Related Experiment Videos
Do Multimodal Vision-Language Models Enhance the Medical Diagnostic Process? A Systematic Review
Lattawat Eauchai1, Laura Otálora González1, Yifan Shi1
1Department of Anesthesiology and Perioperative Medicine, Division of Critical Care, Mayo Clinic, Rochester, MN 55905, USA.
Healthcare (Basel, Switzerland)
|July 15, 2026
Summary
Multimodal vision-language models (VLMs) outperform unimodal models for medical diagnosis. While standalone VLMs show inconclusive results against physicians, copilot models enhance diagnostic accuracy, though further research is needed.
Area of Science:
- Artificial Intelligence in Medicine
- Medical Informatics
- Clinical Decision Support
Background:
- Novel vision-language models (VLMs) integrate patient text and image data for medical diagnosis.
- Conflicting results exist on multimodal VLM performance versus other models and physicians.
- Real-world diagnostic performance of these models requires systematic evaluation.
Purpose of the Study:
- To systematically review the diagnostic performance of multimodal VLMs using both patient text and image data.
- To compare multimodal VLMs against unimodal models and physician performance in hospital settings.
- To assess the quality and risk of bias in studies evaluating VLMs for medical diagnosis.
Main Methods:
- Comprehensive literature search across eight databases (Embase, MEDLINE, SCOPUS) up to December 2025.
- Inclusion of studies on VLMs integrating image and text data for adult patients, compared to other models or physicians.
- Adherence to PRISMA guidelines and assessment using PROBAST + AI tool; protocol registered in PROSPERO.
Main Results:
- 18 studies met inclusion criteria; multimodal VLMs consistently outperformed unimodal models.
- Standalone VLM diagnostic accuracy compared to physicians yielded conflicting evidence.
- VLMs as clinical copilots showed higher accuracy when assisting physicians; meta-analysis was precluded by heterogeneity.
Conclusions:
- Multimodal VLMs surpass unimodal models, but standalone VLM efficacy against clinicians is inconclusive.
- Copilot VLMs demonstrate potential for high diagnostic accuracy, warranting further investigation.
- Methodological limitations (dataset quality, lack of validation) necessitate cautious interpretation and high-quality research for clinical applicability.
