The importance of interpreting machine learning models for blood glucose prediction in diabetes: an analysis using SHAP.

Machine learning has become a popular tool for learning models of complex dynamics from biomedical data. In Type 1 Diabetes (T1D) management, these models are increasingly been integrated in decision support systems (DSS) to forecast glucose levels and provide preventive therapeutic suggestions, like corrective insulin boluses (CIB), accordingly. Typically, models are chosen based on their prediction accuracy. However, since patient safety is a concern in this application, the algorithm should also be physiologically sound and its outcome should be explainable. This paper aims to discuss the importance of using tools to interpret the output of black-box models in T1D management by presenting a case-of-study on the selection of the best prediction algorithm to integrate in a DSS for CIB suggestion. By retrospectively "replaying" real patient data, we show that two long-short term memory neural networks (LSTM) (named p-LSTM and np-LSTM) with similar prediction accuracy could lead to different therapeutic decisions. An analysis with SHAP-a tool for explaining black-box models' output-unambiguously shows that only p-LSTM learnt the physiological relationship between inputs and glucose prediction, and should therefore be preferred. This is verified by showing that, when embedded in the DSS, only p-LSTM can improve patients' glycemic control.

Clinical evaluation of a decision support system for glucose infusion in hypoglycaemic clamp experiments.

AIM: To provide a preliminary evaluation of the accuracy and safety of Gluclas decision support system suggestions in a hypoglycaemic clamp study. METHODS: This analysis was performed using data from 32 participants (four groups with different glucose-insulin regulation: post Roux-en-Y gastric bypass with and without postprandial hypoglycaemia syndrome, postsleeve gastrectomy and non-operated controls) undergoing Gluclas-assisted hypoglycaemic clamps (target: 2.5 mmol/L for 20 minutes at 150 minutes after oral glucose ingestion). Gluclas provided glucose infusion rate suggestions upon manual entry of blood glucose values (every 5 minutes), which were either followed or overruled by investigators after critical review. Accuracy and safety were evaluated by mean absolute error (MAE), mean absolute percentage error (MAPE), average glucose level, coefficient of variation (CV) and minimal glucose level during the 20-minute hypoglycaemic period. RESULTS: Investigators accepted 84% of suggestions, with a mean deviation of 30.33 mg/min. During the hypoglycaemic period, the MAE was 0.16 (0.12-0.24) (median [interquartile range]) mmol/L and the MAPE was 6.12% (4.80%-9.29%). CV was 4.90% (3.58%-7.27%), with 5% considered the threshold for sufficient quality. The minimal glucose level was 2.40 (2.30-2.50) mmol/L. CONCLUSIONS: Gluclas achieved sufficiently high accuracy with minimal safety risks in a population with differences in glucose-insulin dynamics, underscoring its applicability to various patient groups.

Counter-regulatory responses to postprandial hypoglycaemia in patients with post-bariatric hypoglycaemia vs surgical and non-surgical control individuals.

AIMS/HYPOTHESIS: Post-bariatric hypoglycaemia is an increasingly recognised complication of bariatric surgery, manifesting particularly after Roux-en-Y gastric bypass. While hyperinsulinaemia is an established pathophysiological feature, the role of counter-regulation remains unclear. We aimed to assess counter-regulatory hormones and glucose fluxes during insulin-induced postprandial hypoglycaemia in patients with post-bariatric hypoglycaemia after Roux-en-Y gastric bypass vs surgical and non-surgical control individuals. METHODS: In this case-control study, 32 adults belonging to four groups with comparable age, sex and BMI (patients with post-bariatric hypoglycaemia, Roux-en-Y gastric bypass, sleeve gastrectomy and non-surgical control individuals) underwent a postprandial hypoglycaemic clamp in our clinical research unit to reach the glycaemic target of 2.5 mmol/l 150-170 min after ingesting 15 g of glucose. Glucose fluxes were assessed during the postprandial and hypoglycaemic period using a dual-tracer approach. The primary outcome was the incremental AUC of glucagon during hypoglycaemia. Catecholamines, cortisol, growth hormone, pancreatic polypeptide and endogenous glucose production were also analysed during hypoglycaemia. RESULTS: The rate of glucose appearance after oral administration, as well as the rates of total glucose appearance and glucose disappearance, were higher in both Roux-en-Y gastric bypass groups vs the non-surgical control group in the early postprandial period (all p<0.05). During hypoglycaemia, glucagon exposure was significantly lower in all surgical groups vs the non-surgical control group (all p<0.01). Pancreatic polypeptide levels were significantly lower in patients with post-bariatric hypoglycaemia vs the non-surgical control group (median [IQR]: 24.7 [10.9, 38.7] pmol/l vs 238.7 [186.3, 288.9] pmol/l) (p=0.005). Other hormonal responses to hypoglycaemia and endogenous glucose production did not significantly differ between the groups. CONCLUSIONS/INTERPRETATION: The glucagon response to insulin-induced postprandial hypoglycaemia is lower in post-bariatric surgery individuals compared with non-surgical control individuals, irrespective of the surgical modality. No significant differences were found between patients with post-bariatric hypoglycaemia and surgical control individuals, suggesting that impaired counter-regulation is not a root cause of post-bariatric hypoglycaemia. TRIAL REGISTRATION: ClinicalTrials.gov NCT04334161.