以異質網路圖學習病況事件表示法進行死亡風險預測

Abstract

近年來以機器自動學習數據的特徵表示法，已顯示有助於提升預測任務的準確率。本論文以電子病歷資料中相異類型的病況資料，依指定時間區間內病況事件同時發生的關聯，建立病況事件異質網路圖，並搭配不同的病況事件序列生成樣式，從取樣的事件序列中，學習儀器偵測數據特徵的病況事件表示法，用來從加護病房病患入病房後48小時的病況資料，以LSTM類神經網路架構進行死亡風險預測。本論文實驗比較使用同質特徵走訪路徑與異質特徵走訪路徑的擷取策略，所學習到的病況事件表示法對模型預測效果的差異。實驗在院內死亡預測及短期死亡預測的任務，初步顯示由異質特徵走訪路徑中學習的病況事件表示法，對兩個預測模型的預測效果皆有提昇。

In recent years, feature representation learning from data has been shown to be helpful for improving the accuracy of prediction tasks. In this thesis, the various types of attributes combined with the values in the electronic medical record, which implicitly describe patient’s condition, are named clinical events. We constructed a heterogeneous network of clinical events according to their occurring on the same patient within a specified time interval. Then event sequences are sampled by visiting different meta-paths for learning the representations of chart events. The learned representations of chart events are used to input to a framework of LSTM neural network for predicting mortality of ICU patients according to their first 48 hours of in-ICU EMR data. In the experiments, we compared the prediction effectiveness of the learned event representations by changing the time interval of constructing the heterogeneous network and applying homogeneous or heterogeneous meta-path visiting. The preliminary results of experiments show that the representations of chart events learned from the heterogeneous meta-path effectively improve the recall and AUROC on both the tasks of in-hospital mortality prediction and short-term mortality prediction.