Adaptive digital filtering and TreeSHAP feature selection for non-invasive diabetes classification from photoplethysmography signals

Diabetes affects over 537 million adults globally, with early detection critical for preventing complications. This study develops a comprehensive machine learning framework for non-invasive diabetes screening using photoplethysmography (PPG) signals integrated with clinical metadata. We systematically evaluated ten digital filtering techniques across two independent datasets, namely the Guilin PPG-BP and Mazandaran datasets. The Adaptive Butterworth filter achieved optimal performance for Guilin data where SNR was 31.59 dB, while DC Removal with HP and LP Cascade excelled for Mazandaran where SNR was 22.50 dB, demonstrating dataset-specific filtering requirements. Comprehensive feature extraction generated 38 physiological features spanning time-domain, frequency-domain, morphological, and variability characteristics. TreeSHAP analysis identified seven key predictive features, with AI_Surrogate having SHAP importance of 0.926 and SDPPG ratios emerging as robust biomarkers. Five machine learning algorithms were optimized through hyperparameter tuning and SMOTE-based class balancing. On the Guilin dataset, the Ensemble Voting Classifier delivered the best results, achieving 90.91% accuracy, 100% precision, 77.78% recall, and a ROC-AUC of 0.915. On the Mazandaran dataset, models reached up to 90% accuracy. A Gradio-based web application demonstrated real-time prediction capability with a 0.043 s latency, providing clinicians with interpretable diabetes risk assessment. This framework establishes PPG-based screening as a viable non-invasive alternative for diabetes detection in resource-limited settings.