OAII-002 - EXTREME GRADIENT BOOSTING COMPARED TO TRADITIONAL MODELING FOR PREDICTING PROGRESSION-FREE AND OVERALL SURVIVAL IN PATIENTS WITH RENAL CELL CARCINOMA

C. Grant¹, J. Li², M. Shahin³; ¹Mayo Clinic Graduate School of Biomedical Sciences, Rochester, MN, USA, ²Pfizer, La Jolla, CA, USA, ³Pfizer Research & Development, Groton, CT, USA.

Background: Recent studies have demonstrated that extreme gradient boosting (XGBoost) outperforms traditional modeling methods in predicting survival outcomes. However, limited studies have assessed this comparison in renal cell carcinoma (RCC) patients on different therapeutic treatments. Herein, we evaluate the performance of XGBoost compared to traditional models in predicting progression-free survival (PFS) and overall survival (OS) in patients with RCC from four different clinical trials.

Methods: The dataset comprised 1,839 patients with RCC who received interferon-α, sunitinib, sorafenib, axitinib, or avelumab+axitinib. A total of 33 covariates were considered in the modeling, including model-derived tumor-growth inhibition metrics, treatment and dosing variables (dose interruption, dose reduction), and clinical markers. Data were split into training (70%) and testing (30%) sets for building and validating models. Performance, measured by integrated brier score (IBS) and c-index, was compared between XGBoost and traditional approaches (Cox regression, accelerated-failure time [AFT]) using t-tests. For XGBoost, sensitivity analyses were conducted to determine whether the model’s performance could be maintained using fewer covariates (N=10, 7, 5, and 3 top features by variable importance). Model generalizability was evaluated using bootstrap resampling (N=100 bootstraps) on the testing dataset.

Results: XGBoost outperformed (p < 0.05) traditional models in predicting PFS for RCC patients (Figure 1). For OS, XGBoost outperformed traditional methods when measured by c-index, and outperformed cox-regression (p < 0.05) but not AFT (p-value = 0.84) when measured by IBS. Results from the sensitivity analyses revealed that XGBoost maintained its prediction performance for PFS and OS using only 3 and 10 covariates, respectively. For PFS and OS, TGI metrics (tumor growth rate constant, drug resistance, and tumor inhibition rate constant) were amongst the top predictors.

Conclusion: XGBoost outperforms traditional modeling in predicting PFS and OS (dependent on evaluation metric) in patients with RCC. This indicates that ML-based models, like XGBoost, may serve as powerful, predictive tools for survival analysis in the pharmacometrics toolbox.