Towards precision medicine in valvular heart diseases: a systematic review of predictive multimodal machine learning models

Background Machine learning (ML) can facilitate the integration of vast disparate clinical data for precise clinical care. This is especially crucial in areas such as the diagnosis and management of valvular heart diseases (VHDs), where clinical decision-making is inherently reliant on the synthesis of multiple data sources. Purpose To systematically review studies of predictive machine learning models for VHDs, focusing on the use and prevalence of multimodal models. Methods We conducted a systematic review adhering to the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines. PubMed, Web of Science, and Embase were searched between 2014 and 2025. Eligible peer-reviewed studies that developed predictive ML models for VHDs were included. Titles and abstracts were screened independently by two reviewers, with a third reviewer consulted to resolve any conflicts. Information on the primary predictive task, data modality, model architecture and performance were extracted. Risk of bias was independently assessed by two reviewers using the PROBAST+AI tool. Results A total of 195 studies were included in the review. Left-sided lesions, most notably aortic stenosis, predominated, representing 97% of single-disease models and the majority of cases in multi-disease analyses. Diagnostic prediction models accounted for 62% of the studies, while 38% focused on prognostic prediction. Only 8% of the included studies developed multimodal models. Among those comparing multimodal and unimodal approaches, multimodal models demonstrated an average performance gain of 6.3%. This improvement was more pronounced when integrating structured data with unstructured sources (8.0%) compared to combining unstructured modalities alone (4.0%). The largest increases in performance were observed with mitral regurgitation (MR) diagnosis when tabular clinical data were combined with ECG (+22%) and heart sound recordings (+11%). However, despite these performance gains, only 31% of the multimodal models underwent external validation. Additionally, a subset of studies was identified as 'unimodal with multisource learning', where handcrafted variables from heterogeneous sources were consolidated into a single structured input. Conclusion ML is increasingly used for the diagnosis and management of VHDs. While multimodal models offer notable performance advantages over unimodal approaches, their application in VHD remains limited. Future research should prioritize the integration of heterogeneous data sources alongside rigorous external validation to ensure clinical generalizability and facilitate adoption in valvular care.Figure 1Figure 2