An interpretable machine learning framework for early-stage diabetes mellitus prediction using comparative classification models and SHAP

Abstract

Diabetes mellitus is a long-term metabolic disorder that typically goes unrecognised until it has already caused significant harm, making timely identification a critical clinical priority. Unfortunately, conventional diagnostic approaches frequently fall short in detecting the disease before it progresses, particularly within busy healthcare settings. To tackle this gap, the current study introduces a machine learning-based framework enhanced with explainability, built around a structured data preparation process that handles categorical encoding, numerical scaling, and minority class oversampling through the SMOTE technique. A pair of classification models, Logistic Regression and Random Forest, are trained, tested, and directly compared to assess their diagnostic reliability. The findings reveal that Random Forest consistently delivers stronger results, reaching a classification accuracy of 98%, which reflects its capacity to learn intricate relationships within real-world clinical data. To move beyond raw performance, SHAP analysis is integrated to shed light on how individual patient attributes shape each prediction outcome. The resulting system strikes a meaningful balance between diagnostic accuracy and model interpretability, positioning it as a trustworthy tool for assisting medical professionals in data-driven clinical decision-making.