Stroke Prediction Analysis using Machine Learning Classifiers and Feature Technique

Abstract

Stroke is one of the fatal brain diseases that cause death in 3 to 10 hours. However, most stroke mortality can be prevented by identifying the nature of the stroke and reacting to it promptly through smart health systems. In this paper, a machine learning model is approached for predicting the existence of stroke of a patient where the Random forest classifier outperforms the state-of-the-art models, including Logistic Regression, Decision Tree Classifier (DTC), K-NN. We conduct the experiments on datasets which has 5110 observations with 12 attributes. We also applied EDA for preprocessing and feature techniques for balancing the datasets. Finally, a cloud-based mobile app collects user data to analyze and provide the possibility of stroke for alerting the person with the accuracy of precision 96%, recall 96%, and F1-score 96%. This user-friendly system can be a lifesaver as the person gets an essential warning very easily by providing very little information from anywhere with a mobile device.

Full text

International Journal of Electronics and Communications System
Volume 1, Issue 2, 17-22.
ISSN: 2798-2610
http://ejournal.radenintan.ac.id/index.php/IJECS/index
DOI: 10.24042/ijecs.v1i2.10393
 Corresponding author:
Md. Monirul Islam, Uttara University, Dhaka-1230, BANGLADESH. monirul@uttarauniversity.edu.bd
© 2021 The Author(s). Open Access. This article is under the CC BY SA license (https://creativecommons.org/licenses/by-sa/4.0/)
Stroke prediction analysis using machine learning classifiers and
feature technique
Md. Monirul Islam *
Uttara University, Dhaka-1230,
BANGLADESH
Sharmin Akter
Atish Dipankar University of Science
& Technology, Dhaka-1230,
BANGLADESH
Md. Rokunojjaman
Chongqing University of Technology,
Chongqing 400054, CHINA
Jahid Hasan Rony
Dhaka University of Engineering and
Technology, Gazipur, Gazipur-1700,
BANGLADESH
Al Amin
Chongqing University of
Technology, Chongqing 400054,
CHINA
Susmita Kar
Dhaka University of Engineering and
Technology, Gazipur, Gazipur-1700,
BANGLADESH
Article Info
Abstract
Article history:
Received: October 4, 2021
Revised: December 8, 2021
Accepted: December 15 2021
Stroke is one of the fatal brain diseases that cause death in 3 to 10 hours.
However, most stroke mortality can be prevented by identifying the nature of the
stroke and reacting to it promptly through smart health systems. In this paper, a
machine learning model is approached for predicting the existence of stroke of a
patient where the Random forest classifier outperforms the state-of-the-art
models, including Logistic Regression, Decision Tree Classifier (DTC), K-NN. We
conduct the experiments on datasets which has 5110 observations with 12
attributes. We also applied EDA for preprocessing and feature techniques for
balancing the datasets. Finally, a cloud-based mobile app collects user data to
analyze and provide the possibility of stroke for alerting the person with the
accuracy of precision 96%, recall 96%, and F1-score 96%. This user-friendly
system can be a lifesaver as the person gets an essential warning very easily by
providing very little information from anywhere with a mobile device.
Keywords:
Feature Technique,
Random Forest Classifier,
Stroke disease
To cite this article: M. M. Islam, S. Akter, M. Rokunojjaman, J. H. Rony, A. Al Amin, and S. Kar, “Stroke
Prediction Analysis using Machine Learning Classifiers and Feature Technique,” Int. J. Electron. Commun. Syst.,
vol. 1, no. 2, 17-22, 2021.
INTRODUCTION
A stroke happens to interrupt blood flow
to a portion of your brain [1]. A loss of blood
circulation to some brain areas causes a
stroke, which is also known as a brain attack
[2]. Furthermore, clot blocking is the major
cause of stroke in the brain (thrombosis). The
blood vessel delivers the brain portion and is
subsequently run down of blood and oxygen.
The brain cells expire as an outcome of the
lack of blood and O2, and the part of the body
it regulates ceases working [3]. Death and
disability happen for stroke in the United
States badly. Ischemic embolic and
hemorrhagic strokes cause the majority of
strokes. An ischemic embolic stroke happens
when a blood clot exits the patient's brain,
travels through the circulatory system, and
becomes lodged in smaller brain arteries.
Another type is hemorrhagic stroke, which
occurs when leaks or ruptures a blood vessel
in the brain. [4]. The use of various predictive
indicators to predict the outcome of a stroke
could help doctors identify high-risk patients
and reduce morbidity. Overweight, physical
inactivity, diabetics, and other parameters
such as age, sex, race can be used to predict the
possibility of stroke. On the other hand,
machine learning offers an option, particularly
for large-scale multi-institutional data that

