2nd Edition of Diabetes and Endocrinology World Conference 2026

Abstract

Diabetes mellitus (DM) represents a rapidly growing global health burden, with a substantial proportion of cases remaining undiagnosed, particularly in low- and middle-income countries. Conventional glucose screening and monitoring methods rely on invasive blood sampling, which may limit accessibility and adherence in large-scale population screening. Blood Glucose Evaluation and Monitoring (BGEM) is a non-invasive artificial intelligence (AI)–based algorithm designed to estimate blood glucose levels by analysing photoplethysmography (PPG) signals obtained from wearable devices. This study evaluated the performance of BGEM in classifying glucose impairment status in an Indonesian population.

Methods:
A total of 885 participants aged 18–59 years representing different glucose impairment classifications were enrolled. Participants underwent plasma glucose testing paired with PPG recordings obtained from wearable devices after an overnight fast of at least 8 hours. A second paired measurement was collected 2 hours after participants consumed a standardized meal. The paired datasets from fasting and postprandial conditions were used to evaluate the predictive performance of the BGEM algorithm against laboratory plasma glucose measurements.

Results:
BGEM demonstrated a sensitivity and specificity of 79.34% and 79.19% for detecting glucose impairment during fasting conditions, and 79.28% and 76.55% respectively for the 2-hour postprandial condition. The model achieved a Mean Absolute Relative Difference (MARD) of 15.07 for fasting measurements and 17.22 for postprandial measurements. Error grid analysis showed that 98.96% and 99.20% of predictions fell within Zones A and B of the Type I and Type II Parkes Error Grid respectively, indicating clinically acceptable accuracy.

Conclusion:
The BGEM model demonstrated promising performance in estimating blood glucose levels using non-invasive wearable-derived PPG signals. Its convenience and non-invasive nature suggest potential utility as a scalable screening tool for glucose impairment in large populations. Such technology may support early identification of individuals at risk of diabetes and contribute to improved diabetes prevention and population health management.