Deep learning for rare disease: A scoping review.

Area of Science:

• Medical Informatics
• Computational Biology
• Genomics

Background:

• Over 7,000 rare diseases affect 10% of patients, significantly impacting quality of life and incurring societal costs.
• Low prevalence of individual rare diseases presents challenges in diagnosis, patient care, and research participation.
• Deep learning (DL) shows promise in advancing various scientific fields, including healthcare applications.

Purpose of the Study:

• To review the current applications of deep learning in advancing rare disease research.
• To identify trends in deep learning methodologies and disease areas within rare disease research.
• To summarize challenges and outline future research directions for DL in rare diseases.

Main Methods:

• Systematic review of 332 articles on deep learning applications in rare disease research.
• Categorization of studies by rare disease type (neoplastic, genetic, neurological) and DL architecture used.
• Analysis of research focus, with a primary emphasis on diagnostic applications.

Main Results:

• Deep learning is extensively applied to rare neoplastic diseases (250/332), rare genetic diseases (170/332), and rare neurological diseases (127/332).
• Convolutional neural networks (307/332) are the most prevalent DL architecture, likely due to the common availability of image data.
• Diagnosis is the predominant focus of deep learning research in rare diseases (263/332).

Conclusions:

• Deep learning demonstrates significant potential to overcome challenges in rare disease diagnosis and research.
• Future research should focus on leveraging DL to improve patient care, facilitate research participation, and accelerate treatment development for rare diseases.
• Continued exploration of DL architectures and data types is crucial for maximizing its impact on rare disease research.