Interpretable multimodal deep learning improves postoperative risk stratification in intrahepatic cholangiocarcinoma in multicentre cohorts.

Abstract

Surgical resection is the primary curative treatment for intrahepatic cholangiocarcinoma (ICC), yet high postoperative recurrence rates pose a significant challenge. We developed an interpretable, transformer-based deep-learning pipeline that integrates multimodal data-including clinical variables, radiomic features, and whole-slide pathology images-by fusing a pre-trained encoder with a transformer network. To biologically validate our model, we leveraged spatial transcriptomics and proteomics to decipher the attention mechanisms underlying its predictions. It demonstrated robust performance in predicting 2-year overall survival, with area under the curve (AUC) values of 0.952 (95% CI: 0.909-0.983), 0.924 (95% CI: 0.804-1.000), and 0.924 (95% CI: 0.828-0.993) in three independent validation cohorts. Interrogation via spatial multi-omics revealed that the model's attention was preferentially focused on regions histologically and molecularly associated with tumor invasion and aggressive behavior. We present a novel, interpretable multimodal deep-learning framework that achieves superior postoperative risk stratification for ICC patients.