Multi-Target Non-Communicable Disease Prediction via Hybrid Feature Selection and Supervised Learning

Abstract

Abstract— Hospital intake workflows routinely assess chronic illnesses in departmental silos, missing the shared physiological markers that connect systemic disorders. Consequently, clinical diagnoses face systematic delays, patients undergo repetitive blood panels, and care delivery costs escalate. We addressed these systemic bottlenecks by engineering a joint diagnostic framework that performs simultaneous risk screening across four major conditions: Cardiovascular Disease (CVD), Type-2 Diabetes Mellitus, Chronic Kidney Disease (CKD), and Liver Pathology. Raw electronic health records are sanitized using class-stratified median imputation for missing values, interquartile range (IQR) thresholds for artifact clipping, and min-max feature normalization. To resolve collinearity and isolate salient biological markers, a two-phase hybrid selector (MI-RFE) couples Mutual Information filtering with Recursive Feature Elimination driven by Random Forest importance metrics. Five supervised baselines—Support Vector Classifiers (SVC), Random Forest (RF), Logistic Regression, XGBoost, and a Multilayer Perceptron (MLP)—were trained and tuned on benchmark cohorts from the UCI Machine Learning Repository and Kaggle. Top-performing models were combined into an AUC-weighted soft-voting meta-ensemble. Evaluated via 10-fold stratified cross-validation, the ensemble obtained an aggregate classification accuracy of 89.46% (91.80% for CVD, 85.71% for Diabetes, 100.00% for CKD, and 80.34% for Liver Disorders) alongside a macro ROC-AUC of 0.925. Implemented inside a modular user interface, the tool generates holistic multi-organ risk profiles in 18.4 ms per record, demonstrating rapid decision-support utility for frontline triage.