建置人工智慧ICD-10-CM編碼系統輔助疾病分類人員編碼

Using artificial intelligence ICD-10-CM coding system to assist medical coding professionals

為與國際接軌及正確地反映當今的醫療技術、處置，中央健康保險署自民國105年1月1日起，導入美國2014年版國際疾病與相關健康問題統計分類第十版臨床修訂及處置代碼系統（ICD-10-CM/PCS），作為各特約醫療院所門診、急診、住院診斷及處置之申報依據，已實施及持續擴大實施的住院診斷關聯群支付制度（Tw-DRGs）亦以此為基礎。ICD-10-CM/PCS比ICD-9-CM更複雜且困難，其診斷增加側性、細部位置、照護狀況；手術處置增加特材與手術途徑等，影響疾病分類人員需花費更多時間查找代碼。
本研究以某公立醫學中心由疾病分類人員編碼的ICD-10-CM資料及出院病歷摘要電子病歷為特徵資料，應用自然語言處理技術Bio-BERT建置機器學習模型，以學習各疾病分類人員之編碼模式，並探討不同模型於預測準確度上的表現差異。
為提高AI編碼使用率及預測精準度，本研究亦將AI自動化編碼網頁服務整合落地於院內疾病分類系統，以輔助疾病分類人員ICD-10-CM編碼參考，期能提升編碼品質及效率。

To align with international standards and accurately reflect contemporary medical technologies and procedures, the National Health Insurance Administration (NHIA) in Taiwan adopted the 2014 edition of the International Classification of Diseases, 10th Revision, Clinical Modification and Procedure Coding System (ICD-10-CM/PCS) on January 1, 2016. This coding system is now the basis for claims related to outpatient, emergency, and inpatient diagnoses and procedures across contracted healthcare institutions, and underpins the Taiwan Diagnosis-Related Groups payment system (Tw-DRGs). Compared to ICD-9-CM, ICD-10-CM/PCS is more complex, incorporating greater specificity such as laterality, anatomical details, and care circumstances for diagnoses, as well as surgical approaches, device usage, and materials for procedures. This complexity increases the workload for coders, who must spend significantly more time identifying appropriate codes.
This study utilizes ICD-10-CM codes assigned by professional coders at a public medical center, along with structured data extracted from electronic discharge summaries, as input features. By leveraging Bio-BERT, a domain-specific natural language processing model, we developed machine learning models to learn and replicate coders’coding patterns. We also evaluated the predictive performance of different models to identify optimal approaches for automated coding.
To improve the adoption and accuracy of AI-assisted coding, we implemented an AI-based automated coding web service integrated into the hospital’s disease classification system. This service serves as a decision-support tool for coders, aiming to enhance both the quality and efficiency of ICD-10-CM coding.