dictionary-based mechanism. Though straightforward, this method has its drawbacks.
Firstly, since target terms are mostly proper nouns or unregistered words, named entity recognition is tough.
Additionally, new words are produced often, and depending on the context, the same word stream may be
recognized as a variety of named entities. A rule-based approach is based on the conventions and patterns of
named entities that are present in sentences that are spoken. Although the problem of many named entities can
be solved using context in rule-based procedures, each rule must be created before it is used. The third
approach, based on machine learning, links the discovered items to words even though the words are not in the
dictionary and the context is not provided in the rule set. In this domain, popular methods include support vector
machine (SVM) [5], decision tree (DT) [23], Hidden Markov models (HMM) [18], [24], maximum entropy
Markov model (MEMM) [25], and conditional random field (CRF) [24]. When training with test data that isn’t
part of the training data, supervised learning techniques use data with a variety of features based on several
linguistic criteria. Mathu and Raimond [26] used Bi-LSTM with residuals for a drug-named entity recognition task
with only a 40% recognition rate for 4-gram entities. Ramachandran and Arutchelvan [27] have implemented a
hybrid model in which the annotated entities are trained by a blank Spacy ML model validated by this dictionary
and human. The average F1 score was 73.79% [28] used the classical rule-based approach on PubMed articles
achieved an F1 score of 77.65%. However, it has been noticed that the high-end systems built do not produce
adequate results on multi-dimensional and large datasets because no single method can generate higher
performance. There is a pressing need to address these issues. Graphic processing units (GPUs) on the other
hand have become more programmable and emerged as a form of manycore processor with highly parallelized
and multi-threaded capabilities during the last decade [29]. The modern GPU provides great computing
horsepower and higher memory bandwidth when compared to generic x86-based central processing units
(CPUs) at a fraction of the cost. The GPU is now at the cutting edge of chip-level parallelism, with applications
ranging from 3D rendering [30] to general-purpose computing [31].
In this paper, three machine learning (ML) techniques from different paradigms are tested on the
BioNLP/NLPBA 2004 corpus [32], [33], which contains 22,402 sentences. SVM, DT, and K-nearest neighbor
(KNN) with two kernel versions i.e., KNN-1 and KNN-5 are implemented on a CPU. Among these, the KNN
algorithm is a commonly used machine learning technique for pattern recognition, classification, and

TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 4, August 2023: 825-835
TELKOMNIKA Telecommun Comput El Control  827

regression [21], [34]. This algorithm is an instance based-lazy learning technique that is utilized in a variety of
applications like image processing and pattern matching. Even though the algorithm is straightforward, it takes
a while to run. The application’s execution time may become a bottleneck when the train and test data sets are
both very large. In this study, we describe a new parallel implementation of the KNN algorithm based on the
compute unified device architecture (CUDA) and GPU from NVIDIA, and we compare the performance with
that of conventional CPU execution. The research tries to investigate if this parallelized version of KNN
surpasses other machine learning algorithms in terms of accuracy and speedup, which could provide
interesting insights into the development of a reliable Bio-NER system.
Following are the sections of this manuscript: section 2 presents the methodology used in terms of
explaining the three classifiers used and the dataset chosen, the basic principle of the KNN algorithm, and a
note on NVIDIA GPU and CUDA technology and the programming approach adopted on the algorithm
towards a KNN-CUDA version. Section 3 presents both the classification and execution time results and a
comparison with other existing models along with a detailed analysis. Finally, the conclusion is covered in
section 4 followed by references.

2. METHOD
Figure 1 shows a high-level flow chart of the complete procedure. The BioNLP/JNLPS 2004 corpus
is the data set used. The data is pre-processed using natural language processing operations such as stop word
removal, numeric value removal, tokenization, and lemmatization because it contains a lot of undesired
data/noise that can affect the overall accuracy of the model. The preprocessed data is then transformed using
the word2vec word embedding or feature extraction technique into numerical vectors. Before being fed to the
machine learning models, the data is divided into train (2000 MEDLINE abstracts) and test (404 MEDLINE
abstracts) data. Three popular machine learning algorithms, SVM, DT, and KNN are used for the Bio-NER
classification of words in a sentence, and the efficacy of each technique is compared in the first part of the
study. In the second part of the study, all the algorithms executed on a traditional CPU for efficacy are also
checked on their execution time, which is captured in ms (milliseconds). The KNN distance computation is
parallelized using NVIDIA CUDA GPU and a novel data segmentation approach as explained in section 2,
section 3, and section 4.

