Unimodal to multimodal: a systematic review of predictive machine learning models for valvular heart diseases

We aimed to synthesize existing evidence on predictive machine learning (ML) models for valvular heart disease (VHD) and examine how these models have been applied across clinical tasks, data modalities and validation settings.

ML is gaining traction for improving cardiovascular care, particularly in the management of VHDs. However, empirical evidence on how ML models handle the multimodal complexity of valvular pathologies remain sparse.

We conducted a systematic review according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines searching PubMed, Web of Science, and Embase from 2014 to 2025. We included articles that developed ML for clinical prediction in VHD patients. (PROSPERO: CRD42025644167).

We identified 195 studies that met the inclusion criteria. Seventy-five studies (38.5%) developed single-lesion models for aortic stenosis. Retrospective datasets were used in 86% of the included studies and 79% relied on internal validation. Sixteen studies (8.2%) developed multimodal models, integrating different types of ML input data. The multimodal models demonstrated a 6.3 percentage point increase in average performance across tasks compared to their unimodal counterparts within the same cohort.

Across the literature, unimodal ML models for VHDs demonstrate promising performance for disease detection, patient stratification, and risk prediction, but multimodal approaches are emerging with potential advantages for procedural planning and outcome forecasting. Translation to clinical practice will require large, multicenter datasets to validate and standardize data-driven VHD management.

https://www.crd.york.ac.uk/PROSPERO/view/CRD42025644167.