2020/09/21 by Maxime De Bois, De Bois, Maxime, Mounîm A. El‐Yacoubi +3
Computer Science · Health Professions · Medicine · #Artificial Intelligence in Healthcare #Diabetes Management and Research #FOS: Biological sciences #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning in Healthcare #Quantitative Methods (q-bio.QM)
paper · pdf · doi:10.48550/arxiv.2009.10514
openalex publication_date 2020/09/21 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Standard objective functions used during the training of neural-network-based\npredictive models do not consider clinical criteria, leading to models that are\nnot necessarily clinically acceptable. In this study, we look at this problem\nfrom the perspective of the forecasting of future glucose values for diabetic\npeople. In this study, we propose the coherent mean squared glycemic error\n(gcMSE) loss function. It penalizes the model during its training not only of\nthe prediction errors, but also on the predicted variation errors which is\nimportant in glucose prediction. Moreover, it makes possible to adjust the\nweighting of the different areas in the error space to better focus on\ndangerous regions. In order to use the loss function in practice, we propose an\nalgorithm that progressively improves the clinical acceptability of the model,\nso that we can achieve the best tradeoff possible between accuracy and given\nclinical criteria. We evaluate the approaches using two diabetes datasets, one\nhaving type-1 patients and the other type-2 patients. The results show that\nusing the gcMSE loss function, instead of a standard MSE loss function,\nimproves the clinical acceptability of the models. In particular, the\nimprovements are significant in the hypoglycemia region. We also show that this\nincreased clinical acceptability comes at the cost of a decrease in the average\naccuracy of the model. Finally, we show that this tradeoff between accuracy and\nclinical acceptability can be successfully addressed with the proposed\nalgorithm. For given clinical criteria, the algorithm can find the optimal\nsolution that maximizes the accuracy while at the same meeting the criteria.\n