Figure 1. Flow-chart of the proposed methodology

This CUDA version of KNN is run on a CPU+GPU in the system and execution time is noted. This
time is compared with the earlier CPU run and performance compared. The next sections provide a thorough
explanation of the dataset used, the algorithm, and the parallelization of KNN.

Biomedical-named entity recognition using CUDA accelerated KNN algorithm (Manish Bali)
828  ISSN: 1693-6930

2.1. Dataset
We use 22402 sentences in the BioNLP/NLPBA 2004 corpus [31], [32]. The training data set
consists of 18546 and the test data consists of 3856 sentences. The dataset is further divided into groups like
protein, DNA, RNA, cell line, and cell type.

2.2. Support vector machine

This approach for supervised machine learning is often used to solve classification and regression
problems. It is a binary, linear, non-probabilistic classifier that makes use of data patterns. It reduces
generalization error on unobserved examples by employing the structural risk minimization principle.
As depicted in Figure 2, the objective of the SVM algorithm is to construct the optimal decision boundary, or
hyperplane, that can divide 𝑛-dimensional space into classes and make it simple to classify fresh data points.
It selects the extreme points or vectors called support vectors that aid in building the hyperplane. SVM uses the
following function to predict:

𝑌 ′ = 𝑤 × 𝜙(𝑥) + 𝑏 (1)

The non-linear transform is stated in (1) by 𝜙(𝑥), with the primary goal of obtaining the appropriate values
of 𝑤, 𝑏. In (1), 𝑌’ is obtained by minimizing the risk of regression.
1
𝑅𝑟𝑒𝑔 (𝑌 ′ ) = 𝐶 × ∑𝑙𝑖=0 𝛾(𝑌𝑖′ − 𝑌𝑖 ) + × ‖𝑤‖2 (2)
2

Where:

𝑤 = ∑𝑙𝑗=1(𝛼𝑗 − 𝛼𝑗∗ )𝜙(𝑥𝑗 ) (3)

Figure 2. Illustration of support vector machine [35]

In (3), the parameters α and α∗ state the relaxation parameter called the Lagrange multiplier. The output
obtained is:

𝑌 ′ = ∑𝑙𝑗=1(𝛼𝑗 − 𝛼𝑗∗ )𝜙(𝑥𝑗 ) × 𝜙(𝑥) + 𝑏 (4)

𝑌 ′ = ∑𝑙𝑗=1(𝛼𝑗 − 𝛼𝑗∗ ) × 𝐾(𝑥𝑗 , 𝑥) + 𝑏 (5)

In (4) and (5), 𝐾(𝑥𝑗 , 𝑥) states the kernel function. To extract named entities, SVM with linear kernel function
is used in this study.

2.3. Decision tree

A non-parametric supervised learning approach called a decision tree is employed for both
classification and regression. Figure 3 depicts its hierarchical tree structure, which includes a root node,
branches, internal modes, and leaf nodes. We employed the C5.0 model, which performs recursive
partitioning across the biological data, to extract event expressions. Starting at the root node, each node of the
tree divides the feature vector or data into multiple branches based on the association rule between the split
criteria. It is an inductive method that aids to acquire knowledge in classification.

TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 4, August 2023: 825-835
TELKOMNIKA Telecommun Comput El Control  829

2.4. K-nearest neighbor (KNN)

K-nearest neighbor is utilized for both classification and regression issues, and there is no training step;
instead, the entire data set is used to predict/classify fresh data. As shown in Figure 4, when a new data point
needs to be categorized, it determines how far apart from all other data points in categories A and B it is. Next it
determines where the dataset’s nearest neighbor(s) are based on the 𝑘 value; if 𝑘 = 1, all points are assigned to
the same class as the data point with the smallest distance based on the lowest distance between them. A list of
the 𝑘 minimum distances between all of the data points is returned if 𝑘 is higher than one.

2.4.1. Principle of KNN algorithm

According to the method, a query point must fall into a certain category if the majority of the 𝑘
samples that are most similar to it in the feature space also do. This method is known as the KKN algorithm
because the distance in the feature space can be used to assess similarity. A trained data set with precise
classification labels should be known before the technique is started. Then determine the distance between
each point in the train data set and a query data point 𝑞𝑖, whose label is unknown and which is represented by
a vector in the feature space. The labels of the 𝑘 closest points in the train data set may be used to decide the
test point 𝑞𝑖’s class label after sorting the results of the distance computation.

