Computational Modeling of Blood Glucose Variations using Physiological PPG Signals and Clinical Attributes

Abstract

Diabetes is a major global health challenge, with over 537 million adults affected worldwide in 2023, and an estimated 1 in 2 cases remaining undiagnosed, leading to severe complications. Traditional blood glucose monitoring relies on invasive finger-prick tests, which are inconvenient, painful, and unsuitable for continuous monitoring. To address these limitations, this study proposes a supervised learning-based predictive framework using photoplethysmography (PPG) signal patterns and demographic data to estimate glucose levels non-invasively. The dataset includes PPG signal recordings alongside patient demographic attributes such as age, weight, and BMI. In the preprocessing phase, signals undergo denoising, normalization, and feature extraction to generate robust predictive features. Baseline models such as Linear Regression (LR), Ridge Regression (RR), Lasso Regression (Lasso), and Decision Tree Regressor (DTR) are evaluated for performance comparison. The proposed approach leverages Natural Gradient Boosting Regressor (NGBR), which captures non-linear relationships and probabilistic uncertainty in glucose predictions. The model outputs continuous glucose level predictions, offering real-time estimation suitable for personal health monitoring. Experimental results demonstrate that NGBR outperforms traditional regression models in accuracy, robustness, and uncertainty quantification. By integrating this model into a Flask-based web application, patients and healthcare providers can access continuous glucose predictions, enabling proactive management of diabetes, reducing risks of hyperglycemia and hypoglycemia, and improving overall patient quality of life. This approach represents a scalable, non-invasive, and intelligent solution for continuous glucose monitoring using PPG and demographic data.