Endocrine Abstracts

Introduction: Current guidelines emphasize the important role of Continuous Glucose Monitoring (CGM) for type 1 diabetes mellitus (T1DM) management. The aim of the GlucoseML study is the development of a mobile health system for T1DM self-management based on CGM data, physical activity, food intake and insulin dosage. We herein present the development and evaluation of a univariate Autoregressive Moving Average (ARMA) prediction model of interstitial glucose concentration for prediction horizons of 30-, 45- and 60-minutes.

Methods: CGM data (GlucoMen Day, Menarini®) from T1DM patients over a 4-week monitoring period under real life conditions were included in the analysis. Ambulatory Glucose Profile (AGP) report was computed for every patient. Categorical variables are expressed as number (percentage). Continuous variables with or without normal distribution are expressed as mean (standard deviation) or median (range), respectively. An ARMA (p, q) model, where p and q denote, respectively, the order of the AR and MA model of the ARMA equation, was identified upon the glucose data. The partial autocorrelation plot and the Akaike information criterion (AIC) were used to estimate the appropriate values of p and q orders in the model. The root mean square error (RMSE) and the mean absolute error (MAE), were used to evaluate the predictive performance of our models.

Results: Data were included for 29 T1DM patients (38% women) aged 38 years (12). Age at diagnosis was 15 years (2-45) and diabetes duration 20 years (11). Most patients were on insulin treatment with multiple daily injections [19 (66%)] compared to continuous subcutaneous infusion [10 (34%)]. Glycosylated haemoglobin was 7.4% (5.8-10.4). Based on AGP report, time below range (glucose<54,<70 mg/dl), time in range (70-180 mg/dl) and time above range (>180,>250 mg/dl) were 1.4% (0-10), 4% (2), 62% (21-82), 22% (6) and 8% (0.9-49), respectively. Average glucose was 150 mg/dl (130-247), glucose management indicator 6.9% (6.4-9.2) and glucose variability 39% (32-51). The RMSE for examined prediction horizons of 30-, 45- and 60-minutes was 9.04 (2.22), 11.84 (3.18) and 14.82 (3.87) mg/dl, respectively. Similarly, MAE was 6.48 (1.7), 9.04 (2.56) and 11.62 (3.38) mg/dl, respectively.

Conclusions: The predictive performance of the identified ARMA models compares favourably with that of existing models of similar or higher computational complexity. More advanced multivariate adaptive deep learning models are currently under way as part of the GlucoseML study. Further analyses are required, to test the model’s predictive capacity in the critical region of hypoglycaemia.