Objective: We show how the use of time features in a deep learning model improves glucose prediction on T1D patients.

Method: Data of 139 adult T1D patients wearing DBLG1 System, from clinical trials NCT02987556 and NCT04190277, were used in this study. 70% of patients were used as a train set and 30% as a test set. The OHIO test database was used as an extra test set. In all cases, we only kept the data far from meals and physical activities.

The hour of the day may reflect the meal pattern and night/day differences. To consider this, we added (cos ((time in hour) / 24 * 2 pi) , (sin ((time in hour) / 24 * 2 pi) as input to the perceptron.

The day of the week may also reflect different behaviors. To consider this, we input a label for weekday vs. week-end.

We trained 2-layer perceptrons using past glucose levels (13 values with 5 minutes interval) and combinations of these time features to predict the glycemia at 45 minutes.

Results: On the clinical trials data, the baseline model achieved a 13.35% MARD and a 26.56 RMSE. When using the hour, we got a 13.2% MARD and 26.35 RMSE. When using the weekday label, we got a 13.32% MARD and an RMSE of 26.56. When using the hour and the weekday label, we obtained a 13.17% MARD and a 26.39 RMSE. We observe similar results on the Ohio data: the figure shows the clark grids for the baseline (left) and when using hour and weekday label (right) .

Conclusion: The use of the time of the day leads to great improvements on MARD and RMSE, as does, to a lesser extent, the use of the day of the week.


M.Louis: Employee; Diabeloop SA. H.M.Romero-ugalde: Employee; Diabeloop SA. A.Adenis: Employee; Diabeloop SA. L.Daudet: Employee; Diabeloop SA. Y.Tourki: Employee; Diabeloop SA. E.Huneker: Board Member; Diabeloop SA.

Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at http://www.diabetesjournals.org/content/license.