A clinical treatment recommender system optimizing adjuvant chemoradiotherapy benefit in Chinese women with breast cancer using interpretable causal Bayesian networks

Background: Breast cancer (BC) has surpassed lung cancer in becoming the most common cancer among women worldwide. However, few studies have investigated multivariate BC prognosis outcomes. This study aimed to develop and validate a causal model based on Bayesian networks (BN) that can simultaneously predict long-term risk of all-cause mortality and recurrence among Chinese women with BC, and further to estimate individualised expected benefits of model-based treatment recommendations. Methods: This study used data from adult women who were diagnosed with primary invasive T1-4 N0-3 M0 BC at the Affiliated Zhongshan Hospital of Dalian University, Dalian, Liaoning, China, between January 2011 and December 2017. After the whole cohort was randomly divided into 70% development and 30% validation cohorts, a causal BN (CBN) model was built by aggregating extensive expert knowledge and data-driven approaches to arrive at a consensus structure for causal inference, and subsequently validated its ability to predict both overall survival (OS) and recurrence free survival (RFS) by using the area under the curve (AUC), Brier score (BS), accuracy and predicted/observation ratio. Variables included sociodemographics; preoperative clinical, histopathological and molecular biomarkers; operative variables; and postoperative treatment variables. OS and RFS in women with BC were the main outcomes of this study and were assessed at the most recent follow-up on December 31, 2022. Findings: A total of 3512 patients, which consisted of 2458 patients in the development cohort (mean [SD] age, 54.60 [10.68] years) and 1054 patients in the validation cohort (mean [SD] age, 53.77 [11.08] years), were eligible for this study. A total of 175 (5.0%) patients with BC died, and 344 (9.8%) experienced recurrence after at least 5-year follow-up. A CBN model was developed that included the following predictors: age group; preoperative CEA, CA 125 and CA15-3 serum levels; neoadjuvant chemoradiotherapy; surgery type; tumour grade and histology; immunohistochemical expression of oestrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2) amplification, and Ki-67; pT and pN stages; and postoperative treatment with chemotherapy, radiotherapy and endocrine therapy. In the validation cohort, the CBN model had excellent performance with desirable AUCs of 0.92 (95% CI 0.87-0.98) and 0.73 (0.62-0.83) and low BSs of 0.025 and 0.062 for OS and RFS, respectively. The model also achieved excellent accuracy and model calibration for OS and RFS, with accuracies of 96.4% and 92.8% and predicted/observed ratios of 1.01 (0.92-1.11) and 1.05 (0.95-1.15), respectively. Furthermore, the chemoradiotherapy treatment according to our CBN model recommendations was associated with notably better prognosis, with a hazard ratio of 0.63 (0.44-0.90) for OS and 0.81 for RFS (0.75-0.87), respectively. Interpretation: In this prognostic study based on clinical real-world data, a CBN model was developed that accurately predicted two important outcomes for women with BC, and it has the potential to identify patients who could benefit from chemoradiotherapy. The CBN can effectively capture complex interrelationships among risk factors and identify potential causal pathways underlying tumour progression through its graphical representation, and could also provide an adjuvant treatment recommendation system for BC women.