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

Objective: Continuous glucose monitoring (CGM) is used by meal detection algorithms (MDAs) to automatically identify eating events. However, no study has comprehensively validat- ed CGM-based MDAs on a common dataset. We aimed to evaluate all currently published MDAs that require only CGM data to identify the best-performing methods. Methods: We analyzed nine published MDAs using CGM data from 16 healthy young adults under free-living conditions and compared detected meals with logged ground-truth meals. We employed a per-participant holdout design, with 189 meals allocated for training, 198 for validation, and 216 for testing. P 2-2 Portionsgrößendatenbank der KiESEL-Studie Anna Holy, Tobias Höpfner, Oliver Lindtner Bundesinstitut für Risikobewertung (BfR), Berlin Hintergrund: Die Kenntnis über Portionsgrößen und Haus- haltsmaße ist im Ernährungssektor von entscheidender Be- deutung. Portionsgrößendatenbanken ermöglichen eine Ab- schätzung von Lebensmittelmengen und sind essentiell, um den Lebensmittelverzehr zu erfassen und die Nähr- bzw. Schad- stoffaufnahme zu ermitteln. Dabei müssen Portionsgrößen- datenbanken ständig verändernde Verzehrsgewohnheiten berücksichtigen. Methoden: Im Rahmen der KiESEL-Studie wurde der Lebens- mittelverzehr von 1104 Kindern (≤ 5 Jahre) mittels Wiege-/ Schätzprotokoll und Fotobuch erfasst. Das KiESEL-Fotobuch basiert auf Portionsgrößen verschiedener nationaler sowie internationaler Institutionen und beinhaltet 65 Fotoserien, Ein- zelbilder und Umrisse von Lebensmitteln. Im Verlauf der KiESEL- Studie wurden fehlende Gewichtsangaben zu Lebensmitteln After training and hyperparameter tuning, we assessed per- formance on the test set using sensitivity, false positives per day (FP/day), and detection time. Results: Across MDAs, sensitivity ranged from 49 % to 90 %, FP/day from 0.12 to 2.42, and detection time from 37 to 61 minutes. Fuzzy logic and simulation-based detectors achieved the highest sensitivity (90 % and 83 %) but had longer detec- tion times (59 and 61 min) and more false positives (2.42 and 1.28/day). Pattern-recognition classifiers (82 %, 0.39 FP/day, 44 min detection time; 77 %, 0.33, 42 min) and a glucose-in- sulin-model method (77%, 0.22, 41 min) showed balanced performance. Rate-of-change-based models detected meals earlier (37–38 min) but with lower sensitivity (70–72 %). Conclusion: No single MDA outperformed others across all metrics in young adults. Selection should depend on application priorities: pattern-recognition and glucose-insulin-model ap- proaches balance accuracy and false positives, whereas rate- of-change approaches detect meals earlier, which may support real-time monitoring applications.     «

Objective: Continuous glucose monitoring (CGM) is used by meal detection algorithms (MDAs) to automatically identify eating events. However, no study has comprehensively validat- ed CGM-based MDAs on a common dataset. We aimed to evaluate all currently published MDAs that require only CGM data to identify the best-performing methods. Methods: We analyzed nine published MDAs using CGM data from 16 healthy young adults under free-living conditions and compared detected meals with logged ground-tr...     »