Figure 3. Illustration of a decision tree [36] Figure 4. Illustration of KNN algorithm [37]

Each point in 𝑑-dimensional space can be represented as a 𝑑-vector of coordinates, as represented in (6).

𝑝 = (𝑝1 , 𝑝2 , . . 𝑝𝑛 ) (6)

The distance between two points in a multidimensional feature space can be specified in a variety of ways.
The Euclidean distance is frequently calculated [37] using (7).

𝑑𝑖𝑠𝑡(𝑝, 𝑞) = √∑𝑛𝑖=1(𝑝𝑖 − 𝑞𝑖 )2 (7)

The quality of the training dataset has a direct impact on the classification outcomes. The choice of the parameter 𝑘
is crucial since various 𝑘 values can lead to various classification labels. In this study, 𝑘 = 1 and 𝑘 = 5 are used.
High-dimensional data sets can be handled via the KNN approach, which is simple to use. The computational
complexity is significant and the operation time is considerable when the test set, training set, and data dimension
are higher than projected. When the test and training sets consist of 𝑚 and 𝑛 vectors in 𝑑-dimensional feature
space, respectively, the temporal complexity of this method is 𝑂(𝑚 × 𝑛 × 𝑑).

2.4.2. NVIDIA GPU and CUDA

Theoretically, GPU performance outperforms CPU performance by a wide margin. This is because
GPUs are built for computationally demanding, highly parallel computations, like those involved in
rendering graphics. Because of the enormous processing capacity provided by GPU at a relatively low cost,
researchers have experimented to use a GPU for non-graphics tasks. This effort was called a general-purpose
graphics processing unit (GPGPU). Yet, it was challenging to design GPUs for non-graphics applications
before the introduction of CUDA. For general-purpose computation on its GPUs, NVIDIA’s CUDA is a
parallel programming language. With the help of CUDA, programmers can accelerate computationally
extensive parts of the algorithm by making use of GPUs’ capacity to parallelize this portion of the code.
Three key abstractions form the basis of the CUDA parallel programming concept:
− A hierarchy of thread groups
− Shared memories
− Synchronization of barriers

Biomedical-named entity recognition using CUDA accelerated KNN algorithm (Manish Bali)
830  ISSN: 1693-6930

The CUDA abstractions, which are nested beneath coarse-grained data parallelism and task
parallelism, enable fine-grained data parallelism and thread parallelism. They demonstrate to the programmer
how to break the problem down into coarse sub-problems that blocks of threads can solve independently in
parallel and finer sub-problems that the threads in the block can solve jointly in parallel. Figure 5(a)
demonstrates how a kernel is executed concurrently by a collection of threads. Each thread executes the same
code and has an ID that it uses to calculate memory addresses and make control decisions. Threads are
arranged into a grid of thread blocks, as shown in Figure 5(b). Different kernels may employ various block
configurations. The shared memory that is accessible to threads from the same block allows for synchronized
execution. Thread blocks are necessary for each thread to run separately. These must be carried out in any
arrangement, whether sequential or parallel. Programmers can organize thread blocks in any order across any
number of cores, allowing them to create code that scales with the number of cores. Threads can cooperate within a
block by exchanging information via shared memory and synchronizing their execution to manage memory
accesses. Grids of blocks and thread blocks, as shown in Figure 6, might be one, two, or three-dimensional.

(a) (b)

Figure 5. Depiction of (a) CUDA kernel and (b) a grid of CUDA thread blocks

Figure 6. Thread organization in CUDA

2.4.3. KNN algorithm in CUDA

The KNN algorithm can be parallelized in CUDA using the fundamental steps listed below. The
initialization of the test data set and the labeled 𝑑-dimensional train data set is the first stage in the pre-processing
of the data. Next, determine the distances between one test point and each train data point using a sample from the
test data set.
The results of distance computation are sorted (using CUDA-based radix sort) in the following step,
and the 𝑘 smallest results are chosen based on the parameter 𝑘. The test point’s class label is determined in
the fourth step by using the election result of 𝑘 points (simple statistical election). A different point in the test
data set should then be chosen, and step two repeated until the test data set is empty. The Euclidean distance
is applied in this study.

TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 4, August 2023: 825-835
TELKOMNIKA Telecommun Comput El Control  831

A data segmentation approach is used to calculate distance. Each vector in the data set needs 𝑑
features to represent it because both the train data set 𝐴 and the test data set 𝐵 are 𝑑-dimensional. We restore
the train data set as an 𝑛 × 𝑑 matrix in memory and the test data set as an 𝑚 × 𝑑 matrix if the train data set
has 𝑛 vectors and the test data set has 𝑚 vectors. The result set 𝐶, which includes all the distances between
each pair of points in 𝐴 and 𝐵, is described by a 𝑚𝑛 matrix. The distance between the vector in 𝐴 with row
number 𝑥 and the vector in 𝐵 with row number 𝑦 is therefore represented by a column 𝑥 and row 𝑦 element
in data set 𝐶. The distances recovered in C are squared Euclidean distances because the computation of the
square root has no impact on the outcomes of the sorting.
There are many 𝑇-width tiles in the final data set 𝐶. The GPU’s threads each handle one 𝐶 element
and compute distance, with the results from all threads being combined in the output. The CUDA kernel
function pseudo algorithm is described in Table 1.

Table 1. Algorithm for distance computation (train, test, result)

{In 2D, each thread in each block is designated by the letters 𝑏𝑥, 𝑏𝑦, 𝑡𝑥, and 𝑡𝑦}
𝑠𝑢𝑏_𝑟𝑒𝑠𝑢𝑙𝑡 = 0;
t𝑒𝑚𝑝 = 0;
for each subtitle in dimension/𝑇 do
loads 𝑠ℎ𝑎𝑟𝑒𝑑_𝑡𝑟𝑎𝑖𝑛 in the train set to shared memory;
loads 𝑠ℎ𝑎𝑟𝑒𝑑_𝑡𝑒𝑠𝑡 in the test set to a shared memory;
sync threads;
for 𝐾 = 0 to 𝑇 do
𝑡𝑒𝑚𝑝 = 𝑠ℎ𝑎𝑟𝑒𝑑_𝑡𝑒𝑠𝑡[𝑡𝑦][𝑘] 𝑠ℎ𝑎𝑟𝑒𝑑_𝑝𝑟𝑜𝑏[𝑡𝑥][𝑘];
𝑠𝑢𝑏_𝑟𝑒𝑠𝑢𝑙𝑡+= 𝑡𝑒𝑚𝑝 ∗ 𝑡𝑒𝑚𝑝;
end for
sync threads;
add 𝑠𝑢𝑏_𝑟𝑒𝑠𝑢𝑙𝑡𝑠 to the corresponding position in the result set;
end for

3. RESULTS AND DISCUSSION

The experimental setup used for the study is based on an Intel i7, 4-core, 8-thread processor-based
system. The CPU-only runs use this configuration. The CUDA version of the KNN algorithm looks out for a
GPU and the system has a NVIDIA GeForce MX450 2GB GDDR5 graphics memory card with 896 CUDA
cores. The CPU off-loads the most compute-intensive part of the algorithm like the distance calculation onto
the GPU which uses the inherent parallelism to accelerate this computation and provide the interim result
back to the CPU which compiles the overall results.

3.1. Classifier performance on CPU

Five classes of biomedical named entities, protein, DNA, RNA, cell line, and cell type were tagged
in the first part of the study using the BioNLP/NLPBA 2004 dataset. The GENIA version 3.02 corpus, which
contains 2,000 MEDLINE abstracts, is where the train data came from. 404 MEDLINE abstracts from the
GENIA study were used for testing. The findings are presented as F1-scores, which are determined by
applying (8):

𝐹1 − 𝑠𝑐𝑜𝑟𝑒 = (2 × 𝑃𝑅)/(𝑃 + 𝑅) (8)

Where 𝑅 stands for recall and 𝑃 stands for precision. 𝑃 is the quantity of correctly found named entity
chunks divided by the quantity of correctly found named entity chunks, and 𝑅 is the quantity of correctly found
named entity chunks divided by the quantity of truly found named entity chunks. When every component of the
recognized named entity has been correctly identified, it is regarded as the correct named entity. Figure 7(a)
displays the results of the four different individual classifiers on the test corpus. The best F1-score is 71.17 percent
for the decision tree model, followed by 70.16 percent for the KNN-5 model. SVM has a score of 68.02 percent,
whereas KNN-1 has a score of 67.11 percent.
Considering the timings, Figure 7(b) illustrates the time taken by these classifiers on the test corpus.
The support vector machine method, followed by decision tree and K-nearest neighbor algorithms, requires
the least amount of CPU time. The distance computation phase of the K-nearest neighbor (KNN-5) model
consumes the greatest CPU time. We select different feature sets for these classifiers since the sensitivity and
effectiveness of each classifier vary. KNN models employ local features, full-text features, and a fraction of
external resource features; the DT model mostly uses local features, whereas SVM models only use token
features due to SVM’s inefficiency. There is no post-processing applied to these models. The performance of
the CUDA KNN method on this dataset is examined in the following section.