18 Int. J. Electron. Commun. Syst, 1 (2) (2021) 17-22
may be readily included in a forecast [5] based
on freshly available data.
Smartphones can play a vital role in
establishing a between the healthcare system
and the global population. A mobile app is very
user-friendly and popular in the current world.
According to statistics, there are more than 3.2
billion smartphone users. As a result, a mobile
app could be one of the most popular and
effective mediums. In case of stroke, a disease
avoidable through awareness, a smartphone
could be the easier way to reach people.
Machine learners have various
applications expanding within the study of
bioinformatics, a subfield of artificial
intelligence which includes improving
calculations to discover how projections are
dependent on information. Bioinformatics
manages computational and numerical
approaches for comprehension and
manipulating natural data. Six natural
environments have been subjected to machine
learning. To assist in the analysis of stroke,
machine learning algorithms for examining
neuroimaging data are used. The diagnosis and
treatment of stroke disease in underdeveloped
countries is extremely difficult due to a lack of
diagnostic technologies and a scarcity of
doctors and other resources that impede the
accurate prediction and treatment of heart
patients. Recently, computer technology and
machine learning approaches have been
developed with this goal in mind to improve
the system's ability to assist doctors in the
initial phases of disease decision-making [6].
Our motivation is to benefit stroke prediction
to prevent casualty and ensure accessibility for
everyone.
Among various studies in this area, in [7],
stroke prediction directions were designed as
risk assessment and web-based cooperative
Java applets. These Java applets enable risk
calculations and can be run interactively with
any web browser that supports Java 1.1. With
this method, patient data can simply be
entered into a computer that uses complex
statistical models to produce instant
calculations of risk scores. Authors in [8]
examined the utility of the echo planer
magnetic perfusion imaging and diffusion-
weighted imaging in predicting stock with a
critical hemispheric infraction. In [9], type 2
diabetes patients have an increased risk of
stroke. In this approach, they examined the
stroke predictors and effects of atorvastatin on
certain stroke subtypes in type 2 diabetes in
the collaborative atorvastatin diabetes study,
which used Cox regression models to evaluate
atorva's impact statins on stroke, and assess
the risks associated with stroke and
underlying stroke. The authors determined
how many self-measures of blood pressure
they took home compared to their predictive
value for the risk of a stroke. In [10], they have
designed and compared several methods of
learning machines, which can predict the
result of endovascular intervention in the
previous histosa circulation. The authors
developed a CPS to detect the appearance of
the patients who are at high risk or survived a
stroke before [11]. CPS developed send data
registered by the doctor and warned to find a
stroke.
Furthermore, the proposed system works
in data purchased by the patient's brain
electroencephalography sensor. The authors
have developed a model learning model (ML)
calculated by threshold (ML) to predict the
tracking infarction in patients with acute
ischemic stroke [12]. The author determined
the optimal number of self-measurements of
blood pressure at home based on its predictive
value for stroke risk. Therefore, the Cox
proportional hazard regression model [13]
was used to investigate the prognostic
significance of blood pressure for the risk of
stroke, which was adjusted for possible
confounding factors.
The author has developed a very accurate
and highly interpretable predictive model.
These predictive models will be provided in
the form of sparse decision lists [14], which
are derived from a series of if . . .Then . . .
Statements where the if statement defines a
set of feature partitions and the then
statement corresponds to the predicted result
of interest. In [15], the authors predicted
stroke occurrence using a large population-
based EMC database and also compared DNN
with three other ML methods. The authors
compared the Cox proportional hazard model
with an automatic stroke prediction approach
based on a cardiovascular health study (CHS)
dataset [16]. The author developed a hybrid
machine learning method to predict stroke
based on incomplete and unbalanced
physiological data for clinical diagnosis [17].
Using this method, the whole process involves

