Toward a probabilistic model of eating behavior for patients with type 1 diabetes

There is significant motivation for eating behavior profiles in general medical research and in type 1 diabetes mellitus (T1DM) research. We hypothesize that accurate meal behavior profiles can be developed by using meal diary information. We use 42 total days of real patient data to construct a meal profile for each of 12 patients. For each patient, we compile eating behavior into a meal profile for each day consisting of meal regimes (three for a typical patient). Within each regime, we associate a probability distribution which indicates the probability of eating for consecutive 15-minute intervals throughout the day. We also estimate the probability that no meal will arrive in a given regime, and an estimated carbohydrate intake amount for each expected meal. The algorithm shows promising results in predicting meal times, but estimates of regime carbohydrate intakes and the number of regimes per day need significant improvement. We expect a substantial increase in algorithm performance given more data with which to develop profiles.