Improving rare disease classification using imperfect knowledge graph.

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Accurate rare disease recognition is crucial for patient care but challenged by limited data for machine learning.
• Knowledge graphs (KGs) offer a way to supplement sparse training data with medical knowledge.
• This research addresses the challenge of utilizing imperfect KGs for rare disease classification.

Purpose of the Study:

• To develop a rare disease classification algorithm that effectively integrates knowledge graph information.
• To enhance machine learning models for rare diseases by leveraging medical knowledge, even from incomplete KGs.

Main Methods:

• A text classification algorithm combining "bag of words" with "bag of knowledge terms" was developed.
• Knowledge terms were identified as terms shared between patient documents and relevant KG subgraphs.
• The algorithm was evaluated on two Chinese disease diagnosis corpora (HaoDaiFu and ChinaRe).

Main Results:

• The proposed algorithm demonstrated robust performance on both datasets.
• It outperformed various baseline methods, including deep learning and feature selection approaches.
• Performance was assessed using macro-averaged F1 score and mean reciprocal rank.

Conclusions:

• Large-scale knowledge graphs can significantly improve rare disease classification models.
• The developed method effectively leverages medical knowledge from KGs, even when they are incomplete.
• This approach offers a promising solution for diagnosing rare diseases with limited labeled data.