MKE-Coder: Multi-Axial Knowledge with Evidence Verification in ICD Coding for Chinese EMRs

This paper presents MKE-Coder, a novel framework that improves automatic ICD coding for Chinese electronic medical records by integrating multi-axial disease knowledge with a clinical evidence verification module to address challenges in information extraction and code validity.

The Gist

Imagine you are a librarian trying to sort thousands of messy, handwritten notes into a giant, organized library. But there's a catch: the notes are written in a very specific, shorthand style that only doctors understand, and the library's filing system (called ICD codes) is incredibly strict. If you put a book in the wrong drawer, the whole system gets confused.

This is the daily struggle of medical coders working with Chinese Electronic Medical Records (EMRs). The notes are often short, cryptic, and packed with hidden details, making it hard to find the right "filing code" for a patient's illness.

Enter MKE-Coder, a new AI assistant designed to solve this puzzle. Here is how it works, using a simple analogy:

1. The "Four-Point Compass" (Multi-Axial Knowledge)

Think of a patient's diagnosis not just as a single word, but as a location on a map. To find the right spot, you need to know four things: Where (the body part), What (the disease), Why (the cause), and How bad (the severity).

Old AI methods tried to guess the code by just looking at the whole picture at once, often getting lost. MKE-Coder is different. It acts like a detective with a four-point compass. It breaks every diagnosis down into these four specific "axes" (directions) first. This ensures it doesn't just guess the disease; it understands the context of the disease.

2. The "Evidence Detective" (Evidence Verification)

Once the AI has a list of possible codes (candidates), it doesn't just pick the first one that looks good. It goes on a scavenger hunt through the patient's entire medical record.

Imagine the AI is a lawyer building a case. It pulls out specific sentences from the doctor's notes that act as evidence. It asks: "Does this note actually prove that the patient has this specific condition?" If the note is vague or doesn't support the code, the AI throws that code out. It only keeps the codes that have "witnesses" (clinical evidence) backing them up.

3. The "Final Exam" (Masked Language Modeling)

After gathering the evidence, the AI takes a final test. It uses a strategy called "masked language modeling," which is like a fill-in-the-blank game.

The AI takes the candidate code and the evidence it found, then asks itself: "If I hide the code and look at the evidence, does the code fit perfectly into the blank space?" If the answer is "yes," the code is valid. If the answer is "no," it means the code doesn't match the story the medical record is telling. This step ensures the final decision is rock-solid.

The Result

When the researchers tested this new AI on real hospital data from China, it didn't just do well; it shined.

• For the Computers: It was much more accurate than previous methods at finding the right codes.
• For the Humans: In real-world simulations, it acted like a super-powered co-pilot for human coders. It helped them finish their work faster and with fewer mistakes, turning a tedious, error-prone job into a smooth, efficient process.

In short: MKE-Coder is like a super-smart, detail-oriented librarian who doesn't just guess where a book goes. It checks the book's cover, reads the summary, verifies the author's notes, and then places it in the exact right spot with 100% confidence.

Technical Summary

Based on the abstract provided, here is a detailed technical summary of the paper "MKE-Coder: Multi-Axial Knowledge with Evidence Verification in ICD Coding for Chinese EMRs".

1. Problem Statement

The paper addresses the challenge of automatic International Classification of Diseases (ICD) coding specifically within the context of Chinese Electronic Medical Records (EMRs). While automatic coding has achieved success in English-language contexts, it faces significant hurdles when applied to Chinese EMRs due to two primary factors:

• Linguistic and Structural Complexity: Chinese EMRs often feature a concise writing style and specific internal structures that make extracting disease-related information difficult for standard models.
• Limitations of Existing Methods: Previous approaches fail to effectively leverage disease-based multi-axial knowledge (the multidimensional nature of medical diagnoses) and lack a robust mechanism to associate candidate codes with their corresponding clinical evidence within the text.

2. Methodology: MKE-Coder Framework

The authors propose a novel framework named MKE-Coder (Multi-axial Knowledge with Evidence verification), which operates through a three-stage pipeline:

• Stage 1: Candidate Generation and Multi-Axial Categorization
  The system first identifies candidate ICD codes for a given diagnosis. Unlike flat classification, it categorizes the knowledge associated with each candidate code into four distinct coding axes. This step ensures that the semantic richness of the diagnosis is captured across multiple dimensions relevant to ICD coding standards.

• Stage 2: Evidence Retrieval and Filtering
  To ground the coding decisions in the actual patient record, the framework retrieves corresponding clinical evidence from the comprehensive content of the EMRs. A dedicated scoring model is employed to filter this retrieved information, retaining only the "credible evidence" that strongly supports the diagnosis. This step mitigates noise and focuses the model on high-relevance text segments.

• Stage 3: Evidence-Based Verification (Inference Module)
  The core innovation lies in the final inference module, which utilizes a Masked Language Modeling (MLM) strategy. This module performs a rigorous verification process:

  • It checks whether all axis knowledge associated with a candidate code is explicitly supported by the filtered clinical evidence.
  • If the evidence supports the multi-axial requirements, the code is validated; otherwise, it is rejected or adjusted.
  • The system provides final coding recommendations based on this verified alignment between multi-axial knowledge and textual evidence.

3. Key Contributions

• Novel Framework for Chinese EMRs: Introduces MKE-Coder, specifically designed to overcome the linguistic and structural barriers of Chinese medical text.
• Multi-Axial Knowledge Integration: Moves beyond simple text-to-code mapping by explicitly modeling the four coding axes inherent in ICD classification, ensuring a more granular and accurate representation of diseases.
• Evidence Verification Mechanism: Proposes a unique MLM-based inference module that enforces a "proof-based" coding logic, requiring that every aspect of a candidate code be backed by specific evidence in the EMR.
• Scalable Evaluation: The method is tested on a large-scale dataset collected from various hospitals, ensuring the findings are robust and generalizable across different clinical settings.

4. Results

• Quantitative Performance: Experimental results on the large-scale Chinese EMR dataset demonstrate that MKE-Coder exhibits significant superiority over existing methods in the task of automatic ICD coding.
• Practical Utility: In simulated real-world coding scenarios, the framework was shown to significantly aid human coders. Specifically, it improved both coding accuracy (reducing errors) and coding speed (increasing efficiency), highlighting its potential for immediate clinical deployment.

5. Significance

This work represents a critical advancement in medical NLP for non-English languages. By bridging the gap between complex multi-axial medical knowledge and the specific constraints of Chinese clinical documentation, MKE-Coder offers a solution that is not only technically superior but also practically viable. Its ability to verify codes against clinical evidence reduces the risk of misclassification, which is crucial for billing, epidemiology, and healthcare quality management. The framework sets a new standard for how AI systems can integrate structured medical knowledge with unstructured clinical text in resource-constrained or linguistically complex environments.