(LB04-04-26) Cross-Trial Prediction of Tirzepatide-Associated Weight Loss Using Trial-Level Validation

Objectives: To develop and validate a trial-level predictive model for estimating mean placebo-adjusted tirzepatide-associated weight change using aggregated data from randomized controlled trials (RCTs), and to translate this framework into a web-based tool for research and trial-planning.

Methods: Aggregated arm-level data from 17 randomized controlled trials of tirzepatide versus placebo (49 arms; n=9,441) were used to assemble a modeling dataset. Predictors included dose category, follow-up duration, population type (type 2 diabetes mellitus [T2DM] vs non-T2DM), and dose×time interactions. Weighted least squares (WLS), ridge regression, random forest, and histogram gradient boosting models were compared. Ridge regression was trained on 32 placebo-adjusted active-versus-placebo effect rows using trial-level train/test splitting and leave-one-trial-out cross-validation to assess generalizability. Predictor importance was quantified using permutation analysis. The optimized model was implemented as a web-based tool.

Results: Ridge regression demonstrated low held-out test error on placebo-adjusted active-versus-placebo effects (weighted mean absolute error [WMAE] = 2.31 kg; weighted root mean square error [WRMSE] = 2.81 kg). High-dose tirzepatide and extended follow-up durations were associated with greater placebo-adjusted weight loss, while effects were attenuated in T2DM populations. Permutation importance analysis identified population type, dose×time interactions, and treatment duration as the most influential predictors. Nonlinear models exhibited poor generalization on the arm-level dataset and showed evidence of overfitting (test WMAE = 6.27 kg for random forest and 5.76 kg for histogram gradient boosting).

Conclusions: This is the first cross-trial predictive framework for tirzepatide-associated weight change, to our knowledge. Compared with more complex nonlinear models, ridge regression demonstrated superior stability and generalizability under trial-level validation. These findings support the use of ridge regression as the foundation for a web-based prediction tool for research and trial-planning.

Funding Sources: None