Development of personalized interpretable multilevel prediction models for the risk assessment of hypoglycemia in Type 1 Diabetes.

Prognostic tools for the timely identification of hypoglycemia can play a pivotal role in Type 1 Diabetes Mellitus (T1DM) management by facilitating preventive interventions that improve health outcomes and enhance quality of life. The adoption of prediction models for hypoglycemia risk assessment hinges on their reliability, which is founded on predictive accuracy and the models' ability to yield interpretable outputs. By leveraging continuous glucose monitoring (CGM) and insulin data along with meal information from the medical records of 12 T1DM patients, the present paper aims at the development and evaluation of personalized interpretable multilevel prediction models for the risk assessment of hypoglycemia in T1DM. The proposed hybrid approach combines mathematical models (MM) with various tree-based machine learning (ML) techniques and the use of the TreeSHAP interpretability method. Among others, a MM for the activation of the insulin signaling pathway is harnessed towards capturing the multilevel mechanisms of glucose-insulin dynamics. Various strategies for addressing class imbalance are investigated. The results obtained by the best-performing models for the prediction windows of 0-30 min (Random Forest with Random Oversampling, AUC: 92.11±7.03%) and 30-60 min (XGBoost with SMOTE, AUC: 71.42±14.34%) indicate the potential of the proposed models to handle the unbalanced nature of the considered dataset, while providing clinically meaningful insights about the estimated risk scores.