Machine learning-based clinical decision support tool for advanced ESCC in the immunotherapy era: a multi-center study

OBJECTIVE: We aimed to develop clinical decision support tools for patients with treatment-naïve advanced esophageal squamous cell carcinoma (ESCC). METHODS: Patients who received first-line immunochemotherapy with or without radiotherapy between 2018 and 2023 across 9 centers were included in the current study. Candidate predictors were routine clinical variables such as patient demographics, tumor characteristics, and treatment details. Feature selection was conducted using the least absolute shrinkage and selection operator (LASSO), Boruta, and stepwise feature selection method. The models were fitted using 4 machine learning algorithms and a Cox proportional hazards model was set as a benchmark, comparing model predictive performance by computing the time-dependent area under the receiver operator curve (tAUC). Concordance index (C-index), Brier score, calibration plots, precision-recall (PR) curves, decision curve analysis (DCA), and held-out testing and external validation were utilized for evaluating model robustness, transportability, and clinical utility. SHapley Additive exPlanations (SHAP) analysis was used to rank feature importance and explain the models. RESULTS: -adjust < 0.05). This superiority persisted in the external validation cohort (all tAUCs > 0.750). The Boruta-RSF model exhibited excellent calibration and the highest clinical net benefit on DCA. A publicly accessible web calculator (https://escc.shinyapps.io/ESCC/) was implemented, enabling individualized overall survival prediction under varying treatment scenarios. CONCLUSIONS: The Boruta-RSF model, leveraging routinely available clinical variables with the corresponding web calculator, facilitates personalized prognostic assessment and optimization of treatment strategies.