Model Fusion to Enhance the Clinical Acceptability of Long-Term Glucose Predictions

Abstract

This paper presents the Derivatives Combination Predictor (DCP), a novel model fusion algorithm for making long-term glucose predictions for diabetic people. First, using the history of glucose predictions made by several models, the future glucose variation at a given horizon is predicted. Then, by accumulating the past predicted variations starting from a known glucose value, the fused glucose prediction is computed. A new loss function is introduced to make the DCP model learn to react faster to changes in glucose variations. The algorithm has been tested on 10 in-silico type-1 diabetic children from the T1DMS software. Three initial predictors have been used: a Gaussian process regressor, a feed-forward neural network and an extreme learning machine model. The DCP and two other fusion algorithms have been evaluated at a prediction horizon of 120 minutes with the root-mean-squared error of the prediction, the root-mean-squared error of the predicted variation, and the continuous glucose-error grid analysis. By making a successful trade-off between prediction accuracy and predicted-variation accuracy, the DCP, alongside with its specifically designed loss function, improves the clinical acceptability of the predictions, and therefore the safety of the model for diabetic people.

The Reporting Quality of Machine Learning Studies on Pediatric Diabetes Mellitus: Systematic Review

“…Develop an enhanced reinforcement learning model for personalized insulin delivery and glucose con- Most studies (n=6) focused on the discovery of etiologic or prognostic biomarkers of T1DM [77,80,81,85,88,89], followed by studies aiming to predict hypoglycemia using noninvasive methods such as ECG or EEG signals, breath volatile organic compounds (n=5) [21,67,69,70,76], insulin bolus calculators for closed-loop glucose control (n=5) [66,[82][83][84]90], accurate prediction of glucose levels or hypoglycemia from continuous glucose monitor (CGM) data (n=4) [72,73,78,79], etiologic or risk factors for insulin resistance or T2DM (n=4) [71,74,75,86], and other goals such as long-term cardiovascular risk stratification [68], accurate carbohydrate counting via a smartphone app [18], prediction of cystic fibrosis-related DM from CGM signal [87], and the economic evaluation of diabetic retinopathy screening via AI versus standard care [91].…”

“…The characteristics of the training sample (including the validation data set) were not reported in 3 studies [18,90,91]. Of the 25 studies reporting details, in 18 (72%), the training sample involved human patients, in 6 (24%) only in silico patients [66,72,73,[82][83][84], and there were both human and in silico patients in one study [78]. Of the 9 studies focusing on bolus calculation or glucose prediction from CGM data, only 2 (22%) had human patients in the training sample [78,79], 6 (67%) had only in silico patients [66,72,73,[82][83][84], and the training sample was not characterized in one study [90].…”

“…The size of the training sample ranged from 561 [68] to 5 [76]. Of the 25 studies reporting details, only 10 (48%) had training samples with more than 100 patients [68,71,74,77,78,81,[85][86][87][88], whereas 9 (36%) involved 10 or fewer patients [67,69,70,72,73,76,[82][83][84]. Etiologic or prognostic studies for T1DM [77,80,81,85,88,89] and T2DM [71,74,75,86] featured the largest training samples ranging between 56 and 418 and 45 and 373, respectively.…”

“…The test sample was characterized in 21 studies that involved a full ML workflow. Testing was performed on the same patients as the training in 12 (57%) studies, involving time-split in 8 studies using CGM data [66,72,73,76,78,[82][83][84] and cross-validation in 4 studies [21,68,77,80]. In 5 (24%) studies, the test sample involved randomly selected patients from the same center as the training sample [67,[69][70][71]79]; in 2 (10%) studies, patients from one or more external centers [75,81]; and in 2 (10%) studies randomly selected patients from multicenter studies [74,85].…”

“…In 5 (24%) studies, the test sample involved randomly selected patients from the same center as the training sample [67,[69][70][71]79]; in 2 (10%) studies, patients from one or more external centers [75,81]; and in 2 (10%) studies randomly selected patients from multicenter studies [74,85]. The test sample included 10 or fewer patients in 10 (48%) studies [67,69,70,72,73,75,76,[82][83][84], and only 4 (19%) studies featured test samples involving more than 100 patients [68,74,77,78]. The sample sizes of the external test samples ranged between 10 and 186 [74,75,81,85].…”