Deep Learning Assisted Macronutrient Estimation For Feedforward-Feedback Control In Artificial Pancreas Systems

People with type 1 diabetes are required to manually input a feedforward bolus of insulin to compensate for glycemic excursions due to the macronutrient content of ingested meals. Precisely assessing macronutrient contents of complex food types is extremely difficult and time-consuming, whereas inaccurate dietary assessment may result in poor glycemic outcomes. To alleviate this burden, we propose a deep learning based assistive tool that automatically estimates the macronutrient content via real-time image recognition. Concretely, the user provides an image of their meal along with an estimated serving size, based on which a deep convolutional neural network (CNN) predicts the food category and subsequently queries a nutritional database to obtain the macronutrient content. This deep learning framework is integrated with an artificial pancreas (AP) system, and equipped with explicit safety constraints. Upon constraint violation, no automatic feedforward bolus is provided, and appropriate corrective insulin boluses are computed solely using the AP's intrinsic feedback control algorithm. Numerical simulations are performed to demonstrate the potential of the proposed methodology. Glucose is maintained within the safe zone of 70-180 mg/dL for 91.76 +/- 7.20% of the time, which is a significant improvement over control with unannounced meals (78.78 +/- 12.14%). Although glycemic outcomes are moderately lower with deep learning assist than with full meal announcement (97.29 +/- 2.91 %), the proposed method offers the advantage of increased autonomy by leveraging machine intelligence to facilitate macronutrient estimation without compromising safety. Robustness testing is performed by considering meal size estimation errors of up to +/- 40%: our method exhibits a small reduction (<3%) of time within 70-180 mg/dL despite over 30 g error in meal size estimates.