以機器學習模型偵測一般病房住院病人惡化

Abstract

住院期間的心跳驟停（In-hospital cardiac arrest, IHCA）是嚴重的事件，常伴隨高死亡率，各大研究亦強調了早期識別和早期介入對於改善患者預後的重要性。部份的心跳驟停是突然地發生，沒有伴隨明顯徵兆，因此開發自動化的預測模型以識別高風險患者並及時進行介入是非常重要的。本研究引入了兩個創新的預測模型：『時間序列早期預警分數（Time-Series Early Warning Score, TEWS）』和『可解釋的時間序列早期預警分數（Explainable Time-Series Early Warning Score, TEWS-X）』。這兩個模型只使用常規量測的生命徵象資料來提供較為準確且可解釋的IHCA預測，使醫療提供者能夠採取主動措施，提高患者安全性。 TEWS模型通過結合多個時間窗口的特徵，再加上類神經網路對於特徵趨勢和模式的處理能力，實現了更高的預測準確性。此外，TEWS-X模型通過採用基於決策樹的機器學習方法和SHAP值，對醫療照顧者解釋其預測結果，使醫療照顧者可依此結果作出臨床決策。這些模型可以無縫地集成到現有的照護流程中，無需中斷工作流程，進而提升病人安全並優化資源分配。

In-hospital cardiac arrest (IHCA) is a critical event associated with high mortality rates. Early identification and intervention are crucial for improving patient outcomes. This study introduces two innovative predictive models: the Time-Series Early Warning Score (TEWS) and the Explainable Time-Series Early Warning Score (TEWS-X), designed to leverage vital signs data and provide accurate and explainable predictions of IHCA. The TEWS model utilizes vital signs data from six time windows (48 hours) to predict IHCA occurrences and performs superior IHCA prediction performance compared to alternative classification algorithms. Incorporating features from multiple time windows significantly improves prediction accuracy, with an area under the receiver operating characteristic curve (AUROC) of 0.808, surpassing the performance of MEWS (AUROC of MEWS: 0.649). The TEWS-X model incorporates a tree-based machine learning approach and SHAP values to enhance model explainability, enabling insights into feature importance and supporting transparent decision-making, facilitating an understanding of the critical factors influencing IHCA risk. These models can seamlessly integrate into existing care processes, improving patient safety without disrupting workflow. The TEWS and TEWS-X models represent significant advancements in IHCA prediction and explainability. By leveraging vital signs data and incorporating explainable modeling techniques, these models empower healthcare providers to identify patients at risk of IHCA and intervene promptly and proactively. Further research is needed to validate the models in diverse healthcare settings and explore additional data sources for enhanced predictive capabilities. Implementing the TEWS and TEWS-X models can improve patient outcomes and optimize resource allocation in the management of IHCA.