Predicting 12-month overall survival from response rate and 6-month progression-free survival in oncology clinical trials: An embedding-based machine learning approach.

Abstract

e13667 Background: Overall Survival (OS) remains the regulatory gold standard endpoint, but early trial decisions depend on Objective Response Rate (ORR) and intermediate Progression-Free Survival (PFS) to accelerate development timelines. Traditional endpoint surrogacy analyses demonstrate variable correlations across tumor types, stages, and treatment modalities. However, machine learning (ML)-based prediction of OS from early endpoints across heterogeneous trials remains limited. We developed an ML framework incorporating ORR, 6-month PFS rate, and LLM-based clinical and treatment embeddings to predict 12-month OS rate in a comprehensive cross-tumor clinical trials dataset. Methods: Data were extracted from the LARVOL CLIN outcomes database (2004–2023), comprising 1,029 Phase I–III trials (1,520 experimental and control arms) reporting ORR, 6-month PFS rate, and 12-month OS rate. To account for cross-trial heterogeneity, key population descriptors (tumor type, stage, setting, prior lines of therapy, biomarker status) were embedded separately to preserve granularity, while treatment details (regimen and dose/schedule) were embedded into a single representation. Embeddings from OpenAI text-embedding-3-large were used and reduced to 30 principal components (PCs) via Principal Component Analysis (PCA) (retaining 85% variance). ML models evaluated included XGBoost, LightGBM (LGBM), elastic net, and Random Forest (RF), with hyperparameters optimized using Optuna. Trial-grouped five-fold CV was used to prevent data leakage. Performance metrics included Coefficient of Determination (R²), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) on a logit scale. Shapley Additive Explanations (SHAP) analysis was used to quantify relative feature contributions. Results: XGBoost achieved CV R² = 0.731 for 12-month OS rate prediction, outperforming RF CV R² = 0.498 (Table 1). SHAP analysis identified 6-month PFS rate (34.1%) and tumor type (24.2%) as dominant predictors, followed by arm description embeddings (9.5%) and stage (6.9%); ORR demonstrated a smaller relative contribution of 3.9%. Limitations include the modest dataset size and uncertain generalizability to rare tumor subtypes. Conclusions: ML models integrating early endpoints with LLM-based clinical and regimen embeddings achieved robust cross-tumor prediction of 12-month OS rate, supporting trial prioritization and OS hypothesis generation. SHAP analysis confirmed that 6-month PFS rate and tumor type were the dominant predictors. Models performance for 12-month OS prediction (logit scale). Model  R²  RMSE  MAE  XGBoost  0.731  0.465  0.35  LGBM  0.716  0.510  0.35  ElasticNet  0.556  0.769  0.42  Random Forest  0.498  0.817  0.44  Coefficient of Determination (R²), Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE).