Introduction and Objective: Accurate glucose prediction is vital for effective glucose management during physical activity (PA) [1], as PA significantly impacts glucose dynamics. This study evaluates the integration of heart rate (HR) and step count per minute (SCPM) data into deep learning (DL)-based glucose prediction models.Methods: The T1DEXI dataset [2] was used (561 T1D adults assigned to exercise sessions over 4 weeks). Patients were split into training and testing sets. Meal-related data (meal time + 3h) were excluded, and only time periods with reported PA were used for validation. Two DL models [3] were trained: one with glucose data only and the other with glucose, HR, and SCPM, each model with 10 distinct hyperparameter sets to avoid differences due to tuning. Performance on glucose predictions was assessed, on the test set using RMSE, correlation, and the zones A, B, and E of the Clarke error grid.Results: Models with HR and SCPM data outperformed glucose-only models (Table 1) in RMSE (p < 0.001), correlation (p < 0.001), and Zone E (p < 0.01). However, no significant difference was found in the Zone A + Zone B classification (p = 0.92).Conclusion: Including HR and SCPM data significantly improves glucose prediction, though the gain is modest, likely due to variability in patients’ responses to PA. It is likely that adding IOB information could further enhance predictive accuracy with HR and SCPM.
H.M. Romero-Ugalde: Employee; Diabeloop SA. A. Adenis: Employee; Diabeloop. P. Gauthier: Employee; Diabeloop. C. Desir: Employee; DIABELOOP. T. Le Roux-Mallouf: Employee; Diabeloop. E. Huneker: Employee; Diabeloop. P.Y. Benhamou: Employee; Diabeloop SA. Board Member; Eli Lilly and Company, Insulet Corporation, Novo Nordisk.
This publication is based on research using data from ?Jaeb Center for Health Research Foundation that has been made available through Vivli, Inc. Vivli has not contributed to or approved, and is not in any way responsible for, the contents of this publication.
Source link
Leave a Reply