Validating the knowledge represented by a self-organizing map with an expert-derived knowledge structure

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Knowledge Representation

Background:

• Machine learning (ML) adoption in healthcare is hindered by a lack of interpretability.
• Visualizations of medical literature can distill vast information but often lack clear meaning.
• Validating ML-derived insights is crucial for their acceptance in fields like medical curriculum development.

Purpose of the Study:

• To validate the interpretability of a self-organizing map (MedSOM) visualization.
• To assess MedSOM's ability to coherently summarize psychiatric knowledge.
• To link ML outputs to established knowledge standards in psychiatry.

Main Methods:

• A self-organizing map (MedSOM) was trained on Medline/PubMed indexed articles.
• Reference lists from ten editions of a core psychiatric textbook were analyzed.
• K-means clustering was applied to textbook references projected onto the MedSOM.

Main Results:

• MedSOM consistently identified six distinct psychiatric knowledge domains across ten textbook editions (1967-2017).
• Clustering revealed coherent organization at the level of broad psychiatric practice areas.
• The identified domains included General/Adult Psychiatry, Child Psychiatry, and Administrative Psychiatry.

Conclusions:

• The study validates MedSOM's ability to represent and stabilize psychiatric knowledge domains.
• This demonstrates a method for validating ML-driven visualizations of medical literature.
• Successful validation enhances trust and facilitates the use of ML insights in medical education and curriculum development.