Explainable Prediction of Medical Codes With Knowledge Graphs

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Documentation

Background:

• Existing International Classification of Diseases (ICD) coding models lack interpretability and struggle with uneven label distribution and inter-label dependencies.
• Current models often fail to capture the nuanced reasons and details behind specific medical diagnoses, limiting their explanatory power.

Purpose of the Study:

• To introduce G_Coder, a novel method for medical code prediction that enhances interpretability by incorporating knowledge graphs and attention mechanisms.
• To address the limitations of existing ICD coding models, including obscurity of diagnostic reasoning and label distribution challenges.

Main Methods:

• Developed G_Coder, a system combining Multi-CNN, graph representation, attentional matching, and adversarial learning.
• Constructed a medical knowledge graph from Freebase, encompassing entities like diseases, symptoms, medicines, surgeries, and examinations for ICD-9.
• Utilized adversarial learning to generate samples that reconcile diverse clinical writing styles.

Main Results:

• G_Coder achieved a micro-F1 score of 69.2% on the MIMIC-III dataset, outperforming state-of-the-art methods.
• The knowledge graph and attention mechanism enabled explanations for relevant segments of medical codes.
• Demonstrated improved precision in ICD code prediction compared to existing approaches.

Conclusions:

• G_Coder effectively integrates information from medical records and embeds knowledge into ICD codes for enhanced prediction.
• The proposed method significantly improves the precision of medical code prediction and offers better interpretability.
• G_Coder has the potential to reduce the workload for human coders by providing more accurate and explainable coding suggestions.