Learning a Health Knowledge Graph from Electronic Medical Records

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Knowledge representation

Background:

• Increasing demand for clinical decision support systems and symptom checkers.
• Limitations of current knowledge bases (manual compilation or simple statistics).
• Need for automated, high-quality health knowledge graph construction.

Purpose of the Study:

• To explore an automated process for learning disease-symptom knowledge graphs from electronic medical records.
• To evaluate probabilistic models for knowledge graph construction.
• To validate the performance against existing knowledge graphs and expert opinions.

Main Methods:

• Extraction of medical concepts from 273,174 de-identified patient records.
• Construction of knowledge graphs using logistic regression, naive Bayes, and a noisy OR Bayesian network.
• Evaluation and validation against Google's knowledge graph and physician experts.

Main Results:

• Feasible automated construction of high-quality health knowledge graphs from medical records.
• The noisy OR model achieved 0.85 precision and 0.6 recall in clinical evaluation.
• Noisy OR model significantly outperformed all other tested models (p < 0.01).

Conclusions:

• Automated construction of high-quality health knowledge graphs from electronic medical records is achievable.
• Probabilistic models, particularly the noisy OR Bayesian network, are effective for this task.
• This approach offers a scalable and efficient method for building medical knowledge bases.