Estimating individual benefit from adjuvant chemotherapy in hormone receptor–positive breast cancer using causal multimodal artificial intelligence.

Abstract

551 Background: Adjuvant chemotherapy decisions in HR+/HER2– early breast cancer are often guided by prognostic gene expression assays. However, assessment of recurrence risk may not yield a predictive biomarker that accurately estimates treatment benefit. Here, we combine advances in deep learning and causality research to address this problem using routine histopathology and clinical data. We develop and validate a causal multimodal artificial intelligence (AI) model that predicts the patient-specific benefit of adding chemotherapy to endocrine therapy in HR+/HER2– early breast cancer. Methods: De-identified whole-slide histopathology images (WSIs) and clinical covariates from 9,269 patients across 12 observational cohorts were used to develop the Ataraxis Breast model (ATX). The locked model was then externally validated in a cohort from the University of Chicago (n = 435, stage I-III HR+/HER2– breast cancer, median age = 56, median follow-up = 7.2 years), grouped by adjuvant therapy received (endocrine therapy [ET] n = 322, chemoendocrine therapy [CET] n = 113). The primary endpoint was 5-year recurrence-free interval (RFI). WSIs were encoded via a pathology foundation model and integrated with clinical variables (including T/N stage, age, ductal vs lobular histology). ATX predicted counterfactual RFIs, assuming CET or ET, with the difference taken as the predicted treatment-benefit score. Discrimination was evaluated using Harrell’s C-index. Multivariate Cox proportional hazards models were fitted, adjusting for clinicopathological factors (age and T/N stage), to estimate hazard ratios (HR). Subgroup differences were assessed using inverse-propensity-weighted Kaplan-Meier estimates and two-sided log-rank tests. Results: In the external validation cohort, ATX was significantly associated with the primary endpoint (adjusted HR = 1.65, 95% CI 1.05-2.6; p = 0.029), and demonstrated good discrimination at the patient-level (C-index = 0.704, 95% CI 0.588-0.821). Notably, ATX stratified patients into subgroups with differential chemotherapy effects: patients in the top tertile of predicted treatment benefit experienced improved RFI with CET (ET RFI = 0.744, CET RFI = 0.980; p < 0.001), with no difference observed in patients with low predicted treatment benefit (ET RFI = 0.975, CET RFI = 0.970; p = 0.85). The interaction between ATX and the magnitude of chemotherapy benefit, when adjusted for clinicopathological factors, was significant (p-interaction = 0.005). Conclusions: A causal multimodal AI model accurately estimates chemotherapy benefit for patients with HR+/HER2– early breast cancer, meeting the criteria for a predictive biomarker. The causal AI methodology presented here may provide a generalizable framework to optimize scalable, predictive biomarkers for other therapies and cancer types.