Prediction of insulin requirements by explainable machine learning for individuals with type 1 diabetes

Objective: Adverse events related to insulin therapy remain common in individuals with type 1 diabetes. Optimization of insulin dose can reduce the frequency of these events and help to prevent macrovascular and microvascular complications. The aim of the present study was to develop machine learning models to predict the total daily dose (TDD) of insulin on the basis of data available in routine clinical practice, to evaluate the performance of the models, and to interpret the relation between its predictions and features.

Materials and methods: This retrospective observational study conducted at a single center recruited individuals diagnosed with type 1 diabetes who visited Kobe University Hospital and used continuous glucose monitoring in combination with continuous subcutaneous insulin infusion between 1 April 2010 and 29 February 2024. We developed TDD prediction models based on machine learning and evaluated its performance on the basis of the mean absolute percentage error (MAPE). Explainable artificial intelligence frameworks were applied to the machine learning model to facilitate interpretability.

Results: A total of 110 individuals with type 1 diabetes was included in the study. The best-performing model, the Random Forest, achieved a MAPE of 19.8%. The most important feature of the model for prediction of the TDD of insulin was body weight, followed by waist circumference and carbohydrate intake.

Conclusions: We here developed machine learning models that predict the TDD of insulin from clinical information. Such models could contribute to the treatment of many individuals undergoing insulin therapy, with further developments being warranted.