Int. J. Electron. Commun. Syst, 1 (2) (2021) 17-22 19
two steps. First, use random forest regression
to estimate missing values before
classification. Secondly, automatic
hyperparameter optimization (AutoHPO)
based on deep neural networks (DNN)
predicts stroke on unbalanced data sets. The
author applies machine learning principles to
existing large data sets to effectively predict
strokes based on potentially changeable risk
factors [18] to develop applications that
provide personalized warnings and related
information based on each user's stroke risk
level Lifestyle of stroke risk factors. The
authors raised the hypothesis that the degree
of stenosis, the irregularity of the plaque's
surface, eColity, and consistency, complicated
in a total score of risk (TPR), are predictors of
the ischemic blow [19]. Three classification
algorithms that include the decision-making
tree, naive bayes, and neuronal network are
used to predict the stretch based on models
higher than general statistics and obtained an
adequate model for identification [20].
This paper proposed Stroke prediction
analysis using a machine learning algorithm
using a healthcare dataset, including various
kinds of risk factors.
The rest of the paper is organized as
tracks: the methodology is stated in the next
section. Study outcome and discussion are in
the results and discussion section. Finally, the
paper concludes with future scope.
METHOD
Figure 1 shows the detailed block diagram
of the proposed methodology.
Figure 1. Block Diagram of the Proposed
Methodology
Dataset Description
The utilized dataset [21] contains 5110
observations with 12 attributes. The attributes
are gender, age, hypertension, heart_disease,
ever_married, work_type, Residence type,
average glucose_level, BMI, smoking_status,
and stroke. Stroke is a dependent variable, and
others are independent variables.
Exploratory Data Analysis (EDA)
EDA often uses data visualization
approaches to analyze and examine data sets
and summarize their key characteristics. It can
help determine how best data sources can be
handled to get the needed answers, facilitating
the finding of patterns, spot anomalies,
hypotheses, or assume checks for data
scientists. In this part, we defined the missing
values, data counts, dropped the id column,
exploring each variable.
Feature Techniques
Feature engineering means transforming
raw data into features that better signify the
predictive models' underlying problem and
improve model accuracy in unsightly data.
Many techniques can be employed, including
NearMiss, SMOTE, Tomak Links, etc. This
paper utilized the synthetic minority over-
sampling technique (SMOTE) after
preprocessing the datasets in the EDA step.
The target variable has 201 stroke occurrences
and 4908 non-occurrence patients.
Machine Learning Analysis
This paper utilized various machine
learning (ML) models containing Naïve Bayes,
Random Forest, Ada Boost Algorithm. Among
them, the Random Forest model outperforms
the best accuracy. So Random forest model is
described here.
Random Forest (RF)
RF is a supervised learning algorithm. It
creates a "forest" from a series of decision
trees that are usually trained using a "bagging"
process. The basic premise of the bagging
method is that combining different learning
models can improve the overall result. The
advantage of RF is that it can solve
classification and regression problems that
make up most of the existing machine learning
systems. Decision trees or bagging classifiers
have almost the same hyperparameters as
random forests. Fortunately, you can use
random forest classifiers instead of combining
decision trees and bagging classifiers. You can
use the algorithm's suppressor to handle the
regression task of the random forest. The RF
adds additional unpredictability to the model

20 Int. J. Electron. Commun. Syst, 1 (2) (2021) 17-22
as the tree develops. Instead of the most
relevant feature, splitting a node looks for the
optimal function in a random selection of
features. Hence, many types lead to better
models. Therefore, the algorithm for splitting
nodes in the random forest only considers a
random subset of features. Instead of looking
for the best possible threshold, you can make
the tree more random by using random
thresholds for each function [22].
The random forest training algorithm uses
the general aggregation bootstrap technique,
or bags, for train trees students. Figure 2
demonstrates the concept of a random forest
model where Tree 1 and Tree 2 associate Class
X. So, the majority vote/predicted output is
Class X.
Figure 2. Random Forest
Predictions for unseen samples I can be
produced after training by summing the
predictions from all of the separate regression
trees on i':

 



or by taking the majority vote in the case of
classification trees [24].
User Interface
User data are collected through mobile
apps. Users input gender, age, work_type,
heart_disease, hypertension, ever_married,
Residence_type, BMI, avg glucose level,
smoking_status through the mobile app. In
Figure 3, the mobile app interface is shown.
User data are stored in the cloud Firestore
database. After the processing, the result is
stored in the Firestore and shown on the user
end.
Figure 3. Mobile App
RESULTS AND DISCUSSION
Python programming language is used to
classify the proposed model and describe
other models for data analysis. The instrument
is very useful for analysis and includes
different methods. For each model species, we
have used 20% of the values for testing and
80% for training. We take precision, recall, and
f1-score as performance metrics.
Precision (P): P is the ratio of the positive
cases correctly predicted to the positive cases.
The low false positive rate refers to high
accuracy. It is a measure of a classifier's
accuracy. In equation 1, it is defined
mathematically.
  
 

Int. J. Electron. Commun. Syst, 1 (2) (2021) 17-22 21
Recall (R): R refers to the ratio of positive
cases correctly predicted to all positive
classification cases. It is a measure of a
classification's completeness. In Equation 2, R
is defined mathematically.
  
