246 - Identifying Hepatocellular Carcinoma Patients Who Derive Maximal Survival Benefit from Proton Therapy Using Causal Machine Learning

01:00pm - 01:10pm ET

Room 254

Presenter(s)

Ibrahim Chamseddine, PhD - Mass General Brigham, Harvard Medical School, Boston, MA

I. Chamseddine¹, H. J. Roberts², K. Joseph³, J. Y. Wo⁴, E. J. Koay⁵, T. S. Hong⁶, and H. Paganetti¹; ¹Department of Radiation Oncology, Massachusetts General Hospital/Mass General Brigham and Harvard Medical School, Boston, MA, ²Columbia University College of Physicians and Surgeons, New York, NY, ³Massachusetts General Hospital, Harvard Medical School, Boston, MA, ⁴Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA, ⁵Department of Gastrointestinal Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, TX, ⁶Dana-Farber Cancer Institute, Boston, MA

Purpose/Objective(s):

Proton beam therapy (PBT) for hepatocellular carcinoma (HCC) is designated ASTRO Group 1 for coverage based on dosimetric liver-sparing advantages, yet no validated criteria exist to identify which patient subgroups actually benefit. This gap results in considerable initial insurance denials and suboptimal allocation of limited capacity. We developed and externally validated a causal machine learning framework to establish objective criteria identifying patients who derive maximal survival benefit from PBT.

Materials/Methods:

We analyzed 232 HCC patients treated with definitive radiotherapy at INST1 (2008–2021): 63 PBT, 169 photon. External validation included 141 patients from INST2 (40 PBT, 101 photon). Primary outcome was overall survival analyzed as restricted mean survival time at 36 months (RMST₃₆). Inverse probability of treatment weighting (IPTW) estimated average causal effects using five baseline confounders (age, ALBI score, gross tumor volume (GTV), liver volume, platelet count). Heterogeneous treatment effects were estimated using causal forest (50 trees, minimum leaf size 5, honest splitting) incorporating polynomial and interaction terms. A structural breakpoint in the tumor volume–benefit relationship was identified using the Chow test and robustness was checked using outlier exclusion, slope ratio plausibility, and cross-institutional replication at the identical threshold.

Results:

IPTW-adjusted analysis demonstrated significant PBT survival benefit (RMST₃₆: 5.83 months, 95% CI: 1.66–9.50, p=0.006), validated by causal forest (mean ITE: 5.57 months, 95% range: 2.05–10.73). Feature importance identified liver function (ALBI: 29.7%) and tumor burden interactions (GTV²: 13.2%; GTV×liver: 11.9%; log(GTV): 11.7%) as primary drivers of heterogeneity. Among candidate effect modifiers, only tumor burden demonstrated reproducible correlation with benefit (INST1: ?=-0.538, p<0.001; INST2: ?=-0.408, p<0.001), while other factors failed external validation. Chow test analysis identified a structural breakpoint at 54 ml (F=51.99, p<0.001), with a steeper ITE-GTV slope below versus above this threshold. This breakpoint replicated in INST2 (F=21.68, p<0.001) and was robust to exclusion of high-leverage small-tumor outliers (F=49.64, p<0.001). Patients with GTV<54 ml derived nearly double the benefit of larger tumors (INST1: 8.5+/-2.0 vs. 4.6+/-1.5 months, p<0.001; INST2: 8.2+/-1.9 vs. 4.8+/-1.5 months), suggesting proton benefit is greatest when normal liver sparing can be maximized, which is more achievable with smaller tumor volumes.

Conclusion:

Patients with tumor volume below 50 ml derived nearly double the survival benefit from PBT compared to larger tumors after adjusting for confounding, with level 2 evidence. This threshold potentially provides objective criteria for patient selection, warranting prospective validation.