The Application of Multimodal Data Fusion Algorithm MULTINet in Postoperative Risk Assessment of TAVR

<b>Background:</b> Transcatheter aortic valve replacement (TAVR) has emerged as a pivotal minimally invasive interventional therapy for aortic valve disease and has seen increasingly widespread clinical adoption in recent years. Despite its overall safety, the adverse events and even deaths in the postoperative period still account for a certain percentage. Accurate identification of high-risk patients is therefore critical for optimizing preoperative decision making, guiding individualized treatment strategies and improving long-term outcomes. However, existing scoring systems and predictive models fail to fully leverage multimodal clinical data from patients, resulting in suboptimal predictive accuracy that falls short of the demands of precision medicine, indicating substantial room for improvement. <b>Methods:</b> In this study, a multimodal deep learning model named MULTINet (multimodal learning for TAVR risk network) was constructed using data from the MIMIC-IV (Medical Information Mart for Intensive Care) cohort. This model achieved unimodal and multimodal modeling through a dual-branch structure, and, by using an attention pooling fusion module, flexibly handled the input that contained missing modalities, to predict the 30-day all-cause mortality in TAVR patients. The area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUPR) and the recall rate were used for prediction evaluation. The calibration degree was evaluated by calibration diagrams and Brier scores, and its clinical practicability was assessed through decision curve analysis (DCA). And the integrated gradient method was used to identify key predictive features to enhance interpretability of the model. <b>Results:</b> In the postoperative 30-day all-cause mortality prediction task, the MULTINet method achieved an AUC value of 0.9153, AUPR value of 0.5708 and Recall value of 0.8051, which was significantly superior to the XGBoost method (AUC 0.8958, AUPR 0.4053 and Recall 0.7793) and the MedFuse method (AUC 0.5571, AUPR 0.2487 and Recall 0.3089). The MULTINet method demonstrated more robust and reliable probability estimation performance, with a Brier score of 0.0269, outperforming XGBoost (0.0343) and MedFuse (0.2496). It achieved a higher net benefit in decision analysis, reflecting its effectiveness in strategy optimization and actual decision-making benefits. The renal function, cardiac function and inflammation-related indicators contributed greatly in the prediction process. <b>Conclusions:</b> The multimodal deep learning model proposed in this study named MULTINet enables adaptive integration of multimodal clinical information for predicting all-cause mortality within 30 days post-TAVR, substantially improving both predictive accuracy and clinical applicability, providing robust support for clinical decision making and boosting TAVR management toward greater precision and intelligence.