Utilizing prior-data-fitted networks and in-context learning: a transformer-based tabular foundation model for predicting symptomatic intracranial hemorrhage after successful recanalization

Background and purpose Symptomatic intracranial hemorrhage (sICH) is a serious complication after endovascular thrombectomy (EVT) and is strongly associated with poor outcomes in acute ischemic stroke. Existing risk scores show limited predictive accuracy. This study aims to develop and externally validate a transformer-based model for predicting sICH following successful recanalization in anterior circulation large vessel occlusion. Methods A total of 661 EVT-treated patients were retrospectively analyzed as the derivation cohort, and 261 patients from another tertiary center were included as the external test cohort. A tabular prior-data-fitted network (TabPFN), a transformer-based foundation model, was constructed using angiographic biomarkers (basal ganglia blush, early venous filling), baseline ASPECT score, fasting blood glucose, collateral status, and the number of retriever passes. Logistic regression and XGBoost were also developed. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), precision, recall, and F1 score, and subsequently compared with established scores (ASIAN, TAG, IER-sICH). Results In internal validation, TabPFN achieved an AUC of 0.948, comparable to XGBoost (0.953) and logistic regression (0.944), but superior to ASIAN (0.786), IER-sICH (0.687), and TAG (0.670). In external validation, TabPFN demonstrated the highest AUC (0.955), significantly outperforming existing scores (all p < 0.05), and exhibited the best F1 score and precision across cohorts. Conclusion The TabPFN model effectively predicts the risk of sICH in Chinese stroke cohorts, enabling real-time risk stratification for antithrombotic therapy and postoperative blood pressure management.