Smart Health Monitoring Berbasis Machine Learning untuk Prediksi Risiko Stroke sebagai Upaya Pencegahan Dini Kesehatan Masyarakat

Abstract

Stroke is a leading cause of death and disability in Indonesia, with a rising prevalence that extends to the working-age population. Current stroke screening methods which remain largely manual and subjective fail to identify complex risk patterns within patient clinical and lifestyle data. This study develops a machine learning-based Smart Health Monitoring system integrated with a digital clinical dashboard to support early stroke risk detection in primary healthcare facilities, utilizing the Healthcare Stroke Data dataset (Kaggle, 2023) comprising 5,110 patient records. Three algorithms were systematically compared: Random Forest, XGBoost, and a Stacked Ensemble (using Random Forest and XGBoost as base learners and Logistic Regression as the meta-learner). Extreme class imbalance (95.1% vs. 4.9%) was addressed using k-Nearest Neighbors (k=5)-based SMOTE, applied in a leak-free manner across each cross-validation fold. Evaluation metrics included Accuracy, Sensitivity, Specificity, F1-Score, and AUC-ROC. Results indicate that XGBoost achieved the highest AUC-ROC (79.23%) with 99.69% Specificity, while Random Forest yielded the best Sensitivity (22.00%) with an AUC-ROC of 78.74%; the Stacked Ensemble provided a balanced performance (AUC-ROC 78.93%). Model interpretability was assessed via Explainable AI using SHAP value analysis, identifying age, average blood glucose, and BMI as the three most influential features. The study concludes with the development of a Streamlit-based digital clinical dashboard prototype featuring continuous risk prediction capabilities.