ICD-10編碼系統及多標籤分類問題

Abstract

世界衛生組織為了整合全世界疾病、損傷以及死亡的分類，訂定了世界通用的國際疾病分類標準International Classification of Disease (ICD)，使病人在醫院或診所產生的醫療相關資訊，如疾病診斷、手術方法等文字敘述，轉換成電腦可處理的字元，而能做統計分析及其他有價值的應用。 本篇研究的目的即是觀察ICD-10編碼系統在醫院使用情況，以及持續改善優化模型系統，以利減少在疾病分類的時間成本。此外醫療資料是非常隱私的，如果想要有更通用的模型，以傳統的方法需要搜集每間醫院的資料混合做訓練，要將這些資料聚集並不容易。但是聯邦學習能夠在資料不共享的情況下，利用傳遞模型權重的方法，將這些資料一起訓練。 在實驗結果中，使用聯邦學習在ICD-10-CM的編碼上，聯邦學習的模型在台大及亞東兩間醫院測試資料集的結果 (F1-measure 0.7465與0.6813)。而台大及亞東在本地端使用自己資料訓練的模型測試彼此的資料集的結果 (F1-measure 0.5583與0.5116);而在本地端使用自己資料訓練的模型測試自己醫院的資料集的結果 (F1-measure 0.7710與0.7412)。此外在台大、亞東及北榮三間醫院上，結果亦顯示使用聯邦學習的結果在跨院的表現上比本地端模型要好。

In order to integrate the classification of diseases, injuries and deaths around the world, the World Health Organization has established the International Classification of Disease (ICD). The medical-related information generated by patients in hospitals or clinics, such as disease diagnosis, surgical methods and other text descriptions, can be converted into specific characters that can be processed by computers and ICD can be used for statistical analysis and other valuable applications. The thesis aims at observing the real world usage of the ICD-10 coding system and continuously improving the coding system to reduce the time cost of disease classification. In addition, medical data are very private. If we want to have a general model, traditional methods mix the data from different hospitals and then do the training. However, it is not easy to aggregate these data because of privacy issue. But federated learning can train these data together without sharing data with each other. In the experimental results, using federated learning (FL) technique to predict ICD-10-CM can achieve F1-measure 0.7465 and 0.6813 on National Taiwan University Hospital and Far Eastern Memorial Hospital dataset respectively, the results of local training by testing their own data can achieve F1-measure 0.7710 and 0.7412, respectively. When evaluating each other’s dataset, local training model can achieve F1-measure 0.5583 and 0.5116, respectively. In addition, we also conduct the FL experiment of three hospitals including NTUH, FEMH, and VGHTPE. The results of FL model are better than local training models when testing the dataset of other hospitals.