Covariate Feature Importance and Cross-Indication Transferability Analysis for Phase III Oncology Trial Outcome Prediction

Oncology drug development suffers from the lowest clinical success rates among all therapeutic areas, with the likelihood of approval (LOA) from Phase I to regulatory approval estimated at 3.4%--5.3%. Accurate prediction of Phase III trial outcomes is essential for optimizing resource allocation and reducing late-stage attrition costs. This study presents a comparative regression analysis framework designed to identify and rank covariate features that influence Phase III oncology trial success, and to evaluate the temporal stability and cross-indication transferability of these features. Using 1,203 Phase III oncology trials registered between 2005 and 2023, sourced from ClinicalTrials.gov and the BioMedTracker database, 42 candidate covariates spanning trial design, molecular characteristics, patient baseline demographics, and historical performance metrics were extracted and analyzed. SHAP-based importance ranking revealed that prior Phase II efficacy endpoints, biomarker-driven patient selection, and sponsor therapeutic area experience constituted the three most influential predictive features. A comparison of five regression techniques demonstrated that gradient boosted regression trees achieved the highest discriminative performance (AUC = 0.823, 95% CI: 0.791--0.855) on the hold-out validation cohort. Cross-indication transferability analysis across six major cancer types showed moderate to strong covariate consistency for trial design features (Spearman ρ = 0.71--0.89) and lower consistency for molecular target features (ρ = 0.43--0.62). These findings provide quantitative evidence supporting the development of more robust technical success probability (PTS) calculation methods.