運用Rule Base優化ICD-10自動化編碼系統之成效

The Effectiveness of Optimizing the Rule Based Artificial Intelligence ICD-10-CM Automatic Coding System

目的：本院於2020年運用疾病分類人工智慧(AI)編碼系統，2021年落地導入初期發現ICD-10-CM/PCS推薦總代碼數不多，導致運用AI編碼使用率及ICD-10-CM召回率偏低。為了提升使用率及召回率，以Rule Base人工智慧學習模式來優化ICD10自動化編碼系統，以提升預測代碼的準確率及時效性。
方法：運用自然語言處理技術Bio-BERT，讓人工智慧學習每位疾病分類師編碼模式，且輔以實務中ICD-10-CM診斷組合碼或外傷碼、院內計價碼對應ICD-10-PCS等相關的疾病分類編碼規則，經由人工智慧(AI)編碼系統處理分析，將最有可能的ICD-10-CM/PCS代碼呈現於個案編碼輸入畫面以利疾病分類師選取適當代碼。
結果：統計2022年4月至9月AI模型推薦代碼加上Rule Base模型推薦代碼的整合模型效能衡量指標，平均精確率(Precision)為32%、平均召回率(Recall)為74%、平均F1分數(F1 score)為48%。另2022年，疾病分類師人工智慧(AI)編碼使用率已提升至90%以上ICD-10-CM召回率從47.88％逐漸提升至60%以上。建立Rule Base採混合訓練模式，平均每個編碼案件節省47.4秒，以每月平均使用編碼件數5,000筆計算，每月節省65.8小時等於8.2個工作日。
結論：對於疾病分類師而言，優化人工智慧(AI)編碼系統除了推薦出更精確的代碼外，也可以避免因疏忽而漏編碼或編錯代碼的情況發生，有效提升編碼正確性及時效性。

Purpose: In 2020, our hospital implemented an Artificial Intelligence (AI) ICD-10-CM/ PCS coding system. However, during the initial implementation in 2021, it was found that the recommended total number of codes in ICD-10-CM/PCS was insufficient, leading to low utilization and recall rates of the AI coding system. To improve both utilization and recall rates, a Rule-Based AI learning model was adopted to optimize the ICD-10 automated coding system. This was to enhance the accuracy and timeliness of code predictions.
Methods: To improve the AI coding system, the Bio-BERT, Natural Language Processing technique (NLP), is employed to train the AI model for individual coding specialist. In addition, practical coding rules are incorporated, including the ICD-10-CM diagnosis combination codes or injury, poisoning and certain other external cause codes, as well as mappings to ICD-10-PCS. Through analysis and processing by the AI coding system, the most probable ICD-10-CM/PCS codes are presented on the coding entry interface, facilitating coding specialists in selecting the appropriate codes.
Results: From April to September 2022, the performance indicators for the integrated model combining the AI model’s recommended codes and rule base model's recommended codes are as follows: Precision rate is 32% on average, Recall rate is 74% on average, and F1 score is 48% on average. In 2022, the utilization rate of the AI coding system by coding specialists has increased to over 90%. The recall rate of ICD-10-CM codes has gradually improved from 47.88% to over 60%. By establishing a Rule-Based hybrid training mode and conducting indicator analysis, an average time saving of 47.4 seconds per coding case was achieved. Based on an average monthly coding volume of 5,000 cases, this equates to a time saving of 65.8 hours per month, equivalent to 8.2 working days.
Conclusion: For coding specialists, optimizing the AI coding system not only recommends more accurate codes but also helps to prevent the occurrence of missed or incorrect coding due to oversight. This effectively enhances coding accuracy and timeliness.