 
F1-Score: is an average weighted accuracy and
recall. F1, if there is an inconsistent class
distribution in the data set, is usually more
useful than precision. It is displayed in
equation 3 mathematically, and the result of
accuracies can be seen in table 1.
     
   
Table 1. Result Accuracies
ML Model
Accuracies (%)
Preci
sion
Rec
all
F1-
Score
Logistic
Regression [23]
87
87
87
DTC [24]
93
93
93
K-NN [25]
90
91
90
Random Forest
(proposed)
96
96
96
Table 1 describes the result of accuracies.
The random forest model gives the highest
accuracies in all performance metrics as 96%.
K-NN achieves 3rd place as holding 90%
performance metrics, DTC stays 2nd position
as 93% accuracy, and logistic regression
receives 87% accuracy.
CONCLUSION
This paper presented a machine learning
approach to the stroke dataset. The Random
Forest models showed the best accuracy as
precision 96%, recall 96%, and F1-score 96%,
outperforming the state-of-art models
including logistic regression, decision tree
classifier, and K-NN. The utilized dataset is
imbalanced, therefore, SMOTE feature
engineering is used to process the data. In the
future, we will plan to analyze the dataset
using deep learning methods and try to
enhance the accuracy.
REFERENCES
[1] “Stroke (Causes, Symptoms, and
Complications) - Assignment Point.”
https://www.assignmentpoint.com/science
/medical/stroke-causes-symptoms-and
(accessed May 16, 2021).
[2] Mayo Clinic, “Stroke - Symptoms and
Causes,” Mayo Clinic, Nov. 06, 2020.
https://www.mayoclinic.org/diseases-
conditions/stroke/symptoms-causes/syc-
20350113.
[3] B. Wedro, “Stroke Warning Signs,
Symptoms, Treatment, Types & Causes,”
MedicineNet, 2019.
https://www.medicinenet.com/stroke_sym
ptoms_and_treatment/article.htm
[4] H. Rodgers, “Stroke,” Neurological
Rehabilitation, pp. 427–433, 2013, doi:
10.1016/b978-0-444-52901-5.00036-8
[5] H. Asadi, R. Dowling, B. Yan, and P. Mitchell,
“Machine Learning for Outcome Prediction
of Acute Ischemic Stroke Post Intra-Arterial
Therapy,” PLoS ONE, vol. 9, no. 2, p. e88225,
Feb. 2014, doi:
10.1371/journal.pone.0088225.
[6] P. Govindarajan, R. K. Soundarapandian, A.
H. Gandomi, R. Patan, P. Jayaraman, and R.
Manikandan, “Classification of stroke
disease using machine learning algorithms,”
Neural Computing and Applications, vol. 32,
no. 3, pp. 817–828, Jan. 2019, doi:
10.1007/s00521-019-04041-y.
[7] T. Lumley, R. A. Kronmal, M. Cushman, T. A.
Manolio, and S. Goldstein, “A stroke
prediction score in the elderly,” Journal of
Clinical Epidemiology, vol. 55, no. 2, pp.
129–136, Feb. 2002, doi: 10.1016/s0895-
4356(01)00434-6.
[8] P. A. Barber et al., “Prediction of stroke
outcome with echoplanar perfusion- and
diffusion-weighted MRI,” Neurology, vol. 51,
no. 2, pp. 418–426, Aug. 1998, doi:
10.1212/wnl.51.2.418.
[9] G. A. Hitman et al., “Stroke prediction and
stroke prevention with atorvastatin in the
Collaborative Atorvastatin Diabetes Study
(CARDS),” Diabetic Medicine, vol. 24, no. 12,
pp. 1313–1321, Dec. 2007, doi:
10.1111/j.1464-5491.2007.02268.x.
[10] H. Asadi, R. Dowling, B. Yan, and P. Mitchell,
“Machine Learning for Outcome Prediction
of Acute Ischemic Stroke Post Intra-Arterial
Therapy,” PLoS ONE, vol. 9, no. 2, p. e88225,

