A template matching algorithm to identify unreported meals and improve quality of data collected on post-bariatric population: a feasibility study

Background and Aims: A complication of bariatric surgery is post-bariatric hypoglycemia (PBH), an increased post-prandial glycemic variability causing hypoglycemia. The combined use of continuous glucose monitoring (CGM) devices and meal in- formation is essential to develop decision rules to limit PBH occurrence. However, in trials involving PBH monitoring, manually recording meal information is burdensome and error- prone, leading to many unreported meals. Here, we propose a template match algorithm (TMA) for the retrospective identifi- cation of unreported meals using CGM data only. Methods: A template, i.e., a sample curve capturing post- prandial glucose dynamics, has been shaped from an independent dataset. Then, a new dataset consisting of 46 days of CGM data, and 94 self-reported meals collected in a single PBH individual has been selected to build and assess the TMA. After a 60%:40% training-test split to tune the matching threshold, the algorithm slides the template onto the test-set, calculates a template simi- larity score and flags possible unreported meals when the score overcame a predefined threshold. Performances were compared vs a na ̈ıve algorithm based on glucose rate-of-change (nROC). Results: On the considered subject, the TMA detects 33 on a total of 39 recorded meals in the test-set, with no false detections. Compared vs nROC, TMA improves both precision (100% vs 73%) and F1-score (92% vs 81%), and slightly worsens recall (85% vs 90%). Conclusions: The algorithm results promising to retrospec- tively identify unreported meals and improving data quality. Future work will validate the procedure in a wider dataset and will leverage TMA for developing new decision support systems for PBH management.