TL;DR
This paper introduces MKE-Coder, a novel framework for automatically assigning ICD codes to Chinese Electronic Medical Records (EMRs). It tackles the unique challenges of Chinese EMRs by leveraging multi-axial ICD knowledge and verifying candidate codes with clinical evidence, significantly improving coding accuracy and interpretability.
What Problem Does This Paper Address?
The International Classification of Diseases (ICD) is crucial for standardizing disease diagnoses in healthcare. While automatic ICD coding has seen success with English EMRs, it faces significant hurdles with Chinese EMRs. These challenges include the concise and unique writing style of Chinese EMRs, making it difficult to extract relevant disease information. Furthermore, existing methods often overlook the multi-axial nature of ICD codes (e.g., etiology, anatomical site, pathology) and fail to link diagnostic information to concrete clinical evidence, leading to inaccurate and less interpretable coding. This paper aims to bridge this gap by developing a system that effectively handles Chinese EMRs by incorporating both multi-axial knowledge and evidence verification.
Methods
MKE-Coder employs a multi-stage approach:
- Candidate Code Identification: For each diagnosis in an EMR, the system identifies the top N most probable ICD candidate codes. This is done by calculating a similarity score based on edit distance, TF-IDF, and word2vec embeddings between the diagnosis and potential ICD codes.
- Multi-Axial Knowledge Parsing: A "multi-axial parser" is used to break down each candidate ICD code into its constituent knowledge components across four axes: etiology, anatomical site, pathology, and clinical manifestations.
- Evidence Retrieval and Evaluation: The system retrieves relevant clinical evidence from the EMR that supports the knowledge associated with each axis of a candidate code. This retrieval is guided by a "prior knowledge table" that maps ICD code groups to likely evidence locations within Chinese EMRs. A novel scoring model, Clinical-Simbert, is then used to evaluate the reliability and relevance of the retrieved evidence.
- Inference with Prompt-Tuning: The core of the inference process uses a prompt-tuning strategy adapted for masked language modeling (MLM). A template is constructed that combines the candidate ICD code description with its verified evidence. This template is fed into a T5 model, which then predicts whether the evidence sufficiently supports the candidate code. This transforms the binary classification task into an MLM problem, allowing for robust prediction.
Datasets
The authors utilized a large-scale Chinese EMR dataset collected from various hospitals across four cities. This dataset is described as the "Chinese medical insurance version." They also reference the MIMIC-III dataset for comparative analysis of EMR characteristics, highlighting the differences in diagnosis and discharge summary lengths between Chinese and English EMRs.
Key Insights and Ideas
- Chinese EMRs are fundamentally different: The paper strongly emphasizes that Chinese EMRs are significantly shorter and more concise than their English counterparts (like MIMIC-III), rendering direct application of English-centric methods ineffective.
- Multi-axial knowledge is key: ICD codes aren't just hierarchical; their structure reflects distinct clinical dimensions (etiology, site, etc.). MKE-Coder's success hinges on explicitly parsing and verifying evidence for each of these axes.
- Evidence is paramount: Simply relying on discharge summaries is insufficient. The paper demonstrates that incorporating verified clinical evidence significantly boosts coding accuracy and interpretability.
- Clinical-Simbert for evidence quality: A dedicated model is needed to assess the credibility of retrieved evidence, ensuring that only reliable information influences the final coding decision.
- Prompt-tuning for verification: Framing the evidence verification as an MLM task via prompt-tuning proves effective for making robust binary decisions about code support.
Key Results
- Superior Performance: MKE-Coder significantly outperforms several strong baselines (LAAT, JointLAAT, MSMN, T5) on their Chinese EMR dataset, achieving an average F1 score of 72.61% and accuracy of 74.32% when using evidence sets.
- Evidence Boost: Using evidence sets as input, rather than just discharge summaries, improved F1 scores by an average of 12.08% and accuracy by 15.15% across baseline methods.
- Practical Impact: In simulated coding scenarios, MKE-Coder assisted novice coders in improving accuracy by up to 8.65% and speed by up to 34.15% compared to independent coding. Intermediate coders also saw improvements in speed by up to 37.38%.
- Interpretability: ICD coders found that MKE-Coder's evidence explanations led to a higher level of "fully supported" evidence (48% after assessment vs. 21% before), increasing trust and acceptance.
Limitations and Future Work
- Dataset Specificity: While the dataset is large, it's from a specific "Chinese medical insurance version." Future work could explore broader datasets or different versions of Chinese ICD coding.
- Computational Cost: Training the Clinical-Simbert model took approximately 156 hours on a TITAN RTX GPU, indicating a significant computational requirement.
- Template Design: The effectiveness of the prompt-tuning approach is heavily reliant on template design. While the paper outlines EMR priors for this, further research into automated or more robust template generation could be beneficial.
- Generalizability: The paper focuses on Chinese EMRs. Investigating how MKE-Coder's principles can be adapted or extended to other languages or EMR structures would be a valuable future direction.
Summary updated