Physiological PPG Signals with Clinical Attributes revealing Blood Glucose Variation Dynamics

Abstract

Diabetes mellitus (DM) remains a major global health challenge, affecting over 537 million adults worldwide, with many cases undiagnosed and at risk of severe complications. Conventional blood glucose monitoring (BGM) methods rely on invasive finger-prick (FP) techniques, which are uncomfortable and unsuitable for continuous glucose monitoring (CGM). To address these limitations, this study proposes a supervised machine learning (ML) framework for non-invasive glucose estimation using Photoplethysmography (PPG) signals combined with patient demographic data. The dataset includes PPG recordings along with attributes such as age, weight, and Body Mass Index (BMI). Preprocessing involves denoising, normalization, and feature extraction (FE) to ensure high-quality inputs for model training. Several baseline regression models, including Linear Regression (LR), Ridge Regression (RR), Lasso Regression (Lasso), and Decision Tree (DT), are implemented to establish performance benchmarks. The proposed approach utilizes Natural Gradient Boosting (NGB), which effectively captures complex non-linear relationships and provides uncertainty quantification (UQ). Experimental results demonstrate that NGB outperforms baseline methods in terms of accuracy, robustness, and predictive reliability. Additionally, the model is integrated into a Flask-based webbased (WB) application, enabling both patients and healthcare professionals (HCPs) to access real-time glucose predictions. This supports early detection, improved glycemic control (GC), and better diabetes management. The proposed system offers a scalable, intelligent, and non-invasive solution for continuous glucose monitoring (CGM) using PPG signals and ML techniques.