Machine learning (ML) offers promising tools to support clinical decision-making, particularly in vascular medicine where data heterogeneity is a key challenge. Unimodal ML models, built on a single data source, may fail to capture the full complexity of clinical scenarios. This work investigates the added value of multimodal ML models for preventive vascular care.

A systematic review of the literature was performed on PubMed/MEDLINE from inception to January 2025. We included original studies that compared multimodal versus unimodal ML models for diagnosis, prognosis, or therapeutic decision-making, specifically in clinical settings. Three reviewers independently screened articles using the Rayyan tool, with consensus resolution for conflicts. Extracted data included model architecture, data modalities, performance metrics, and external validation practices.

Ninety-seven studies were included across 12 specialties, including vascular medicine. Neural networks were the most frequent algorithms used (57%). Multimodal models integrate tabular data, imaging, signal data, and free text. Structured data combined with imaging was the most common pairing (67%). In 91% of cases, multimodal models outperformed unimodal baselines, but the median AUC improvement was modest (0.04). Only 33% of studies performed external validation. High heterogeneity was observed across datasets, metrics, and fusion strategies.

Despite the high rate of reported performance gains, the modest median improvement and the lack of external validation raise concerns about the translational readiness of these models. The use of real-world data and context-specific architectures remains limited.

Multimodal ML shows promise in vascular medicine, but rigorous external validation and domain-specific model design are needed to translate these gains into clinical impact.