深度隨機時間序列補值法於電子醫療病例的應用

近年，深度學習被應用在醫療領域，帶來許多進步。其中一項應用為電子病歷的研究。深度神經網路能夠從病歷解析出疫病的資訊，並擁有提供新穎且與眾不同見解的潛力。然而，電子病歷的不完整性阻礙深度模型的發展，限制其應用上的預測表現。電子病歷中檢驗的採檢頻率不同，各項目需要採檢的情況也不同，每個病例中不可能記錄所有項目。因此，在整合病例資料時，那些未檢測的項目就會成為大量的缺失值，造成電子病歷的不完整。該如何取代缺失的數值，正是影響深度模型能否精準預測疾病的關鍵。本研究提出深度隨機時間序列補值法 (Deep STI) 解決電子病歷缺值帶來的挑戰。其核心概念為運用深度補值網路取代缺失值，同時利用補值的電子病歷預測疾病。藉由融合補值網路以及預測網路，我們提出的模型能被應用於變動長度的電子病歷序列，產生其中的缺失值，且同時產生疾病的預測。我們以預測肝病病人一年後是否會罹患肝細胞癌 (HCC) 來評估預測模型的表現，並在本研究中展示我們的模型確實能夠有效增進預測表現，達到 40.74%平均 AUPRC 以及 89.46%平均 AUROC。相較於邏輯斯回歸僅有 22.68%以及平均 AUPRC 83.39%平均 AUROC，我們的預測模型更能顯其優勢。除此之外，我們還發現模型的補值網路能夠增進模型的穩定度，降低隨機性對模型的干擾。我們相信本研究將會成為於深度時序列補值應用於電子病歷研究中重要的一步。

關鍵字

深度學習；電子病歷；補值；癌症預測；肝細胞癌

並列摘要

In recent years, deep learning has brought advances to various medical applications. One of them is exploring electronic health records (EHRs) via deep learning. Deep neural networks (DNNs) possess the ability to extract abstract information from patients’ medical history and the potential to provide new and intelligent insights. However, the lack of complete EHRs hinders model development and limits model performance. Features in EHRs are observed in different frequencies or under different conditions; not all features are recorded in every EHR. After feature alignment, unobserved features in every EHR turns into numerous missing values. Determining a good representation of missing value is one of the keys to improving prediction performance. This study proposes the Deep Stochastic Time-series Imputation (Deep STI) model to address the challenge. The central concept is to infer missing values from observed values by an imputation network and simultaneously predict target disease according to imputed data by a prediction network. By integrating both the imputation network and prediction network into an end-to-end architecture, our model can simultaneously generate missing values in dynamic length EHR sequence and predict target probability. We evaluated our model via the prediction of real-world hepatocellular carcinoma (HCC) patients. Numerical experiments showed that Deep STI could improve model performance in predicting HCC in one year. Our model yielded the top mean AUROC of 89.46% and mean AUPRC of 40.74%. Our model significantly outperforms logistic regression, with a mean AUROC of 83.39% and a mean AUPRC of 22.68%. Besides, our imputation mechanism can efficiently improve model stability and reduce variance caused by randomness. We believe that this study can be an important piece of research iniv developing deep learning models for time-series EHR imputations.