Causal machine learning personalizes endocrine therapy selection for ductal carcinoma in situ

Background: Standard treatment of ductal carcinoma in situ (DCIS) in post-menopausal women is breast conserving surgery/mastectomy, radiotherapy, and adjuvant endocrine therapy (ET) with either tamoxifen or an aromatase inhibitor. Despite the need to individualize selection of ET based on patient characteristics and side effects, limited tools exist for this purpose. NSABP-B-35, a phase III trial evaluating anastrozole vs tamoxifen in post-menopausal women with DCIS, found that younger patients may benefit more from anastrozole. We hypothesized that causal machine learning methods, which predict treatment effect by accounting for complex interactions between baseline characteristics, would identify treatment benefit more accurately than age alone. Methods: Individual-level data from the NSABP-B-35 trial (n = 3104) were obtained from the NCI NCTN Data Archive, with administrative approval from NCI. Trial eligibility included: DCIS with no invasive component, hormone receptor positive, and previous lumpectomy (with clear margins and negative nodes), followed by whole-breast irradiation. Baseline variables included age, tumor palpability, presence of comedo necrosis, body mass index (BMI), and black race. The outcome was disease-free survival (DFS). The causal machine learning method, a T-learner with accelerated failure time-Bayesian Additive Regression Trees (AFT-BART), predicted the individualized treatment effect (ITE) of anastrozole vs tamoxifen on the difference in restricted mean survival time at 116 months, conditional on patient characteristics. Three repeats of five-fold cross-validation were used to generate out-of-sample predictions for each patient. Statistical significance (p < 0.05) was determined by an aggregate Cauchy association test (ACAT) calculated on the out-of-sample Qini coefficient p-values, a metric that quantifies how well the model orders patients by most to least benefit. Variable importance was quantified using kernelSHAP. Results: The AFT-BART ITE model was able to significantly prioritize patients in order of most to least benefit from anastrozole vs tamoxifen (p-value = 0.03). Predicted ITE ranged from an increase of DFS by 2.9 months to a decrease in DFS by 4.5 months when on anastrozole vs tamoxifen. Younger age, presence of comedo necrosis, and black race best predicted anastrozole benefit. No statistically significant heterogeneity was found when patient age alone was used to rank patients from most to least benefit (p-value = 0.6). Conclusions: Our causal machine learning model accurately predicted who would benefit from anastrozole vs tamoxifen, outperforming treatment selection based on age alone. Once validated, this model may help clinicians optimize treatment selection for post-menopausal women with DCIS. This work was supported by NLM 5T15LM007359.