3592 - Estimating Individual Survival Benefit from Proton Therapy in Liver Cancer Using Causal Forest
Presenter(s)
I. Chamseddine1, K. Joseph2, J. P. Schuemann3, H. J. Roberts4, J. Y. Wo1, E. J. Koay5, T. S. Hong1, and H. Paganetti6; 1Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, 2Massachusetts General Hospital, Harvard Medical School, Boston, MA, 3Department of Radiation Oncology, Massachusetts General Hospital & Harvard Medical School, Boston, MA, 4Columbia University College of Physicians and Surgeons, New York, NY, 5Department of Gastrointestinal Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX, 6Massachusetts General Hospital, Boston, MA
Purpose/Objective(s): Proton therapy offers dosimetric advantages over photon therapy, but its survival benefit remains unclear. This study applies causal inference to identify factors driving survival benefit from proton therapy in liver cancer.
Materials/Methods: Two independent datasets of liver cancer patients treated with proton or photon therapy were analyzed (INS1: n=392 for training; INS2: n=143 for validation). Key clinical and tumor variables included baseline albumin-bilirubin (ALBI) score (liver function), gross tumor volume (GTV), alpha-fetoprotein (AFP, tumor aggressiveness), and mean liver dose adjusted to 2 Gy equivalent. To address confounding, propensity score matching was performed to balance baseline characteristics between proton and photon cohorts. Causal forest modeling was applied to estimate individual treatment effects (ITE) and quantify the survival benefit at patient level. Kaplan-Meier was used to assess treatment effect modification by key clinical factors. External validation was performed using the INS2 cohort.
Results: Casual forest analysis revealed that proton therapy provides a survival benefit, but only in selected subgroups. Patients with large GTVs (>median) demonstrated a significant survival improvement with proton therapy patients in both institutions (INS1: p=0.0025, HR=0.39; IN2: p=0.0167, HR=0.47). In patients with high AFP (>median), proton therapy was associated with a significant survival benefit in INS2 (p=0.0007, HR=0.30) but was borderline significant in INS1 (p=0.1489, HR=0.66). For patients with ALBI > -1.39, proton therapy was significantly associated with improved survival in INS1 (p=0.0028, HR=0.49) and showed a borderline significant trend in INS2 (p=0.14, HR=0.65). When patients were stratified ITE estimates, Kaplan Meier analysis demonstrated a strong survival advantage in the subgroup predicted to benefit most from proton therapy (HR=0.48, p<0.0001).
Conclusion: Proton therapy is not universally superior to photon therapy but provided a survival benefit in patients with large tumors and possibly in those with aggressive tumor biology (high AFP) and poor liver function (high ALBI). Causal inference techniques offer a robust approach to identifying treatment effect heterogeneity and developing a selection strategy for proton therapy in liver cancer.