Risk prediction model for brain metastasis risk among patients diagnosed with primary gastrointestinal cancers.

BACKGROUND: The risk of radiotherapy-treated brain metastasis (BrM) in gastrointestinal (GI) cancers is poorly defined. We aimed to determine its incidence and develop a clinically interpretable prediction model. METHODS: Patients with GI cancers diagnosed between 2005-2020 were identified from our registry. Receipt of brain radiotherapy was used to define radiotherapy-treated BrM. Logistic regression was performed on a training cohort to identify BrM predictors, which were then used to construct a risk score, and validate performance in a held-out cohort. Model performance was assessed with receiver operating characteristic (ROC) analysis, calibration, Brier score, and Multiple imputation by chained equations (MICE) sensitivity analysis. RESULTS: Among 14,863 eligible patients, the 15-year cumulative incidence of clinically treated BrM was 1.89%. Median OS following BrM was 3.7 months. Logistic regression was performed in 11,022 patients with complete data. In the complete-case training cohort (n = 7348), significant predictors included primary site, clinical stage, chemotherapy, and recurrence at presentation. The AUC was 0.80, with good calibration and Brier score 0.012. Medium- and high-risk groups had higher BrM risk in validation (HR 4.57 and 11.35, respectively). CONCLUSIONS: We developed an internally validated prediction model for BrM in GI cancers. External validation is required.