Integrating Machine Learning and Tumor Immune Signature to Predict Oncologic Outcomes in Resected Biliary Tract Cancer.

BACKGROUND Improved methods are needed to predict outcomes in biliary tract cancers (BTCs). We aimed to build an immune-related signature and establish holistic models using machine learning. METHODS Samples were from 305 BTC patients treated with curative-intent resection, divided into derivation and validation cohorts in a two-to-one ratio. Spatial resolution of T cell infiltration and PD-1/PD-L1 expression was assessed by immunohistochemistry. An immune signature was constructed using classification and regression tree. Machine learning was applied to develop prediction models for disease-specific survival (DSS) and recurrence-free survival (RFS). RESULTS The immune signature composed of CD3+, CD8+, and PD-1+ cell densities and PD-L1 expression within tumor epithelium significantly stratified patients into three clusters, with median DSS varying from 11.7 to 80.8 months and median RFS varying from 6.2 to 62.0 months. Gradient boosting machines (GBM) outperformed rival machine-learning algorithms and selected the same 11 covariates for DSS and RFS prediction: immune signature, tumor site, age, bilirubin, albumin, carcinoembryonic antigen, cancer antigen 19-9, tumor size, tumor differentiation, resection margin, and nodal metastasis. The clinical-immune GBM models accurately predicted DSS and RFS, with respective concordance index of 0.776-0.816 and 0.741-0.781. GBM models showed significantly improved performance compared with tumor-node-metastasis staging system. CONCLUSIONS The immune signature promises to stratify prognosis and allocate treatment in resected BTC. The clinical-immune GBM models accurately predict recurrence and death from BTC following surgery.