Janciel Fidel M. Pedrano, Ralph Matthew M. Gobui, Angie M. Ceniza-Canillo
Stroke remains a critical public health concern in the Philippines, often leading to severe disability or death due to delayed detection and intervention. This study evaluates three supervised machine learning models for stroke risk classification: (1) Classical Logistic Regression, (2) a proposed Deep Logistic Regression (DLR) model that combines a linear logistic layer with hidden layers to capture nonlinear interactions, and (3) a Multilayer Perceptron (MLP). A publicly available dataset of 40,909 individual records with clinical, lifestyle, and demographic features was used. Temporal dependencies were explored through lagged predictors identified via the PCMCI algorithm and applied exclusively to the DLR model. Model performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and the Youden Index. The MLP achieved the best performance (95% accuracy, AUC 0.986, Youden Index 0.9025, stroke recall 98.3%), outperforming Classical Logistic Regression (69% accuracy, AUC 0.751) and DLR (64% accuracy, AUC 0.669). Temporal features did not improve performance, likely due to noise introduced by synthetic lag construction. Results highlight the promise of neural network–based models for stroke risk assessment but also underscore the need for external validation to confirm generalizability. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
University of San Carlos, Cebu City, Philippines
