Artificial intelligence (AI) and machine learning (ML) are increasingly used to support clinical decision-making. Unimodal models, based on a single data type (e.g., imaging, structured clinical data, genomics), are widely adopted but may fall short in capturing clinical complexity. Multimodal approaches, which integrate multiple heterogeneous data sources, could enhance diagnostic, prognostic, and therapeutic performance. However, their added value remains debated.

This systematic review aims to assess whether multimodal ML algorithms outperform unimodal models in clinical settings. It explores three main axes: (1) the most frequently used data combinations, (2) multimodal integration strategies, and (3) the benefits and limitations of these approaches.

A systematic review of academic papers published from inception to January 2025 was performed. Three reviewers (AB, SO and FL) independently screened each title and abstract. The authors (SO, PB, FL and AB) then performed a full-text analysis for assessment.

Among 352 studies identified in MEDLINE up to January 2025, 97 met the inclusion criteria. Of these, 91% reported better performance with multimodal models, with a median AUC improvement of 0.04 (≥ 0.10 in 10% of cases, Fig. 1 ). The most common combinations involved structured clinical data and imaging (65%), followed by omics-imaging (8%) and signal-clinical data (4%). The main clinical fields were oncology (35%), cardiovascular diseases (14%), and neurology (13%). Integrating free-text data and physiological signals showed particularly promising results. Nevertheless, high methodological heterogeneity was observed (sample sizes, metrics, fusion strategies). Late fusion (66%) was predominant, with early and hybrid fusion methods underrepresented. Real-world implementation remains limited: only 1% referenced FDA- or CE-cleared algorithms, and none included medico-economic evaluation. Reproducibility was also a concern, with only 10% of studies sharing their code.

Multimodal approaches outperform unimodal models in most cases, but heterogeneity, limited reproducibility, and a lack of regulatory validation hinder their clinical adoption.