Comparison of Machine Learning and Deep Learning Techniques for Stroke Prediction

Abstract

The early diagnosis and management of diseases in medicine have become critically important in today's world. This comparative study focuses on evaluating models developed using various machine learning and deep learning techniques (Logistic Regression, Decision Trees, Random Forest, Support Vector Machine, CNN 1-D, LSTM , BİLSTM ) to determine the risk of stroke. The obtained highest accuracy values are as follows: LR (0.96), DT (0.95), RF (0.95), SVM (0.96), CNN (0.9442), LSTM (0.9442), BİLSTM (0.9442). The study analyzes a dataset containing various clinical parameters (age, gender, hypertension, heart disease, marital status, occupation type, residence type, average glucose level, BMI, and smoking) using the healthcare-dataset-stroke-data/Fedesoriano. This comparative research aims to make a significant contribution to the field of health by evaluating the effectiveness of different machine learning and deep learning models in determining the risk of stroke.

Dataset

Dataset Properties

The data set used contains 5510 data and includes individuals with the following characteristics: https://www.kaggle.com/datasets/fedesoriano/stroke-prediction-dataset

id	Gender	Age	Hypertension	Heart Disease	Marital Status	Occupation	Residence Type	Average Glucose Level	Body Mass Index	Smoking Status	Stroke
1	Male	45	0	1	No	Private	City	87.5	28.3	have smoked before	0
2	Female	35	1	0	Yes	Government	Rural	110.2	32.1	never smoked	1
...	...	...	...	...	...	...	...	...	...	...	...

Methods

Machine Learning: Logistic Regression, Decision Trees, Random Forest, SVM

Deep Learning: CNN 1-D, LSTM, Bi-LSTM

Results

1-) Accuracy Values at Different Testing and Training Rates

Accuracy	Logistic Regression	Decision Tree	Random Forest	SVM
10% Test - 90% Training	0.9491	0.8884	0.9491	0.9491
20% Test - 80% Training	0.9500	0.9080	0.9540	0.9559
30% Test - 70% Training	0.9562	0.9145	0.9543	0.9562
50% Test - 50% Training	0.9495	0.9021	0.9487	0.9495

2-) Accuracy Values Changing According to Outliers

Accuracy Values	Logistic Regression	Decision Tree	Random Forest	SVM
Before Removing Outliers	0.9491	0.8884	0.9491	0.9491
After Removing Outliers	0.9557	0.9305	0.9578	0.9578
Taking Mean of Outliers for the Dataset	0.9491	0.8884	0.9491	0.9491

3-) Accuracy Values After Applying Lasso and PCA

Accuracy Values	Logistic Regression	Decision Tree Model	Random Forest Model	SVM
Accuracy 1	0.9557	0.9305	0.9578	0.9578
Accuracy 2 (LASSO)	0.96	0.95	0.95	0.96
Accuracy 3 (PCA)	0.9586	0.9158	0.9579	0.9586

4-) Ensemble Learning Accuracy Values

Ensemble	Accuracy Values
LR, DT, RF, SVM	0.9585
LR, DT, RF, SVM, with LASSO	0.96
LR, DT, RF, SVM, with PCA	0.9586
CNN, LSTM, BiLSTM	0.9442

5-) Accuracy Values Varying at Different Epoch Values

	CNN	LSTM	BiLSTM
Epoch-10	0.9431	0.9442	0.9421
Epoch-50	0.9442	0.9442	0.9442

6-) Accuracy Values of Hybrid Models in Deep Learning Algorithms

Model	Accuracy Value
CNN, LSTM	0.9442
CNN, BiLSTM	0.9442
CNN, LSTM, BiLSTM	0.9442

Conclusion

In this study, the effectiveness of machine learning and deep learning algorithms used for stroke prediction was comprehensively evaluated and the accuracy rates between various models were compared. According to the research results, machine learning algorithms such as Logistic Regression, Support Vector Machines, Random Forest, Decision Trees and deep learning algorithms such as Convolutional Neural Network, Long-Short Term Memory and Bidirectional Long Short Term Memory have significant potential in stroke prediction. Additionally, dimension reduction techniques such as Lasso and Principal Component Analysis (PCA) and Ensemble Learning methods were used to increase the accuracy and generalizability of the model. As a result of the studies, the highest accuracy rate was obtained in the post-LASSO, Logistic Regression, post-LASSO Support Vector Machines and Machine Learning Algorithms Ensemble Learning (LR, KA, RO, SVM) method, with an accuracy rate of 0.96.