Multimodal Machine Learning for Stroke Prognosis and Diagnosis: A Systematic Review

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Stroke poses significant mortality and sensorimotor deficit risks.
• Machine learning (ML) is increasingly used for stroke outcome prediction.
• Multimodal ML is gaining traction due to diverse clinical data types (images, bio-signals, clinical data).

Approach:

• Systematic literature review following PRISMA guidelines.
• Focused on state-of-the-art multimodal ML methods for stroke prognosis and diagnosis.
• Analyzed dominant fusion paradigms (early, joint, late) and less explored methods (translation, alignment).

Key Points:

• Fusion techniques (early, joint, late) dominate current multimodal ML for stroke.
• Opportunities exist in exploring multimodal translation and alignment.
• Current datasets often lack diversity in scale and modality types.

Conclusions:

• Advancements in multimodal ML are crucial for improving stroke diagnosis and prognosis.
• Developing more diverse multimodal datasets is essential.
• Further research into novel multimodal learning paradigms is recommended.