- 學位論文
基於深度學習的心臟驟停預警系統在不同長度和不規則時間序列上的內部及外部驗證
Internal and External Validation of a Deep Learning-based Early Warning System of Cardiac Arrest with Variable-Length and Irregularly-Measured Time-Series Data
本文將於2027/08/06開放下載。若您希望在開放下載時收到通知，可將文章加入收藏。
摘要
院內心跳驟停(IHCA)在急診室中是一個很重要的議題，跟患者的生命率息息相關。此外，在醫院的急診室中常會發生過度壅塞的問題，造成醫護人員人手不足。因此，如何設計出一個有效且可靠的預警系統就成了當今重要的議題。 然而，醫療資料的取得並不容易，通常不規則且擁有大量缺失值，這使如何處理資料並妥善運用於訓練模型變得十分困難。此外，大多數關於深度學習的研究使用時間序列模型(例如:GRU, LSTM)設計系統，但該類模型不適合應用於短序列資料，這也就意味者對於在急診室暫留時間短的患者，模型的表現可能會不如預期。 在這篇論文中，我們使用了台大醫院急診室的資料庫，並且同時利用該資料庫中的靜態特徵及動態特徵來訓練並驗證模型。靜態特徵包含了病患的背景資訊、檢傷級數等，動態特徵包含了患者的生命徵象，例如血壓、血氧及心率。對於動態特徵，我們會產生對應的缺失值遮罩，並且基於TCN設計一個可以同時考慮生命徵象及其缺失值的模型。此外，我們設計了一個混合模型來舒緩時間序列模型遇到短序列樣本時準確率下降的問題。實驗結果顯示出我們的系統在資料不平衡的資料集上獲得了優秀的表現。在內部驗證的實驗中，模型的AUPRC跟AUROC分別達到0.2150以及0.9831的表現。此外，我們也在FEMH-Death以及MIMIC-IV-ED的資料集上進行了外部驗證，分別獲得了AUPRC 0.1336跟AUROC 0.9734以及AUPRC 0.0533跟AUROC 0.8428的表現，證實了我們提出的模型是有效且可被應用於不同環境的。
並列摘要
In-hospital cardiac arrest (IHCA) in the emergency department (ED) is an important issue that significantly impacts patient survival rates. Additionally, overcrowding in the ED often leads to a shortage of medical staff. Consequently, developing an effective and reliable early warning system (EWS) has become a critical concern. Acquiring medical data remains a formidable obstacle often characterized by irregularities and a substantial amount of missing data, making processing and application to deep learning challenging. Additionally, most studies employ RNN-based models (such as GRU and LSTM) to design EWS. Nevertheless, these models are not well-suited for short sequence data. This implies that the model's performance may not meet expectations for those patients with brief stays in the ED. This study utilizes the database from the National Taiwan University Hospital's ED and employs static and dynamic features for training and validating the model. Static features encompass the patient's background and triage information, whereas dynamic features include the patient's vital signs, such as blood pressure, blood oxygen, and heart rate. We will generate the corresponding missing value mask for dynamic features and design a model based on TCN that can consider both vital signs and their missing values. Additionally, we design a hybrid model to address the issue of decreased accuracy when the RNN-based model encounters short sequence samples. Experimental results show that our system achieves excellent performance on imbalanced datasets. In internal validation, the model reached an AUPRC of 0.2150 and an AUROC of 0.9831. Additionally, we conducted external validation on the FEMH-Death and MIMIC-IV-ED datasets, obtaining an AUPRC of 0.1336 and an AUROC of 0.9734 for FEMH-Death, and an AUPRC of 0.0533 and an AUROC of 0.8428 for MIMIC-IV-ED. These results confirm the robustness of the proposed system for application in diverse environments.
參考文獻
Joon-myoung Kwon,, Youngnam Lee, Yeha Lee, Seungwoo Lee, and Jinsik Park. An algorithm based on deep learning for predicting in-hospital cardiac arrest. Journal of the American Heart Association, 7(13):e008678, 2018.
Adel Darraj, Ali Hudays, Ahmed Hazazi, Amal Hobani, and Alya Alghamdi. The association between emergency department overcrowding and delay in treatment: a systematic review. In Healthcare, volume 11, page 385. MDPI, 2023.
Gary B Smith, David R Prytherch, Paul Meredith, Paul E Schmidt, and Peter I Featherstone. The ability of the national early warning score (news) to discriminate patients at risk of early cardiac arrest, unanticipated intensive care unit admission, and death. Resuscitation, 84(4):465–470, 2013.
Christian P Subbe, Michael Kruger, Peter Rutherford, and L Gemmel. Validation of a modified early warning score in medical admissions. Qjm, 94(10):521–526, 2001.
Chu-Lin Tsai, Tsung-Chien Lu, Cheng-Chung Fang, Chih-Hung Wang, Jia-You Lin, Wen-Jone Chen, and Chien-Hua Huang. Development and validation of a novel triage tool for predicting cardiac arrest in the emergency department. Western Journal of Emergency Medicine, 23(2):258, 2022.