Transfer learning for non-image data in clinical research: A scoping review

Area of Science:

• Machine Learning in Healthcare
• Clinical Data Science
• Artificial Intelligence in Medicine

Background:

• Transfer learning reuses pre-trained models for new tasks, gaining traction in medical imaging.
• Its application to clinical non-image data remains less explored.
• This scoping review investigates transfer learning's use with non-image clinical data.

Purpose of the Study:

• To explore and summarize the current use of transfer learning for non-image data in clinical research.
• To identify trends and applications of transfer learning in diverse medical specialties.
• To highlight areas for future development and adoption.

Main Methods:

• Systematic search of medical databases (PubMed, EMBASE, CINAHL) for relevant peer-reviewed studies.
• Inclusion of 83 studies focusing on transfer learning applied to human non-image clinical data.
• Analysis of data types, model applications, author affiliations, data sources, and code sharing.

Main Results:

• A significant increase in transfer learning applications for non-image data in recent years.
• Dominant use in time series data (61%), followed by tabular, audio, and text data.
• 40% of studies transformed non-image data into images (e.g., spectrograms) for model application.
• Many studies utilized public datasets (66%) and models (49%), but code sharing was limited (27%).

Conclusions:

• Transfer learning adoption in clinical non-image data research is rapidly expanding.
• Demonstrated potential across various medical fields, highlighting its versatility.
• Emphasizes the need for increased interdisciplinary collaboration and reproducible research practices.