Research Project

PPG-Based Blood Glucose Estimation

BGL Estimation

Overview

The paper introduces a standardized, extensible pipeline to evaluate five representative machine learning approaches using two publicly available datasets, VitalDB and MUST. It compares performance under three data-splitting protocols (random window-level, participant-aware, and leave-some-participants-out) and additionally tests generalization across datasets. Evaluation combines regression metrics with clinical error measures and a mean-prediction baseline.


Models that appeared promising under random splitting showed limited predictive ability under stricter evaluation, often performing similarly to the baseline. Nevertheless, more than 90% of predictions fell within the clinically acceptable Clarke Error Grid zones, even for the baseline. These findings highlight why favorable clinical-zone scores alone do not establish reliable glucose estimation.

Our Goal

To determine whether existing PPG-based blood glucose estimation methods generalize beyond their training data and provide meaningful predictive value. The study also aims to establish a reproducible evaluation framework that exposes data leakage, supports more reliable comparisons, and helps distinguish apparent performance from practical utility.

Contact People