Integrating multimodal data with artificial intelligence to predict postoperative recurrence in intrahepatic cholangiocarcinoma after curative resection.

Abstract

e16294 Background: Intrahepatic cholangiocarcinoma (ICC) is the second most common primary liver malignancy and carries a poor prognosis. Although curative resection remains the only potentially curative treatment, nearly half of patients experience postoperative recurrence, which substantially compromises long-term survival. Therefore, we aimed to develop and validate a real-world, multimodal artificial intelligence (AI)–based prognostic framework that integrates imaging, pathology, and clinical information while accommodating incomplete data, to enable more precise prediction of recurrence after curative resection in ICC. Methods: This retrospective multicenter study included 426 patients with pathologically confirmed ICC who underwent curative resection at 10 medical centers between January 2013 and August 2024. The multimodal data included clinical variables, preoperative contrast-enhanced CT (CECT) images, and hematoxylin and eosin (H&E)–stained whole-slide pathology images (WSIs). Patients were randomly assigned to training and validation cohorts at an 8:2 ratio. An AI–based multimodal prognostic framework was developed to integrate heterogeneous data sources for postoperative recurrence prediction. To address the pervasive issue of incomplete multimodal data in real-world clinical settings, modality-specific attention mechanisms were incorporated to capture salient features within each modality. Cross-modal interaction layers were employed to model inter-modality relationships and extract complementary information, enabling compensation for missing modalities. Model performance was evaluated in terms of discrimination, calibration, and clinical utility. Results: The multi-model demonstrated superior performance for recurrence prediction. It achieved an AUC of 0.912 (95% CI: 0.831–0.926) in the training cohort and 0.893 (95% CI: 0.834–0.907) in the validation cohort, outperforming models based on radiomics alone or pathomics alone (AUCs of 0.874 [95% CI: 0.789–0.960] and 0.79 [95% CI: 0.687–0.916], respectively). Moreover, the multimodal model showed good calibration and yielded greater net clinical benefit on decision curve analysis, supporting its potential utility for clinical risk stratification. Conclusions: This study developed a multimodal AI framework designed for real-world clinical settings with incomplete data. The framework integrates CECT imaging, histopathological features, and clinical variables. By leveraging complementary macroscopic and microscopic tumor characteristics, it enables accurate prediction of postoperative recurrence in ICC.