Biomedical-named entity recognition using CUDA accelerated KNN algorithm (Manish Bali)
832  ISSN: 1693-6930

3.2. CUDA accelerated KNN performance

GPU and CPU refer to the CUDA implementation and the serial CPU approach compared in this
paper respectively. The initialization and output components of the program are essentially identical. The
core components of KNN algorithms’ execution times are shown in Figure 8(a) and in comparison to other
algorithms in Figure 8(b). As can be seen from the numbers, the GPU technique has a distinct advantage in
terms of overall execution time. An execution time speedup of 32× is achieved on the CUDA implementation
of the KNN-1 algorithm as compared to the CPU algorithm execution time, while an execution time speedup
of 35× is reached on the CUDA implementation of KNN-5 algorithm when compared to the CPU approach
execution time. This is because, in CPU implementation, the distance computing step takes up most of the
execution time, which gets accelerated on GPU implementations the most.

(a) (b)

Figure 7. Classifier performance results on CPU and test corpus in terms of (a) accuracy and (b) execution
time in ms

(a)

(b)

Figure 8. Result on GPU: (a) KNN algorithm execution time and (b) execution time for all algorithms in ms

TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 4, August 2023: 825-835
TELKOMNIKA Telecommun Comput El Control  833

As a classification algorithm, the approaches’ efficiency should be mandatorily considered, as well

as if GPU implementation has had any effect on F1-score. However, since (i) their principles are identical
and (ii) they are trained for the same number of epochs, the classification accuracy of both approaches is
always similar. The only difference and advantage are that the higher-performance GPU version of KNN
finishes faster, allowing for a larger data set to be processed in the same amount of time. If we train all the
algorithms for the same amount of time, the accuracy will vary because the higher-performance GPU version
of KNN will complete a larger number of training iterations.

3.3. Performance comparison with existing models

Table 2 compares the results achieved with state-of-the-art existing models on the common
BioNLP/NLPBA 2004 corpus to check their efficacy with the proposed models. The F1-score of the DT algorithm
is 71.17%, which is the highest of all the findings. This is because it compels the analysis of every possible
decision consequence and tracks every path to a conclusion. It generates a thorough analysis of the outcomes along
each branch and indicates the decision nodes that need more research. KNN-5 comes in second with 70.16%,
followed by SVM with 68.02%. Maximum entropy was utilized by [25] to acquire an F1-score of 67.41%. Using
SVM and CRF [32] were able to attain a score of 66.28%. HMM was used by [28], who achieved a 65.7 percent
score. The models proposed by Krauthammer and Nenadic [19] are also compared, with F1 scores of 60.75 and
58.80 percent, respectively. Although the CUDA version of KNN performs second best, its faster execution time
and accuracy provide significant insights into developing a robust and fast-named entity recognition system for
larger data sets. Also, it holds steady for more recent hybrid [27] and classical rule-based [28] approaches.

Table 2. Performance comparison with existing models

Reference research Methodology F1-score (%) Time taken (in ms)
[19] Jordan type 60.75 NA
[19] Elman type RNN 58.80 NA
[32] SVM, CRF 66.28 NA
[28] HMM 65.70 NA
[25] Maximum entropy 67.41 NA
Proposed models Decision tree 71.17 65050
SVM 68.02 49860
KNN-1, GPU 67.11 3350
KNN-5, GPU 70.16 4380

4. CONCLUSION
Since technical terms in biological texts have distinctive properties, Bio-NER has more challenges
than regular NER. In the first study, we compared the performance of three machine-learning approaches from
various paradigms and measured their execution time on a standard CPU. We next GPU or CUDA enables the
distance computation and present the GPU-version of the KNN algorithm in the second study. Results show that
the CUDA enabled solution fully utilizes the GPU’s computational capabilities and multi-leveled memory
architecture, resulting in a 35× speedup, a notable performance improvement over the traditional CPU
implementation. And this is achieved on a sub-hundred-dollar graphics card, which is worth noting.
The suggested GPU version of the KNN algorithm is valuable for applications having high dimensionality
and large amounts of data. Comparing proposed models with existing research, though DT returns the best results
in terms of accuracy in the true sense, the GPU version of KNN with 35× speedup over a CPU and near-best levels
of accuracy is a good option to consider in building a robust and fast NER system for large multi-dimensional
datasets. As part of a future study, we intend to investigate deep neural networks for biomolecular event extraction
tasks for which named entity recognition is a sub-task. To develop a quick and efficient BioNLP system,
paradigms like transfer learning and bidirectional encoder representations from transformers (BERT) models will
also be investigated along with other newer hardware.

