Disease classification and coding play a critical role in healthcare administration and health insurance claims. The internationally adopted ICD-10 system contains approximately 69,000 diagnostic codes and 72,000 procedural codes. Traditionally, coding has relied on specialists manually reviewing medical records to determine classifications—an approach that is both time-consuming and susceptible to subjective interpretation, potentially impacting the accuracy of medical reimbursements and research outcomes.

To address this "high knowledge intensity, low automation" challenge, the National Taiwan University Hospital (NTUH) Medical Informatics Innovation Team has taken the lead in developing a proprietary, locally trained generative Large Language Model (LLM). By incorporating natural language processing (NLP) and deep learning technologies, the team devised a specialized training strategy combining internal and external medical corpora to create an AI model capable of clinical interpretation. This approach overcomes the limitations of traditional NLP models, particularly in understanding long-form text, interpreting sequences of medical events, and performing professional logical reasoning based on clinical records.

AI Accuracy Breakthrough Sets a New Benchmark for Smart Healthcare

The NTUH team trained and deployed the language model entirely on local infrastructure. Experiments showed that in the task of primary diagnosis prediction, the system achieved an accuracy rate of 82.29%, a nearly 30% improvement (+29.7%) over the previous JointLAAT model. The F-score for full-code prediction reached as high as 86.67%. This achievement is attributed to several key innovations:

• Dual-task architecture: Separating the training of primary and secondary diagnoses enhances the model’s ability to focus on identifying the main diagnosis.
  
• Prioritized medical record segments: Based on practical experience, critical sections such as discharge diagnoses and surgical records are prioritized for model input.
  
• Medical history integration: The model incorporates patients’ previous diagnoses and the textual definitions of ICD codes to improve understanding of clinical chronology and causal relationships.
  
• Version-specific output: The model is restricted to ICD-10 2014+ codes, with readiness to support the upcoming 2023 revision, minimizing version-related errors.
  
• Standardized formatting and semantic mapping: In addition to outputting ICD codes, the model also provides corresponding textual definitions, enhancing semantic consistency and interpretability.

For real-world application, the user interface has been designed with both coders and clinicians in mind. The system offers a streamlined workflow and intuitive operation, and it has already been fully implemented across all NTUH branches, earning a user satisfaction rate of 95%. Beyond improving coding quality, the system directly contributes to more precise health insurance reimbursements, reliable data for research and statistical analysis, and greater accuracy in public health policymaking.

Integrating Technology and Clinical Practice to Establish a Global Smart Healthcare Model

Dr. Chuang, Chiou-Hwa, the project’s principal investigator, emphasized that under Taiwan's National Health Insurance system, ICD coding is closely tied to the allocation of medical resources. Coding quality has a direct impact on healthcare system efficiency and the utility of clinical information. The development of this large language model system represents not only a technological breakthrough but also a successful integration of cross-disciplinary expertise.

In preparation for the full implementation of the 2023 ICD-10 update in 2025, Dr. Chuang highlighted that the system's flexible architecture enables rapid adaptation. Updating code definitions and training data allows for a seamless transition to new coding rules. Looking ahead, the team plans to collaborate with the National Health Insurance Administration, major healthcare systems, and health insurance organizations to expand the application of this technology. The goal is to further enhance the quality of disease coding across institutions and extend its use into DRG analysis, automated auditing, and clinical decision support.

Resource: AI輔助疾病分類編碼 提升品質助健保給付精準落點