Multimodal clinical-imaging deep learning model for predicting refractory hypersplenism after liver transplantation.

Refractory hypersplenism after liver transplantation (LT), defined in this study as persistent cytopenia and splenomegaly beyond 6 months after LT, is a clinically important postoperative condition. However, reliable methods for preoperative risk stratification remain limited. This study aimed to identify preoperative risk factors and to develop a multimodal deep learning model integrating clinical and CT-derived imaging features for predicting refractory hypersplenism after LT. Patients who underwent LT between 2013 and 2023 were retrospectively analyzed and classified into remission (n = 120) and refractory (n = 44) hypersplenism groups. Clinical, laboratory, and CT-derived splenic parameters were compared. Radiomic features were extracted from preoperative contrast-enhanced CT images. Logistic regression identified independent predictors, and single-modality (clinical or CT) and multimodal (clinical–radiomic) deep learning models were developed. Model performance was evaluated using AUC, F1-score, and accuracy across training, validation, and test cohorts. Preoperative platelet count (OR = 1.04, p = 0.006), splenic length (OR = 0.96, p < 0.001), and splenic vein diameter (OR = 0.81, p = 0.001) were independent predictors. The multimodal model showed the best overall performance (AUC = 0.857, F1 = 0.842), outperforming logistic regression (P = 0.022). Attention maps highlighted the spleen and splenic hilum were among the most influential regions for prediction. Preoperative splenic morphology and imaging features related to splenic hemodynamic burden were associated with refractory hypersplenism after LT. The proposed clinical–radiomic multimodal deep learning model showed promising predictive performance and may provide useful supplementary information for preoperative risk stratification. Further validation in larger multicenter cohorts is needed before broader clinical application.