Multimodal machine learning in precision health: A scoping review

Area of Science:

• Biomedical Machine Learning
• Health Informatics
• Data Science in Medicine

Background:

• Machine learning (ML) is increasingly used for clinical decision support, traditionally with single-source data.
• Fusing diverse data types (multimodal ML) aims to enhance predictive accuracy and mimic clinical expert decision-making.
• Existing research has focused on technical aspects of data fusion with limited exploration of clinical integration.

Purpose of the Study:

• To systematically review and summarize current studies on multimodal data fusion in health.
• To identify key trends, common applications, and prevalent methodologies in the field.
• To highlight gaps in research and suggest future directions for multimodal ML in healthcare.

Main Methods:

• A scoping review was conducted following the PRISMA extension for Scoping Reviews.
• Searches were performed in PubMed, Google Scholar, and IEEEXplore databases from 2011 to 2021.
• 128 articles were included, focusing on multimodal data fusion for health-related diagnosis and prognosis.

Main Results:

• Neurology and oncology are the most frequent application areas for multimodal ML.
• Early fusion, a strategy of merging data at the initial stage, is the most common approach.
• Multimodal data fusion demonstrated improved predictive performance, with an average accuracy increase of 6.4% compared to unimodal methods in studies that performed such comparisons.

Conclusions:

• Multimodal machine learning offers enhanced predictive capabilities in healthcare but faces challenges in scalability and data processing time.
• Key areas for future research include developing clear clinical deployment strategies and obtaining regulatory approval (e.g., FDA).
• Further investigation is needed into how multimodal approaches can mitigate biases and reduce healthcare disparities across diverse subpopulations.