Optical Pulse Signal Patterns and Population Attributes in Blood Glucose Estimation using Probabilistic Boosting

Abstract

This research explains that accurate and continuous monitoring of blood glucose levels is crucial for effective diabetes management. Conventional invasive techniques, however, remain uncomfortable and unsuitable for frequent use. This study presents a non-invasive blood glucose estimation framework using wrist-based photoplethysmography (PPG) signals. The acquired PPG waveform contains both an AC (pulsatile) component, reflecting cardiac-driven blood volume changes, and a DC (non-pulsatile) component associated with baseline tissue absorption. PPG signals undergo Signal Quality Index (SQI) evaluation, filtering, and baseline correction before feature extraction. Key features include systolic peak, diastolic peak, heart rate, pulse area, and PPG amplitude measures derived mainly from the AC component, along with DC-related baseline features. Additionally, empirical mode decomposition (EMD) is applied to extract intrinsic mode function (IMF)-based features, which, to the best of our knowledge, have not previously been used for noninvasive glucose prediction. Multiple machine learning models like Linear Regression, Lasso, Ridge, Decision Tree, Random Forest, and Natural Gradient Boosting (NG-Boost) are trained on dataset features such as PPG Signal, Heart Rate, Systolic Peak, Diastolic Peak, Pulse Area, Age, Height, Weight, and Gender. Experimental results demonstrate that NG-Boost outperforms all baseline models, making it the proposed algorithm for glucose level prediction. The findings confirm the feasibility of advanced PPG feature engineering and gradient boosting methods for non-invasive blood glucose monitoring.