REFERENCES
[1] P. Bose, S. Srinivasan, W. C. Sleeman, J. Palta, R. Kapoor, and P. Ghosh, “A Survey on Recent Named Entity Recognition and
Relationship Extraction Techniques on Clinical Texts,” Applied Sciences, vol. 11, no. 18, 2021, doi: 10.3390/app11188319.
[2] B. Alshaikhdeeb and K. Ahmad, “Biomedical Named Entity Recognition: A Review,” International Journal on Advanced Science
Engineering and Information Technology, vol. 6, no. 6, pp. 889-895, 2016, doi: 10.18517/ijaseit.6.6.1367.
[3] K. B. Cohen and L. Hunter, “Natural language processing and systems biology,” Artificial Intelligence Methods and Tools for
Systems Biology, Computational Biology, Dordrecht: Springer, 2004, vol. 5, pp. 145–173, doi: 10.1007/978-1-4020-5811-0_9.
[4] E. F. T. K. Sang and F. D. Meulder, “Introduction to the CoNLL-2003 Shared Task: Language-Independent Named Entity
Recognition,” Proceedings of the seventh conference on Natural language learning at HLT-NAACL, 2003, pp. 142-147,
doi: 10.48550/arXiv.cs/0306050.

Biomedical-named entity recognition using CUDA accelerated KNN algorithm (Manish Bali)
834  ISSN: 1693-6930

