3465 - Doubly Robust Causal Estimation of Proton Therapy vs. IMRT/3DCRT on 90-Day Unplanned Hospitalization-Free Time In La-NSCLC

Purpose/Objective(s):

Unplanned hospitalizations during/after concurrent chemoradiation (cCRT) for locally advanced NSCLC (LA-NSCLC) can interrupt therapy and delay consolidation immunotherapy. Prior durvalumab-era observational work reported fewer 90-day unplanned hospitalizations with proton beam therapy (PBT) versus IMRT. We aimed to estimate the causal effect of PBT versus photon radiotherapy (IMRT/3DCRT) on early hospitalization burden using an interpretable time-based endpoint and causal inference designed for nonrandomized data.

Materials/Methods:

We analyzed N=786 consecutive LA-NSCLC patients treated with definitive cCRT: photon IMRT/3DCRT n=491 and PBT n=295 (30 baseline covariates; PBT 37.5%). The primary endpoint was hospitalization-free days through day 90 (RMST90), computed as 90 minus the observed time to first unplanned hospitalization, capped at 90 days, capturing both whether and when hospitalization occurs in units of days. Censoring before day 90 was rare (0.41% photons; 0% PBT). Because modality selection was not randomized, we used a doubly robust (DR) estimator combining (1) a propensity model to assess overlap/positivity and improve covariate balance and (2) treatment-specific outcome models for RMST90; the average treatment effect (ATE) is consistent if either component is correctly specified. To reduce overfitting and strengthen generalizability, we additionally performed 5-fold cross-fitting with out-of-fold (OOF) estimation and summarized overlap (propensity AUC and score range), ATE, and conditional average treatment effects (CATE).

Results:

Overlap was acceptable (AUC=0.772; propensity 0.132–0.579) without near-deterministic assignment. Mean observed RMST90 was 75.46 days for PBT vs 72.39 for photons (naïve +3.07 days). DR estimation yielded ATE +3.91 hospitalization-free days favoring PBT, concordant in direction and magnitude with the naïve comparison. Individualized effects supported personalized management: CATE mean +3.85 days (IQR 1.78–6.07), enabling risk-adapted modality selection based on predicted gains in hospitalization-free time. Exploratory explainability analyses suggested variation in predicted benefit was driven primarily by baseline comorbidity and smoking burden, supporting clinically interpretable personalization. With cross-fitting, OOF ATEs remained consistently positive (mean +3.54; SD 0.56; range +2.89 to +4.23); validation propensity AUC was 0.64±0.03, supporting robustness to sample splitting.

Conclusion:

In this cohort, DR causal analysis suggests PBT increases 90-day hospitalization-free time by ~4 days compared with photon IMRT/3DCRT. Incorporating doubly robust estimation with cross-fitting strengthens comparative effectiveness inference when randomization is unavailable. This is achieved by explicitly addressing measured confounding, diagnosing overlap, and demonstrating stability across folds, while acknowledging potential unmeasured confounding.