925-P: Using Machine Learning to Predict Dynamic Cardiovascular Risks and Identify the Treatment Responses of Canagliflozin in Type 2 Diabetes Mellitus

Evaluating cardiovascular risks and choosing the right pharmacotherapy is crucial for personalized treatment in patients with type 2 diabetes. Our study aims to dynamically estimate cardiovascular risks in patients with type 2 diabetes and predict their heterogeneous treatment responses to canagliflozin. We employed data from ACCORD and CANVAS for model derivation and CANVAS-R and CREDENCE for external tests. The ML-CVD model used baseline clinical features and interim factors to predict 4-year major adverse cardiovascular events (MACE). ML-HF for heart failure were also developed. Complication-specific XGBoost survival models were applied. Our model included 13,961 patients for development and 9,609 for testing. The area under the receiver operating characteristic curve of ML-CVD to predict MACE was 0.75 (95%CI 0.71-0.78) in the testing cohort of ACCORD and CANVAS, 0.75 (0.65-0.84) in CREDENCE, and 0.71 (0.65-0.76) in CANVAS-R. This model outperformed traditional models Framingham, ASCVD, ADVANCE, and model using only baseline features (ML-CVD (base)). ML-CVD more effectively distinguished the cardiovascular risk reduction between canagliflozin and placebo in patients with higher baseline risks. Patients who were treated with canagliflozin and showed a greater risk reduction than the median of the placebo group at one year, as predicted by ML-CVD, were identified as "SGLT2i responders." Those who did not meet this criterion were labeled as 'non-responders'. The actual hazard ratio (HR) for MACE, assessed by Cox regression, significantly decreased in SGLT2i responders compared to non-responders in both derivation and testing cohorts (p-values: CANVAS <0.001, CREDENCE 0.03, CANVAS-R 0.004). Applying ML-HF showed comparable effects on heart failure outcomes, respectively. Machine-learning models for dynamic cardiovascular outcome prediction yield innovative strategies for risk assessment and drug response identification.