Main Session
Sep 28
PQA 04 - Breast Cancer, Patient Reported Outcomes/QoL/Survivorship, Functional Radiation Medicine, Hematologic Malignancies, Palliative Care, and International/Global Oncology

2913 - Individualizing Post-Mastectomy Radiotherapy Benefit in Breast Cancer Using a Causal Survival Forest: Development and External Validation

03:00pm - 04:00pm ET
Poster Hall - Exhibit Hall A
Screen: 7
POSTER

Presenter(s)

Jingnan Wang, MD - Cancer Hospital Chinese Academy of Medical Sciences, Beijing, Beijing

J. Wang1, C. Wang2, Y. Chu3, L. Y. Bian1, Y. Zhang1, and Y. Zhou1; 1Department of Radiation Oncology, The First Affiliated Hospital of Bengbu Medical University, Bengbu, Anhui, China, 2Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, 3The First Affiliated Hospital of Bengbu Medical University, Bengbu, Anhui, China

Purpose/Objective(s): Post-mastectomy radiotherapy (PMRT) decisions remain uncertain for many intermediate-risk breast cancer patients because population-level guidelines do not quantify individualized overall survival (OS) benefit. We hypothesized that a causal survival forest (CSF), a causal machine-learning method, could estimate individualized PMRT benefit on 5-year OS and identify actionable heterogeneity.

Materials/Methods: We performed a retrospective cohort study of mastectomy patients from SEER (2010-2018, n=5,941), split into development (n=4,752) and internal validation (n=1,189), plus an independent multicenter external cohort (n=409, 42 events). Fourteen pre-treatment covariates were prespecified based on temporal availability and clinical plausibility; a directed acyclic graph documented causal assumptions. The CSF was trained within a causal inference framework with propensity-score overlap diagnostics and inverse probability of treatment weighting (IPTW). Predicted 5-year OS individualized treatment effect (ITE) was stratified using prespecified thresholds: high (>3.5%), moderate (>0%-3.5%), and low predicted benefit (=0%). Robustness was assessed via E-values for unmeasured confounding.

Results: At median follow-up of 67 months (SEER) and 45 months (external), population-level PMRT was associated with improved OS (IPTW hazard ratio 0.801, 95% CI 0.706-0.910; p<0.001). In internal validation (n=1,189), 20.7% were high benefit, 55.8% moderate, and 23.5% low predicted benefit (mean ITE 1.68 percentage points; 76.5% positive). Statistically significant benefit was observed only in the high-benefit subgroup (adjusted HR 0.652, 95% CI 0.429-0.991; p=0.045); neither the moderate (p=0.275) nor low-predicted-benefit subgroup (p=0.330) showed significant association. For the moderate-benefit majority, the model estimated individual benefit averaging 1.6 percentage points, providing patient-specific quantification to support treatment counseling. External ITE distributions were directionally consistent (mean 1.11 percentage points). The high-benefit E-value was 2.50. Predictions were implemented in a publicly available online calculator providing individualized benefit estimates, group classification, and clinical guidance.

Conclusion: A CSF-based framework demonstrated that PMRT benefit on 5-year OS is heterogeneous and individually quantifiable, with significant survival advantage concentrated in a predicted high-benefit minority. These patient-specific estimates can inform treatment intensification for high-benefit patients, de-escalation discussions for low-benefit patients, and shared decision-making for the moderate-benefit majority, operationalized through a publicly available online calculator. Prospective validation is warranted.