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

The document is the protocol of a scoping review.

Objective:

The objective of this scoping review is to explore and characterize studies using transfer learning for non-image data in clinical research.

Introduction:

Transfer learning is the use of a pre-trained model for a task different to what it was originally trained for. The method has garnered considerable attention in recent years, especially in computer vision and natural language processing. However, despite the increasing interest in machine learning among clinical researchers, the use of transfer learning is not well studied in the medical literature, especially for non-image data.

Inclusion criteria:

We will include clinical studies, published in English in the medical literature, that used transfer learning (fine-tuning or feature-representation transfer) for the analysis of non-image data (e.g. tabular, time series, text, audio). Studies using synthetic data of these types will also be included if they represent or are based on human participants.

Methods:

We will search PubMed, EMBASE and CINAHL for peer-reviewed articles (original articles and brief reports). Each abstract/full-text article will be screened by at least two independent reviewers. Study characteristics will be extracted using a piloted data extraction form. Results will be presented using descriptive statistics and visualizations.