22 Int. J. Electron. Commun. Syst, 1 (2) (2021) 17-22
Feb. 2014, doi:
10.1371/journal.pone.0088225.
[11] A. Laghari, Z. A. Memon, S. Ullah and I.
Hussain, "Cyber Physical System for Stroke
Detection," in IEEE Access, vol. 6, pp.
37444-37453, 2018, doi:
10.1109/ACCESS.2018.2851540.
[12] H. Kuang et al., "Computed Tomography
Perfusion–Based Machine Learning Model
Better Predicts Follow-Up Infarction in
Patients With Acute Ischemic Stroke,"
stroke, vol. 52, no. 1, pp. 223–231, Jan.
2021, doi: 10.1161/strokeaha.120.030092.
[13] T. Ohkubo et al., “How many times should
blood pressure be measured at home for
better prediction of stroke risk? Ten-year
follow-up results from the Ohasama study,”
Journal of Hypertension, vol. 22, no. 6, pp.
1099–1104, Jun. 2004, doi:
10.1097/00004872-200406000-00009.
[14] B. Letham, C. Rudin, T. H. McCormick, and D.
Madigan, “Interpretable classifiers using
rules and Bayesian analysis: Building a
better stroke prediction model,” Annals of
Applied Statistics, vol. 9, no. 3, pp. 1350–
1371, Sep. 2015, doi: 10.1214/15-
AOAS848.
[15] C. Hung, W. Chen, P. Lai, C. Lin and C. Lee,
"Comparing deep neural network and other
machine learning algorithms for stroke
prediction in a large-scale population-based
electronic medical claims database," 2017
39th Annual International Conference of
the IEEE Engineering in Medicine and
Biology Society (EMBC), 2017, pp. 3110-
3113, doi: 10.1109/EMBC.2017.8037515.
[16] A. Khosla, Y. Cao, C. C.-Y. Lin, H.-K. Chiu, J.
Hu, and H. Lee, “An integrated machine
learning approach to stroke prediction,”
Proceedings of the 16th ACM SIGKDD
international conference on Knowledge
discovery and data mining - KDD ’10, 2010,
doi: 10.1145/1835804.1835830.
[17] T. Liu, W. Fan, and C. Wu, “A hybrid machine
learning approach to cerebral stroke
prediction based on imbalanced medical
dataset,” Artificial Intelligence in Medicine,
vol. 101, p. 101723, Nov. 2019, doi:
10.1016/j.artmed.2019.101723.
[18] M. Monteiro et al., "Using Machine Learning
to Improve the Prediction of Functional
Outcome in Ischemic Stroke Patients," in
IEEE/ACM Transactions on Computational
Biology and Bioinformatics, vol. 15, no. 6,
pp. 1953-1959, 1 Nov.-Dec. 2018, doi:
10.1109/TCBB.2018.2811471.
[19] P. Prati et al., “Carotid Plaque Morphology
Improves Stroke Risk Prediction:
Usefulness of a New Ultrasonographic
Score,” Cerebrovascular Diseases, vol. 31,
no. 3, pp. 300–304, 2011, doi:
10.1159/000320852.
[20] T. Kansadub, S. Thammaboosadee, S.
Kiattisin and C. Jalayondeja, "Stroke risk
prediction model based on demographic
data," 2015 8th Biomedical Engineering
International Conference (BMEiCON), 2015,
pp. 1-3, doi:
10.1109/BMEiCON.2015.7399556.
[21] “Stroke Prediction Dataset,”
kaggle.com.https://www.kaggle.com/fedes
oriano/stroke-prediction-dataset.
[22] MM. Islam, MA. Kashem, J. Uddin, “Fish
survival prediction in an aquatic
environment using random forest model,”
IAES International Journal of Artificial
Intelligence (IJ-AI), vol. 10, no. 3, pp. 614–
622, 2021, doi: 10.11591/ijai.v10.i3.pp614-
622.
[23] Md. M. Islam, J. Uddin, M. A. Kashem, F.
Rabbi, and Md. W. Hasnat, “Design and
Implementation of an IoT System for
Predicting Aqua Fisheries Using Arduino
and KNN,” Intelligent Human Computer
Interaction, pp. 108–118, 2021, doi:
10.1007/978-3-030-68452-5_11.
[24] A. Esmael, M. Elsherief, and K. Eltoukhy,
“Predictive Value of the Alberta Stroke
Program Early CT Score (ASPECTS) in the
Outcome of the Acute Ischemic Stroke and
Its Correlation with Stroke Subtypes,
NIHSS, and Cognitive Impairment,” Stroke
Research and Treatment, vol. 2021, pp. 1–
10, Jan. 2021, doi: 10.1155/2021/5935170.
[25] C. Y. Baek, W. N. Chang, B. Y. Park, K. B. Lee,
K. Y. Kang, and M. R. Choi, “Effects of dual-
task gait treadmill training on gait ability,
dual-task interference, and fall efficacy in
people with stroke: A Randomized
Controlled Trial,” Physical Therapy, Feb.
2021, doi: 10.1093/ptj/pzab067.