Tumor-type–stratified causal machine learning to identify heterogeneous benefit of PD-(L)1+CTLA-4 versus PD-(L)1 monotherapy: A hypothesis-generating analysis from the MSK pan-cancer cohort.

Abstract

e14505 Background: Combination immune checkpoint inhibition (ICI) can improve survival but increases immune-related adverse events; the incremental benefit varies across tumors and patients, and traditional subgroup analyses are underpowered and multiplicity-prone. Methods: Using the publicly available MSK pan-cancer ICI cohort (Samstein et al., Nat Genet 2019), we compared PD-(L)1+CTLA-4 versus PD-(L)1 monotherapy for 12-month overall survival, defining time zero as first ICI to mitigate immortal time bias. Primary analyses were tumor-type–stratified. We estimated conditional and average treatment effects with a doubly robust learner (DRLearner) using 3-fold cross-fitted propensity and outcome models, evaluated overlap, and trimmed propensity scores to 0.05–0.95. We computed tumor-specific 12-month survival differences with 1,000 bootstrap samples for 95% CIs and bootstrap z-test p-values. We learned an interpretable treatment-selection rule via decision trees and estimated counterfactual policy value with inverse-probability weighting; sensitivity analyses included E-values and inverse probability of censoring weighting. Results: Among 1,285 eligible patients (1,066 monotherapy; 219 combination), 848 remained after trimming (34% trimmed). The doubly robust average treatment effect favored combination by +14.1 percentage points (pp) in 12-month survival (95% CI 11.2–17.0; p < 0.001; E-value 1.86). Marked heterogeneity emerged: renal cell carcinoma showed a robust benefit (+21.1pp; 95% CI 10.6–30.5; p < 0.001; 96.4% vs 75.4% 12-month survival), and non–small cell lung cancer showed a large but imprecise benefit (+30.5pp; 95% CI 10.0–48.3; p = 0.002; combination n = 21). Melanoma showed no significant incremental benefit (+6.0pp; 95% CI −4.7 to 17.0; p = 0.28), as did esophagogastric cancer (+6.1pp; 95% CI −16.9 to 28.9; p = 0.62) and bladder cancer (−1.4pp; 95% CI −25.8 to 21.8; p = 0.91). Decision trees highlighted tumor histology and tumor mutational burden as key effect modifiers. The learned policy recommended combination for 66.7% of patients and achieved 84.7% expected 12-month survival versus 83.2% for treat-all combination and 63.4% for treat-all monotherapy, suggesting potential de-escalation in ~10% without compromising survival. Conclusions: Tumor-type–stratified doubly robust causal ML identified heterogeneous 12-month survival benefit of PD-(L)1+CTLA-4 versus PD-(L)1 monotherapy, strongest in RCC and NSCLC and minimal in melanoma. Despite possible residual confounding from unavailable variables (e.g., ECOG, PD-L1, line), the signal persisted after cross-fitting, overlap/trimming, and sensitivity analysis (E-value 1.86), supporting a robust, hypothesis-generating benefit.