Unlocking stroke prediction: Harnessing projection-based statistical feature extraction with ML algorithms.

Area of Science:

• Medical Informatics
• Biostatistics
• Machine Learning in Healthcare

Background:

• Non-communicable diseases cause 71% of global deaths, with stroke being a leading cause.
• Early stroke recognition and risk factor identification are crucial for prevention and management.
• Traditional machine learning models face challenges with high-dimensional data and varying data scales in stroke prediction.

Purpose of the Study:

• To identify significant stroke risk factors using rigorous statistical tests.
• To propose novel feature representation techniques (PCA-FA and FPCA) for enhanced machine learning model performance.
• To develop and validate a robust stroke prediction model.

Main Methods:

• Utilized a dataset of 5110 patient records with clinical, lifestyle, and genetic attributes.
• Applied chi-square and independent sample t-tests to identify risk factors (age, heart disease, hypertension, etc.).
• Developed predictive models using Random Forests with PCA-FA feature extraction, validated with a stacking ensemble algorithm.

Main Results:

• Identified age, heart_disease, hypertension, work_type, ever_married, bmi, and smoking_status as significant stroke risk factors (P<0.05).
• The Random Forests model with PCA-FA achieved a 92.55% accuracy and 98.15% AUC score.
• Achieved prediction accuracy improvements of 2.19% to 19.03% compared to existing methods.

Conclusions:

• The proposed PCA-FA feature extraction significantly enhances machine learning model performance for stroke risk prediction.
• The developed model offers a robust and reproducible tool for identifying individuals at high risk of stroke.
• A web-based application was created to assist physicians in stroke risk diagnosis.