A Machine Learning Framework for Stroke Prediction: Balancing Precision and Recall in Healthcare Analytics

Stroke prediction is an essential challenge in health-care, where the timely identification of at-risk individuals can considerably improve patient outcomes and reduce long-term disabilities. In this work, we propose a comprehensive machine learning framework to predict stroke occurrence using a Kaggle dataset that integrates a wide array of health and demographic features. The dataset poses significant challenges due to a severe class imbalance, with stroke cases representing only 4.9 % of the overall samples. We evaluate multiple predictive models including Logistic Regression, Support Vector Machines, Random Forests, K-Nearest Neighbors, Naive Bayes, and an LSTM-based Neural Network specifically adapted to process tabular data. Our innovative approach involves reshaping the traditional tabular inputs into a pseudo-sequential format, thereby enabling the LSTM model to capture subtle relationships between risk factors such as age, blood pressure, glucose levels, and lifestyle habits. Experimental results, measured by robust metrics such as the F2-score and Matthews Correlation Coefficient (MCC), indicate that the LSTM-based model achieves superior performance compared to conventional classifiers, particularly in balancing recall and precision. Overall, our framework not only enhances stroke prediction accuracy, but also provides valuable insights into the interdependencies of various risk factors, offering a promising tool for healthcare professionals to implement early intervention strategies and ultimately improve patient care.