Diabetes Prediction using PPG Signal and Clinical data with XGBoost Classification

Abstract

Diabetes mellitus (DM) is a long-term metabolic disease marked by high blood sugar levels brought
on by either inadequate insulin synthesis or inefficient insulin usage. It is a significant worldwide
health issue linked to serious consequences like neuropathy, renal failure, and cardiovascular
disorders. For diabetes to be well managed and long-term problems to be avoided, early detection
and precise prediction are essential. This work combines advanced machine learning techniques
with physiological signal analysis to propose an effective model named as PPG-XGei for the
prediction of diabetes mellitus. The main input of the proposed model is Photoplethysmography
(PPG) digitized signals, which are processed using a deep representation model (PapaGei) to
produce PPG embeddings. To provide a complete fused representation, these learnt features are
combined with 86 hand-crafted PPG features and 13 clinical information elements. The model can
capture both physiological patterns and clinical signs related to diabetes due to this multimodal
feature fusion. For the final prediction, the merged feature set is fed to XGBoost classifier. With an
accuracy of 96.34%, precision of 94.12%, recall of 97.79%, F1-Score of 95.92%, and an AUC of
98.3%, the experimental findings show excellent performance. Overall, the work shows how well
hand crafted features and clinical data can be combined with deep learning based embeddings to
provide scalable and accurate diabetes prediction.