Objective: This paper analyses the glycemia distribution impact in a deep learning (DL) model for glucose forecast on T1D patients.

Method: Data of 139 adult T1D patients equipped with DBLG1 System are used. Data of 70% of the patients is used for hyperoptimizing a DL model (hyperop set) and 30% for testing (test set) . The model accuracy is evaluated using the RMSE for each patient of the test set. The interquartile range (IQR) is calculated for the entire hyperop set (baseline) and for each patient. The difference between each patient IQR and the baseline is computed. Then the correlation is computed between the differences and the RMSE.

Results: A correlation of -0.77 is found between the difference of IQR and the RMSE (Figure 1 B) . It indicates that when IQR (hyperop set) > IQR (patient) the prediction is more accurate. Conversely if IQR (patient) is higher, the accuracy is worse. High differences are observed between the distributions of the hyperop set and of the two patients on which the model performed worse (Figure 1 A) .

Conclusion: RMSE is larger for patients whose glycemia distribution is really different from glycemia distribution on the hyperop set. IQR difference may be used as a predictor of the prediction accuracy for a given patient.


H.M. Romero-Ugalde: Employee; Diabeloop SA. A. Adenis: Employee; Diabeloop SA. L. Daudet: Employee; Diabeloop SA. M. Louis: 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.