Performance of continuous glucose monitoring-based meal detection algorithms in young healthy adults

Abstract Continuous glucose monitoring (CGM) enables automated detection of eating events via meal detection algorithms (MDAs); however, CGM-only MDAs have not been comprehensively evaluated using a shared dataset. We compared nine published CGM-only MDAs using standardized metrics by testing them with CGM data from 16 young, healthy, normal-weight adults under free-living conditions. We employed a per-participant holdout design, with separate training, validation, and testing sets, and assessed performance on the test set (216 meals) using sensitivity, false positives per day (FP/day), and detection time (Δt). Sensitivity ranged from 49 to 90%, FP/day from 0.12 to 2.42, and Δt from 37 to 61 min. Fuzzy logic and simulation-based approaches showed the highest sensitivity (90% and 83%) but slower detection (> 59 min) and higher FP rates (> 1.28/day). Pattern-recognition classifiers (82%, 0.39 FP/day, 44 min; 77%, 0.33, 42 min) and a glucose-insulin-model-based method (77%, 0.22, 41 min) showed more balanced performance, while rate-of-change detectors were faster (37–38 min) but less sensitive (70–72%). No single MDA consistently outperformed others across metrics. Pattern-recognition and physiological modeling approaches demonstrated the most balanced performance, whereas rate-of-change methods enabled faster detection with reduced accuracy. Algorithm choice should reflect application priorities, such as early detection versus reliability.