Evaluating Machine Learning Models for Stroke Prediction Based on Clinical Variables

Patrick O. Akinwumi^1*Stephen Ojo^2*Thomas I Nathaniel^3*James Wanliss²Olukayode Karunwi⁴Mercy Sulaiman⁵

• ¹Clemson University, Clemson, United States
• ²College of Engineering, Anderson University, Anderson, SC, United States
• ³School of Medicine Greenville, University of South Carolina, Columbia, SC, United States
• ⁴College of Arts and Sciences, Anderson University, Anderson, SC, United States
• ⁵Department of Psychology University of New Hampshire, Durham, NH, United States

Stroke remains one of the leading causes of global mortality and long-term disability, driving the urgent need for accurate and early risk prediction tools. Traditional models such as the Framingham Stroke Risk Score have provided foundational insights into stroke prevention but are constrained by linear assumptions and limited adaptability to complex real-world data. In contrast, machine learning (ML) techniques offer the ability to model non-linear relationships and interactions among diverse clinical and demographic variables, supporting more personalized and flexible risk prediction. This study evaluates five supervised ML algorithms, Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine (SVM), and K-Nearest Neighbours (KNN), using a publicly available dataset from Kaggle. Following class imbalance correction, models were assessed using multiple metrics including accuracy, ROC-AUC, and confusion matrices. Logistic Regression and Gradient Boosting achieved the highest accuracy (95.11%) and ROC-AUC (0.836), although all models demonstrated poor recall, reflecting challenges in identifying rare stroke cases. Feature importance analysis using the Random Forest model identified age, average glucose level, and BMI as the most influential predictors of stroke, aligning with the Metabolic Syndrome Hypothesis and previous epidemiological findings. These findings underscore both the promise and current limitations of ML in stroke risk prediction and highlight the need for future research leveraging multimodal datasets and advanced algorithmic strategies to enhance sensitivity and clinical utility.