Predicting Treatment Failure With Sodium-Glucose Cotransporter-2 Inhibitors in People With Type 2 Diabetes: Novel Artificial Intelligence and Machine Learning Approach

Abstract Background The rate of treatment failure with sodium-glucose cotransporter-2 inhibitors (SGLT2i) is high among individuals with type 2 diabetes (T2D). Accurately predicting SGLT2i treatment failure is important for improving the clinical management of T2D. Objective The study aimed to use machine learning (ML) models to identify factors predicting treatment failure with SGLT2i in T2D and to evaluate model performance. Methods This retrospective observational cohort study included adults with T2D treated with SGLT2i (2016-2024). The primary outcome was overall treatment failure with SGLT2i during follow-up (≥180 days after SGLT2i initiation). The secondary outcome was subtypes of treatment failure with SGLT2i (treatment discontinuation, failure with action, and inertial failure) or nonfailure, which was defined as not meeting the definition for one of the failure subtypes. Variables potentially associated with treatment failure were assessed during the year before SGLT2i treatment initiation (analysis 1) and the year before SGLT2i treatment failure (analysis 2). Using these variables, ML models—logistic regression (LR), multilayer perceptron (MLP), extreme gradient boosting (XGBoost), and Transformer—were used to identify significant predictors of the outcomes. Model performance metrics (accuracy, area under the curve, precision, recall, and F1-score) were calculated. Using Shapley Additive Explanations methodology, key features were identified based on their impact on model predictions. LR and Transformer models using key features were further evaluated for their potential to support the development of a risk score for predicting treatment failure with SGLT2i. Results Among all individuals in the study (N=62,222), 71% (n=44,156) had treatment failure with SGLT2i. Across subtypes, failure with action (n=23,839, 38.3%) was more common than treatment discontinuation (n=16,449, 26.4%) and inertial failure (n=3868, 6.2%). Model performance was moderate in both analyses. In analysis 1, the accuracy ranged from 0.72 to 0.73 for predicting overall treatment failure and from 0.56 to 0.57 for predicting the subtype of treatment failure. In analysis 2, the accuracy ranged from 0.74 to 0.75 for predicting overall treatment failure and from 0.61 to 0.63 for predicting the subtype of treatment failure. XGBoost, MLP, and Transformer models showed small improvements compared with LR. Using the top 9 key features identified from the Shapley Additive Explanations analysis, the Transformer model performed similarly in accuracy and area under the curve to its counterpart using the full feature set. Conclusions Performance across the LR, MLP, XGBoost, and Transformer models was moderate. The advanced ML models performed slightly better than LR. Overall, the results suggest that further model advancements and increased data availability are needed to better predict treatment failure with SGLT2i. The LR coefficients from the key features model may inform the development of a risk score to predict SGLT2i treatment failure. Accurate prediction could inform individualized treatment planning for individuals with T2D.