[5] T. Meenachisundaram and M. Dhanabalachandran, “Biomedical Named Entity Recognition Using the SVM Methodologies and
bio Tagging Schemes,” Revista de Chimie, vol. 72, no. 4, pp. 52-64, 2021, doi: 10.37358/RC.21.4.8456.
[6] A. Ekbal, S. Saha, and U. K. Sikdar, “Biomedical named entity extraction: some issues of corpus compatibilities,” SpringerPlus,
vol. 2, no. 601, 2013, doi: 10.1186/2193-1801-2-601.
[7] X. Wang, C. Yang, and R. Guan, “A comparative study for biomedical named entity recognition,” International Journal of
Machine Learning and Cybernetics, vol. 9, pp. 373–382, 2018, doi: 10.1007/s13042-015-0426-6.
[8] M. C. Cariello, A. Lenci, and R. Mitkov, “A Comparison between Named Entity Recognition Models in the Biomedical Domain,”
Translation and Interpreting Technology Online, pp. 76–84, 2021, doi: 10.26615/978-954-452-071-7_009.
[9] T. M. Thiyagu, D. Manjula, and S. Shridhar, “Named Entity Recognition in Biomedical Domain: A Survey,” International
Journal of Computer Applications, vol. 181, no. 41, 2019, doi: 10.5120/ijca2019918469.
[10] D. Cutting, J. Kupiec, J. Pedersen, and P. Sibun, “A practical part-of-speech tagger,” in Proceedings of the third conference on
applied natural language processing, 1992, pp. 133-140, doi: 10.3115/974499.974523.
[11] A. R. Aronson, T. C. Rindflesch, and A. C. Browne, “Exploiting a large thesaurus for information retrieval,” RIAO ‘94:
Intelligent Multimedia Information Retrieval Systems and Management, vol. 94, pp. 197–216, 1994. [Online]. Available:
https://lhncbc.nlm.nih.gov/ii/information/Papers/riao94.final.pdf
[12] P. G. Baker, C. A. Goble, S. Bechhofer, N. W. Paton, R. Stevens, and A. Brass, “An ontology for bioinformatics applications,”
Bioinformatics, vol. 15, no. 6, pp. 510–530, 1999, doi: 10.1093/bioinformatics/15.6.510.
[13] N. Perera, M. Dehmer, and F. E. -Streib, “Named Entity Recognition and Relation Detection for Biomedical Information
Extraction,” Frontiers in Cell and Developmental Biology, vol. 8, no. 673, 2020, doi: 10.3389/fcell.2020.00673.
[14] K. Wang et al., “NERO: a biomedical named-entity (recognition) ontology with a large, annotated corpus reveals meaningful
associations through text embedding,” npj systems biology and applications, vol. 7, no. 38, 2021, doi: 10.1038/s41540-021-00200-x.
[15] C. Blaschke, M. A. Andrade, C. Ouzounis, and A. Valencia, “Automatic extraction of biological information from scientific text:
protein-protein interactions”, Proceedings International Conference on Intelligent Systems for Molecular Biology, 1999,
pp. 60–67. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/10786287/
[16] H. Liu, Z. Zhi, M. Torii, C. Wu, and C. Friedman, “Quantitative assessment of dictionary-based protein named entity tagging,”
Journal of the American Medical Informatics Association, vol. 13, no. 5, pp. 497-507, 2006, doi: 10.1197/jamia.M2085.
[17] N. Ponomareva, F. Pla, A. Molina, and P. Rosso, “Biomedical named entity recognition: a poor knowledge HMM-based
approach,” International Conference on Application of Natural Language to Information Systems, 2007, vol. 4592, pp 382–387,
doi: 10.1007/978-3-540-73351-5_34.
[18] M. Craven and J. Kumlien, “Constructing biological knowledge bases by extracting information from text sources,” in
Proceedings International Conference on Intelligent Systems for Molecular Biology, 1999, vol. 7, pp. 77–86. [Online]. Available:
https://www.biostat.wisc.edu/~craven/papers/ismb99.pdf
[19] M. Krauthammer and G. Nenadic, “Term identification in the biomedical literature,” Journal of Biomedical Informatics, vol. 37,
no. 6, pp. 512–538, 2004, doi: 10.1016/j.jbi.2004.08.004.
[20] P. D. Soomro, S. Kumar, Banbhrani, A. A. Shaikh, and H. Raj, “Bio-NER: Biomedical Named Entity Recognition using Rule-Based
and Statistical Learners,” International Journal of Advanced Computer Science and Applications (IJACSA), vol. 8, no. 12, 2017,
doi: 10.14569/IJACSA.2017.081220.
[21] K. Fukuda, A. Tamura, T. Tsunoda, and T. Takagi, “Toward information extraction: identifying protein names from biological
papers,” Pacific Symoposium Biocomputing, 1998, pp. 707-725. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/9697224/
[22] D. Campos, S. Matos, and J. L. Oliveira, “Biomedical Named Entity Recognition: A Survey of Machine-Learning Tools,” Theory
and Applications for Advanced Text Mining, United Kingdom: IntechOpen, 2012, doi: 10.5772/51066.
[23] A. S. Al-Hegami, A. M. F. Othman, and F. Bagash, “A Biomedical Named Entity Recognition Using Machine Learning
Classifiers and Rich Feature Set,” International Journal of Computer Science and Network Security, vol. 17, no. 1, 2017.
[Online]. Available: https://www.semanticscholar.org/paper/A-Biomedical-Named-Entity-Recognition-Using-Machine-Al-
Hegami-Othman/e6e2d078d06fcc25f51a866b4e45f1400a4c56e2
[24] S. Zhao, “Named entity recognition in biomedical texts using an HMM model,” in Proceedings of the international joint
workshop on natural language processing in biomedicine and its applications, 2004, pp. 87–90. [Online]. Available:
https://aclanthology.org/W04-1216
[25] S. K. Saha, S. Sarkar, and P. Mitra, “Feature selection techniques for maximum entropy based biomedical named entity
recognition,” Journal of Biomedical Informatics, vol. 42, no. 5, pp. 905-916, 2009, doi: 10.1016/j.jbi.2008.12.012.
[26] T. Mathu and K. Raimond, “A novel deep learning architecture for drug named entity recognition”, Telecommunication,
Computing, Electronics and Control (TELKOMNIKA), vol. 19, No. 6, pp. 1884-189, 2021,
doi: 10.12928/telkomnika.v19i6.21667.
[27] R. Ramachandran and K. Arutchelvan, “Named entity recognition on bio-medical literature documents using hybrid based
approach,” Journal of Ambient Intelligence and Humanized Computing, 2021, doi: 10.1007/s12652-021-03078-z.
[28] R. Ramachandran and K. Arutchelvan, “ArRaNER: A novel named entity recognition model for biomedical literature
documents,” The Journal of Supercomputing, vol. 78, pp. 16498-16511, 2022, doi: 10.1007/s11227-022-04527-y.
[29] R. Kobus, C. Hundt, A. Müller, and B. Schmidt, “Accelerating metagenomic read classification on CUDA-enabled GPUs,” BMC
Bioinformatics, vol. 18, no. 11, 2017, doi: 10.1186/s12859-016-1434-6.
[30] C. Xie, S. L. Song, J. Wang, W. Zhang, and X. Fu, “Processing-in-Memory Enabled Graphics Processors for 3D Rendering,” 2017
IEEE International Symposium on High Performance Computer Architecture (HPCA), 2017, pp. 637-648,
doi: 10.1109/HPCA.2017.37.
[31] D. Zhao and Q. Chen, “Current Prediction Model of GPU Oriented to General Purpose Computing,” in IEEE Access, vol. 7,
pp. 127920-127931, 2019, doi: 10.1109/ACCESS.2019.2939256.
[32] Y. Song, E. Kim, G. G. Lee and B. -K. Yi, “POSBIOTM-NER in the shared task of BioNLP/NLPBA,” in Proceedings of the
International Joint Workshop on Natural Language Processing in Biomedicine and its Applications (NLPBA/BioNLP), 2004,
pp. 100–103. [Online]. Available: https://aclanthology.org/W04-1220.pdf
[33] K. Hakala and S. Pyysalo, “Biomedical Named Entity Recognition with Multilingual BERT,” in Proceedings of the 5th Workshop
on BioNLP Open Shared Tasks, 2019, pp. 56–61, doi: 10.18653/v1/D19-5709
[34] S. Wang et al., “kNN-NER: Named Entity Recognition with Nearest Neighbour Search,” Computation and Language, 2022,
doi: 10.48550/arXiv.2203.17103.
[35] A. Yenkikar, C. N. Babu, and D. J. Hemanth, “Figure 5: Illustration of support vector machine,” PeerJ Computer Science, 2020,
doi: 10.7717/peerj-cs.1100/fig-5.

TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 4, August 2023: 825-835
TELKOMNIKA Telecommun Comput El Control  835

[36] A. Yenkikar, C. N. Babu, and D. J. Hemanth, “Figure 4: Decision tree used for sentiment classification,” PeerJ Computer
Science, 2020, doi: 10.7717/peerj-cs.1100/fig-4.
[37] JavaTpoint, “K-Nearest Neighbour (KNN) Algorithm for Machine Learning-Javatpoint.” www.javatpoint.com.
https://www.javatpoint.com/k-nearest-neighbor-algorithm-for-machine-learning (accessed Nov. 15, 2022).

BIOGRAPHIES OF AUTHORS

Manish Bali received a B.E. degree in electronics and communications

engineering from Bangalore University, in 1995, and a master’s degree in computer science
from Ramaiah University of Advanced Sciences, in 2021. He has 28+ years of Industry
experience and has worked as an Adjunct Professor at Presidency University, Bengaluru. He is
currently a CSE research scholar, pursuing Ph.D. He has published 7 papers in various national
and International Journals and has a patent to his name His current research interests include
machine learning, deep learning, and emerging technologies. He can be contacted at email:
balimanish0@gmail.com.

Anandaraj Shanthi Pichandi received the received B. E (CSE) degree from

Madras University, Chennai in the year 2004, M.Tech (CSE) with Gold Medal from Dr. MGR
Educational and Research Institute, University in the year 2007 (Distinction with Honor), and
Ph.D. in August 2014. He is presently working as Professor under CSE at Presidency
University, Bengaluru. He has 15+ Years of Teaching Experience. His areas of interest include
Information security, Data Science, Machine Learning, and Networks. He has written two book
chapters in IGI Global Publications, USA, and authored five books. He has filed 7 patents and
published 50+ papers. He is a member of ISTE, CSI, Member of IACSIT, and Member of
IAENG. He can be contacted at email: anandsofttech@gmail.com.

Jude Hemanth Duraisamy received a B.E. degree in ECE from Bharathiar

University, in 2002, an M.E. degree in communication systems from Anna University, in
2006, and the Ph.D. degree from Karunya University, Coimbatore, India, in 2013. He is
currently working as an Associate Professor with the Department of ECE, at Karunya
University. He has authored more than 130 research articles in reputed SCIE-indexed
international journals and Scopus-indexed international conferences. His cumulative impact
factor is more than 160. He has published 33 edited books with reputed publishers, such as
Elsevier, Springer, and IET. His research interests include computational intelligence and
image processing. He can be contacted at email: judehemanth@karunya.edu.

Biomedical-named entity recognition using CUDA accelerated KNN algorithm (Manish Bali)