Is Multimodal Better? A Systematic Review of Multimodal versus Unimodal Machine Learning in Clinical Decision-Making

Machine learning has demonstrated success in clinical decision-making, yet the added value of multimodal approaches over unimodal models remains unclear. This systematic review evaluates studies comparing multimodal and unimodal ML algorithms for diagnosis, prognosis, or prescription. A comprehensive search of MEDLINE up to January 2025 identified 97 studies across 12 medical specialties, with oncology being the most represented. The most common data fusion involved tabular data and images (67%). A risk of bias assessment using PROBAST revealed that 57% of studies had a low risk of bias, while 41% had a high risk. Multimodality outperformed unimodality in 91% cases. No correlation between dataset sample size and added performance has been observed. However, considerable methodological heterogeneity and potential publication bias warrant caution in interpretation. Further research is needed to refine evaluation metrics and hybrid model architectures based on specific clinical tasks.