Digital biomarkers for interstitial glucose prediction in healthy individuals using wearables and machine learning

Abstract

A personalized low-glycemic diet, maintaining stable blood glucose levels, aids in weight reduction and managing (pre-)diabetes and migraines in individuals. However, invasiveness, high cost, and limited lifecycle of continuous glucose monitoring (CGM) devices restrict their widespread use. To address these issues, we investigated machine learning (ML) approaches for glucose monitoring using data from non-invasive wearables. Our study comprised two phases involving healthy participants: The main study included two experimental sessions lasting 7–8 h with two standardized test meals, totaling over 1550 interstitial glucose (IG) measurements with CGM, and high-frequency multimodal data collected by two different non-invasive sensor devices. The follow-up study involved more than 14,400 IG measurements. Using ML approaches, correlations between glycemic measures and sensor data were assessed to estimate the feasibility of accurately predicting personalized IG alterations in real-time. An ensemble feature selection-based light gradient boosting machine (LightGBM) algorithm, omitting the need for food logs, was developed. This algorithm achieved a root mean squared error (RMSE) of 18.49 ± 0.1 mg/dL and a mean absolute percentage error (MAPE) of 15.58 ± 0.09%, demonstrating the feasibility of non-invasive glucose monitoring with high accuracy, which paves the way for novel approaches in the objective prevention of diet-related diseases.