Interpreting Deep Glucose Predictive Models for Diabetic People Using RETAIN

This is the authors' abstract. We don't add key points for this paper.

Progress in the biomedical field through the use of deep learning is hindered the lack of interpretability of the models. In this paper, we study the architecture for the forecasting of future glucose values for diabetic. Thanks to its two-level attention mechanism, the RETAIN model is while remaining as efficient as standard neural networks. We the model on a real-world type-2 diabetic population and we compare it a random forest model and a LSTM-based recurrent neural network. Our results that the RETAIN model outperforms the former and equals the latter on accuracy metrics and clinical acceptability metrics, thereby proving its in the context of glucose level forecasting. Furthermore, we propose to take advantage of the RETAIN interpretable nature. As informative for patients as for the practitioners, it can enhance the understanding of the made by the model and improve the design of future glucose models.