(P19-010-24) Macronutrient Constraints and Priors Improve Carbohydrate Prediction From Continuous Glucose Monitors

Objectives: Diet monitoring is a crucial intervention step for patients with diabetes to monitor their food intake. Conventional diet monitoring methods such as manually logging meals are however cumbersome and error-prone. To reduce the burden on the user, in a prior work, we proposed an automatic diet monitoring solution that analyzed the glucose response after a meal, also called the postprandial glucose response (PPGR). The rationale behind our approach was that the shape of the PPGR depends on not just the carbs in the meal but also on the amount of protein and fat. However, protein and fat have similar effects on PPGR, so predicting their individual amounts can be challenging. Further, the macros in conventional meals tend to be correlated. To overcome these limitations, we propose a diet monitoring approach that compresses macronutrients into a latent variable and uses a prior distribution of macronutrients to recover the original macronutrients.

Methods: We generate a latent variable that captures domain knowledge about how macros affect PPGRs (as the amount of carbs increases so does the magnitude of the latent variable whereas adding protein and fat reduces it). We then predict the latent variable from a meal’s PPGR using machine learning. Once the model predicts the latent variable, we recover the original macros using an inverse function that also takes prior information of macros based on dietary habits of an individual.

Results: We evaluated our proposed approach against prediction directly from PPGRs in two ways. Using average amount of protein and fat across all meals in our dataset as the prior, we found a significantly lower prediction error compared to predicting net carbs (carbs – fiber) directly from PPGRs. We found a further significant reduction in net carb prediction when we used actual amounts of protein and fat as the prior.

Conclusions: Our results suggest that predicting carbs indirectly through the latent variable reduces prediction error compared to predicting carbs directly from PPGR without any prior information. Even an approximate average estimate of protein and fat can lead to improvement, a significant result since most people do not have access to ground truth amounts.

Funding Sources: This work was supported by the National Science Foundation award IIS #XXXXXXX.