First Assessment Of The Performance Of A Personalized Machine Learning Approach To Predicting Blood Glucose Levels In Patients With Type 1 Diabetes The Cddiab Study

BackgroundPatients with type 1 diabetes (T1D) make their decisions for insulin delivery from available past and present blood glucose (BG) data and the expected effects on BG of forthcoming meals and activities according to education rules and their own experience. Enriched information on predicted BG glucose evolution could help them in better tuning insulin therapy. CDDIAB studyu2019s objective was to evaluate a new machine learning approach to predicting BG levels of each individual from a collection of personal BG measurements with contextual data.MethodsFourteen patients with T1D (8F/6M, age: 51+/-15, T1D duration: 26+/-17 years, HbA1c: 7.09+/-0.82%), treated by insulin pump (n=11) or multiple daily insulin injections (n=3) volunteered to track BG using FreeStyle Libre (n=12), Enlite (n=1) or Dexcom G4 (n=1) CGM devices and log manually meal intakes and insulin doses for 30 days. Collected data were used to design patient-specific prediction models with 30- to 90-min horizons. The algorithms were initially fitted on a training dataset corresponding to an average of 9 days, using a 5-fold cross-validation method. The remaining days of available data were used to provide an unbiased evaluation of final models.ResultsThe MARD (Mean Absolute Relative Deviation) and the consensus Error Grid Analysis were used to evaluate accuracy of BG predictions for 30- to 90-min horizons, Our results, detailed below, show the MARD and percentage of points in zones A+B on a Parkes EGA:- At 30 minutes: MARD of 6.98%u00b12.0, and 99.93%u00b10.13,- At 60 minutes: MARD of 14.78%u00b13.25, and 98.56%u00b11.00,- At 90 minutes: MARD of 20.78%u00b14.08, and 96.29%u00b12.15.ConclusionPrediction algorithms showed promising results since 99.9, 98.6 and 96.3% of computed BG values were in EGA A+B zones at 30-, 60- and 90-min horizons, respectively. The integration into the training process of collected data by an activity tracker could further improve accuracy in future developments of the algorithm.Integrated inside a mobile application to support decision-making process, this technology could help patients anticipate and avoid upcoming occurrence of hypoglycaemia and hyperglycaemia, in particular during night time. It could also be used on top of an Artificial Pancreas MPC model, allowing for more personalization and better regulation of the system, particularly during unstable phases with rapid glucose changes.