透過改良式HAN模型訓練於Tw-DRGs分類下ICD-10-CM人工智慧輔助編碼之可行性

Feasibility of ICD-10-CM AI-Assisted Coding within the Tw-DRGs Classification Unsing an Enhanced HAN Model

ICD-10-CM/PCS編碼精確度對於產出正確Tw-DRGs十分重要。編碼錯誤造成醫院營收虧損，在美國因編錯臨床代碼和後續更正所產生的成本估計為每年250億美元。本醫院與某科技大學合作研發專屬的ICD-10-CM人工智慧編碼模組。藉由本研究主要探討的問題有：驗證AI編碼模型編碼後的Tw-DRGs結果，證明人工智慧編碼模組與人工編碼是否一致。本研究以疾分人員角度驗證人工智慧編碼，研究期間是從2023年2月開始；因為本研究中人工智慧模組需先學習疾分人員ICD-10-CM編碼且病歷需書寫完整，所以研究對象為排除非Tw-DRGs、有處置（ICD-10-PCS）Tw-DRGs出院個案及書寫不完整病歷，透過「病歷回溯」方式進行資料收集。某科大的人工智慧模組是以階層式注意力網路（Hierarchical Attention Network，簡稱HAN）建模，HAN的特性在於模型聚焦於對預測結果最關鍵的部分，捕捉詞與句之間的關聯性，並對處理長文本有較佳的效果。以2及3月份共1739筆出院資料統計；描述性統計方面以有主要診斷但次要診斷錯誤的筆數最多（2月為61.8%；3月為48.1%）；推論性統計Kappa統計結果以主要疾病類別（Major Diagnostic Category，簡稱MDC）分析結果，傳染疾病寄生蟲病及呼吸系統之疾病兩大系統人工智慧編碼與疾分人員為幾乎一致（Kappa值高於0.8），其他高度一致有耳鼻喉及口腔之疾病與疾患（MDC 3）、循環系統之疾病與疾患（MDC 5）等共7項MDC。總結而言，本研究驗證了人工智慧輔助編碼模組的可行性。

Accurate ICD-10-CM/PCS coding is essential for generating correct Tw-DRGs, as coding errors may lead to substantial financial losses for healthcare institutions. In the United States, the annual cost associated with incorrect clinical coding and subsequent revisions has been estimated at approximately USD 25 billion. In response to coding-related challenges, our hospital has collaborated with a technology university to jointly develop a proprietary ICD-10-CM artificial intelligence (AI)-assisted coding module. This study therefore focuses on the following core questions: (1) evaluating the Tw-DRGs results produced by the AI coding model, and (2) determining whether the outputs of the AI-assisted coding module are consistent with those generated through manual coding.
This study evaluated the performance of AI-assisted coding from the perspective of disease classification personnel. The study period began in February 2023. Because the AI module in this study was trained using the ICD-10-CM codes assigned by disease classification experts and required complete medical records for reliable learning, the study population excluded non- Tw-DRGs cases, Tw-DRGs cases involving procedures (ICD-10-PCS), and medical records with incomplete documentation. Data were collected through a retrospective medical record review apporach. The AI-assisted coding module developed by the collaborating University of Science and Technology was constructed using a Hierarchical Attention Network (HAN). HAD offers the advantage of focusing on the most informative components of the input text for prediction, capturing relationships both within and across words and sentences. This architecture is particularly effective for processing long text and complex clinical narratives.
A total of 1,739 discharge records from February and March were included in the analysis. Descriptive statistics indicated that the most frequent type of discrepancy involved cases with a correct primary diagnosis but incorrect secondary diagnoses (61.8% in February and 48.1% in March). For inferential statistics, Kappa analysis was performed based on the Major Diagnostic Category (MDC). The results demonstrated that AI-assisted coding achieved near perfect agreement with disease classification personnel for infectious and parasitic diseases and for respiratory system diseases, both with Kappa values exceeding 0.8. High levels of agreement were also obserbed in seven other categories, including diseases and disorders of the ear, nose, mouth, and throat (MDC 3), and diseases and disorders of the circulatory system (MDC 5). In summary, this study demonstrates the feasibility of the AI–assisted coding module.