Heart Stroke Prediction

Problem Statement

Stroke is a leading cause of death and disability worldwide. Early identification of at-risk patients enables preventative interventions. Clinical datasets are typically imbalanced — strokes are rare (~5% of cases) — requiring careful handling of class imbalance and a focus on recall to avoid missing true stroke cases.

Technical Approach

Model Development

• Random Forest: Ensemble of 200 decision trees with SMOTE oversampling to handle class imbalance. Achieved 94% recall on stroke cases, correctly identifying nearly all at-risk patients.
• Logistic Regression: Interpretable baseline for odds-ratio analysis of individual risk factors. Used for calibration and clinical validation.
• Decision Trees: Single-tree model pruned for maximum interpretability — used as a communication tool with clinical stakeholders.

Feature Engineering

Engineered 30+ features from patient health records: body mass index, average glucose level, age, hypertension status, heart disease history, smoking status, and work type. Created interaction features between age and glucose, and between BMI and hypertension, which significantly improved model performance. Identified top predictors: age (strongest signal), average glucose level, and hypertension status.

SHAP Explainability

Built an interactive SHAP dashboard for clinical use. Each prediction is accompanied by a waterfall plot showing exactly which factors pushed the risk score up or down. This transformed the model from a black box into a tool clinicians could trust and act on.

Key Results

• 94% recall on stroke cases — only 6% of strokes missed by the model
• Age, glucose, and hypertension identified as strongest predictors — consistent with medical literature
• SHAP dashboard enabled clinical validation and built stakeholder trust
• Model outperformed baseline stroke risk calculators (Framingham) by 18% in recall

Tech Stack