Enabling fully automated insulin delivery through meal detection and size estimation using Artificial Intelligence

Abstract

We present a robust insulin delivery system that includes automated meal detection and carbohydrate content estimation using machine learning for meal insulin dosing called robust artificial pancreas (RAP). We conducted a randomized, single-center crossover trial to compare postprandial glucose control in the four hours following unannounced meals using a hybrid model predictive control (MPC) algorithm and the RAP system. The RAP system includes a neural network model to automatically detect meals and deliver a recommended meal insulin dose. The meal detection algorithm has a sensitivity of 83.3%, false discovery rate of 16.6%, and mean detection time of 25.9 minutes. While there is no significant difference in incremental area under the curve of glucose, RAP significantly reduces time above range (glucose >180 mg/dL) by 10.8% (P = 0.04) and trends toward increasing time in range (70–180 mg/dL) by 9.1% compared with MPC. Time below range (glucose <70 mg/dL) is not significantly different between RAP and MPC.

Continuous glucose monitoring for automatic real-time assessment of eating events and nutrition: a scoping review

“…The publications covered a wide age range, including pediatric ( 62 , 73 ), adolescent ( 62 , 64 , 65 ), and adult ( 59 – 61 , 63 , 66 , 68 , 70 , 71 , 74 , 75 , 82 – 85 , 87 ) populations. Fourteen publications included participants with T1DM ( 59 , 62 – 66 , 68 , 70 , 71 , 73 , 74 , 84 , 85 , 87 ), one publication included a sample of participants with T1DM or T2DM ( 75 ), one publication included participants with prediabetes or moderately controlled T2DM ( 83 ) and three publications included participants without diabetes ( 60 , 61 , 82 ). Publications included both controlled/inpatient ( 59 , 62 – 65 , 68 , 74 , 82 , 85 ) and free-living settings ( 60 , 61 , 66 , 70 , 71 , 73 , 75 , 83 , 87 ).…”

“…Fourteen publications included participants with T1DM ( 59 , 62 – 66 , 68 , 70 , 71 , 73 , 74 , 84 , 85 , 87 ), one publication included a sample of participants with T1DM or T2DM ( 75 ), one publication included participants with prediabetes or moderately controlled T2DM ( 83 ) and three publications included participants without diabetes ( 60 , 61 , 82 ). Publications included both controlled/inpatient ( 59 , 62 – 65 , 68 , 74 , 82 , 85 ) and free-living settings ( 60 , 61 , 66 , 70 , 71 , 73 , 75 , 83 , 87 ).…”

“…Moreover, 7/19 (36.8%) publications used meal announcements and/or meal-accompanying insulin boluses ( 66 , 68 , 70 , 71 , 73 , 75 , 87 ). Finally, 8/19 (42.1%) publications utilized other inputs besides CGM, e.g., heart rate or the insulin sensitivity factor ( 59 – 61 , 63 , 64 , 70 , 74 , 82 ). Overall, 15/19 (78.9%) publications elicited methodological concerns in two or more appraised domains ( 59 – 61 , 63 , 64 , 66 , 68 , 70 , 71 , 73 – 75 , 83 , 84 , 87 ), and all publications had methodological concerns in at least one of the appraised domains ( 59 – 66 , 68 , 70 , 71 , 73 – 75 , 82 – 85 , 87 ).…”

“…Our review identified a wide range of methods to automatically detect IA, including fuzzy logic ( 59 , 63 ), model predictive control ( 74 ), support vector machine ( 60 ), random forest ( 60 , 61 , 82 ), (extreme) gradient boosting trees ( 60 , 61 , 82 ), backward difference ( 62 ), Kalman filter estimation ( 62 ), second derivative of glucose ( 62 ), Kalman filters ( 64 , 68 ), switching signal generator ( 68 ), simulation-based explanation ( 83 ), classification of horizons ( 73 ), analysis of the glucose trajectory ( 84 ), pattern recognition using linear discriminant analysis ( 73 ), and threshold violation-based approaches ( 73 ). Further, adaptive model-based ( 64 ), super-twisting-based ( 66 , 71 ), feedback scheme-based ( 70 ), and physiological parameter-invariant-based ( 87 ) meal detection approaches were applied.…”

NIH Fifth Artificial Pancreas Workshop 2023: Meeting Report: The Fifth Artificial Pancreas Workshop: Enabling Fully Automation, Access, and Adoption

“…It is also important to consider that the existence of large intra-patient and inter-patient physiologic variability demands adaptive algorithms in the pursuit of effective FCL solutions. Mosquera-Lopez et al 12 have evaluated a single-hormone robust artificial pancreas (RAP) with automated meal detection using machine learning in a randomized, single-center, crossover trial. Results indicated that RAP significantly and safely reduced the time above 180 mg/dL when compared with an MPC algorithm in HCL.…”