Harnessing Deep Neural Architectures for Precise Diabetes Risk Forecasting: An Interpretable Framework with Multimodal Biomarker Insights

This study presents a robust deep learning framework for predicting diabetes onset using a comprehensive dataset of 100,000 patient records, encompassing demographic and clinical attributes such as age, BMI, HbA1c level, and blood glucose. Data preprocessing involved encoding categorical variables, scaling numerical features, and outlier mitigation via the interquartile range method. Three advanced neural architectures—Deep Neural Network (DNN), 1D Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM)—were implemented and evaluated. The DNN achieved the highest performance, with 97% accuracy and 0.96 ROC-AUC, followed by CNN (96%) and LSTM (95%). Visualizations, including correlation heatmaps, loss curves, ROC plots, confusion matrices, and SHAP feature importance, elucidated key predictors like HbA1c and BMI. The framework addresses challenges in model interpretability and class imbalance, offering a scalable tool for early diabetes detection. Results highlight the efficacy of deep learning in capturing complex health patterns, with implications for proactive screening and personalized interventions in clinical settings.