Mortality Outcome Prediction of Endemic Diseases in Davao City, Mindanao Region, Philippines using Unsupervised and Supervised Machine Learning

Dengue, Leptospirosis, and Acute Meningitis and Encephalitis Syndromes (AMES) are co-circulating endemic diseases in the Philippines that share overlapping early symptoms — sudden fever, headache, and myalgia — complicating differential diagnosis in resource-limited settings where laboratory confirmation is not always available. Machine learning (ML) has been applied to predict mortality outcomes and identify patient subgroups with similar clinical profiles, offering a data-driven complement to symptom-based diagnosis. This study uses patient-level surveillance data from the Davao City Health Office (2015–2025), enriched with NASA POWER daily climate variables (temperature, precipitation). Data preprocessing included feature engineering of onset-to-admission lag, wet/dry season classification, and administrative district mapping from barangay. Severe class imbalance — particularly in dengue (221:1 Alive-to-Died ratio) — was addressed through majority class undersampling (retaining all 225 deaths) and class weighting within each classifier. The $k$-prototypes algorithm was applied per disease for patient clustering, validated using silhouette scores on Gower distance matrices. Factor Analysis of Mixed Data (FAMD) was used for two-dimensional cluster visualization. Optimal clustering at $k$=2 for all three diseases revealed a primary patient division along seasonal lines (wet vs. dry season), with low-to-moderate silhouette scores (0.17–0.29) indicating meaningful but overlapping cluster structure consistent with the shared symptom profiles of these diseases. Cluster membership was evaluated as an additional classification feature. Five classifiers — Decision Tree, Random Forest, XGBoost, LightGBM, and CatBoost — were trained per disease using stratified 70/30 splits and Bayesian hyperparameter optimization (Optuna, 200 trials, PR-AUC objective). Results varied substantially by disease. Leptospirosis showed the strongest predictive signal, with CatBoost achieving the highest PR-AUC of 0.24 and F1-score of 0.28 for the Died class. AMES demonstrated moderate predictability, with Random Forest performing best (F1=0.23, PR-AUC=0.17). Dengue mortality proved the most difficult to predict across all classifiers, with PR-AUC below 0.07, consistent with the severity of class imbalance and the absence of clinical severity markers in routine surveillance records. Cluster membership provided marginal and inconsistent improvement across diseases. This study demonstrates a replicable cluster-then-classify framework for endemic disease mortality risk stratification using routinely collected surveillance data, with findings underscoring the need for richer clinical features to improve minority class detection in low-resource epidemiological settings.