PercolationDF: A percolation-based medical diagnosis framework.

Area of Science:

• Medical Informatics
• Computational Biology

Background:

• The shortage and unequal distribution of medical resources pose significant challenges.
• Prevailing diagnostic models often require vast amounts of electronic medical records (EMRs), which are difficult for many institutions to provide.

Purpose of the Study:

• To develop a general diagnosis framework that uses a smaller volume of EMRs.
• To alleviate the data volume requirement for medical diagnosis models.

Main Methods:

• Constructing an EMR-based medical knowledge network (EMKN) using extracted knowledge.
• Expanding the EMKN to represent medical knowledge comprehensively.
• Modifying percolation theory for disease diagnosis on the EMKN.

Main Results:

• The proposed framework outperforms traditional models like naive Bayes, neural networks, and logistic regression, particularly in top-10 recall.
• 51.7% of 207 test cases achieved 100% top-10 recall, a 21% improvement over previous studies.
• The framework effectively addresses data scarcity by inferring knowledge from the EMKN.

Conclusions:

• The developed framework is effective for medical knowledge representation and diagnosis.
• It offers a viable solution for medical institutions facing data volume limitations.
• The framework's potential extends to other domains for knowledge representation and inference.