基於監督式深度學習之 ICD-10 自動編碼與訓練系統

Abstract

背景: 目前，醫療系統及保險申報已廣泛使用ICD 編碼作為給付依據，然而ICD的分類作業仍主要依靠人力閱讀大量的文字資料作為分類的依據，耗時且耗力。自2014 年台灣醫療健保申報的依據由ICD-9 改成ICD-10 後，ICD 的分類又變得更細節、容易混淆，即使是專業的疾病分類師都需要平均至少20 分鐘才能完成一個案例的編碼。 目標: 本篇研究的目標即是建構一個能夠自動根據醫療診斷資料進行ICD-10 編碼的深度學習模型，以利降低耗費在疾病編碼的時間及人力成本。 方法: 在此篇研究中，我們使用台大醫院的診斷資料並應用自然語言處裡的技術(包含Glove, Word2Vec, ELMo, BERT, SHA_RNN)於深度學習網路中以實現ICD-10 的自動編碼。此外，我們亦導入注意力機制於模型中來視覺化決定編碼的重點文字依據，以提供ICD-10 的新手使用者編碼訓練的服務。 結果: 在各個實驗結果中，使用BERT 詞嵌入和Gated recurrent unit (GRU) 的分類模型在ICD-10-CM 與ICD-10-PCS 的編碼上達到最好的結果 (F1-Score 0.715 與0.615)。訓練完的模型亦導入ICD-10 的網頁服務中，以提供所有的ICD-10 使用者自動編碼及訓練的服務。 結論: 目前，這些模型及相關的網頁服務以提供所有ICD-10 使用者使用，將編碼時間從最多40 分鐘降低至約莫2 分鐘即可得到答案，大幅降低ICD-10 使用者編碼的時間。

Background: Nowadays, ICD code is widely used as the reference in medical system and billing purposes. However, classifying diseases into ICD codes still mainly relies on humans reading a large amount of written materials as the basis for coding. Coding is both laborious and time consuming. Since the conversion of ICD-9 to ICD-10, the coding task became much more confusing, even a disease coder with professional abilities takes about 20 minutes per case on average. Objective: This thesis aims at constructing a deep learning model for ICD-10 coding, where the model is to automatically determine the corresponding diagnostic and procedure codes based solely on free-text medical notes to reduce human effort. Methods: We used NTUH diagnosis records as the resources and applied NLP techniques, including Glove, Word2Vec, ELMo, BERT, and SHA-RNN, on the DNN architecture to implement ICD-10 auto-coding. Besides, we introduce the attention mechanism into the model to extract the key points from the diagnosis and visualize the coding reference for training freshman in ICD-10. Results: In experiments on the NTUH dataset, our predicting result could achieve F1-score of 0.715 and 0.615 on ICD-10-CM and ICD-10-PCS code with BERT embedding approach on GRU classification model. The well-trained models are applied on the ICD-10 web service for coding and training to all ICD-10 users. Conclusions: The proposed model and web service significantly reduces manpower in coding time; coding time is reduced from 20 mins ~ 40 mins to less than 2 mins.