TLDR: A new research paper introduces patient-level knowledge graphs (KGs) to improve automated ICD coding. These KGs efficiently represent patient data (23% of original text, 90% information retained) and, when integrated into a state-of-the-art model, boost coding accuracy (up to 3.20% Macro-F1 improvement) and provide better, more concise explanations for predictions.
Automating the process of assigning International Classification of Diseases (ICD) codes to clinical documents is a crucial task in healthcare. These codes provide structured data vital for research, hospital administration, and enhancing patient care. However, manual coding is time-consuming, prone to errors, and difficult to scale. This challenge is compounded by the vast and uneven distribution of ICD codes, making automation particularly complex.
While previous efforts have focused on using external knowledge to improve the representation of the output codes, the potential of structured information for representing the input clinical documents themselves has been largely overlooked. A recent research paper, “Structured Information Matters: Explainable ICD Coding with Patient-Level Knowledge Graphs,” introduces an innovative approach to address this gap. You can read the full paper here.
Patient-Level Knowledge Graphs: A New Approach
The core of this research lies in creating patient-level knowledge graphs (KGs). These graphs offer a comprehensive, structured view of a patient’s condition by extracting key information from clinical documents. Instead of processing raw text, the system converts patient records into a network of entities (like diseases, treatments, tests, and drugs) and their relationships. This structured representation is remarkably efficient, distilling the essential information from patient notes into a more concise format, retaining 90% of the original information while reducing the content size by 77%.
The knowledge graphs are constructed using named-entity recognition (NER) and relation extraction (RE) models from the Healthcare NLP library. These models identify 14 distinct types of entities and connect them through five types of relationships: Diagnostic Information (interrelationships among problems, treatments, tests), Temporal Information (sequence of clinical events), Posology Information (drug regimens), and Anatomical Information (connections between problems/directions and body parts).
Enhanced ICD Coding Performance and Explainability
To assess the effectiveness of these patient-level knowledge graphs, the researchers integrated them into PLM-ICD, a state-of-the-art ICD coding architecture. Experiments conducted on popular benchmarks, MIMIC-III Full and MIMIC-III Top-50 datasets, demonstrated significant improvements. The model achieved improved Macro-F1 scores by up to 3.20% on the Top-50 dataset and 1.36% on the Full dataset. This enhancement is attributed to the rich, structured information provided by the knowledge graphs.
Beyond performance, the approach also significantly improves the explainability of automated ICD coding. By analyzing the impact of different entities and relationships within the graph, the researchers found that ‘Clinical Relationship’ and ‘Problem’ entities were most critical for accurate coding. The model can highlight specific text spans that contribute most to a prediction, offering clear and concise evidence for its assignments. For instance, in cases related to ‘Arterial Catheterization’, the model accurately identified keywords like ‘hypotensive’, ‘blood pressure’, ‘invasive procedure’, and ‘femoral line’, providing a transparent rationale for its predictions.
Also Read:
- Enhancing Clinical Reasoning in AI: A Compartmentalized Approach for Language Models
- Small Language Models Show Promise in Formal Logic Reasoning for Ontology Engineering
Future Implications for Healthcare
This work highlights the immense potential of structured information in improving automated clinical coding. By providing a more accurate, efficient, and explainable system, patient-level knowledge graphs can offer healthcare providers more precise, data-driven insights. This could lead to better treatment plans, earlier diagnoses, and overall enhanced patient outcomes, ultimately alleviating the burden of manual coding and improving the availability and accuracy of structured clinical data.
