機器學習輔助急診分流:以缺氧－年齡－休克指數為基礎之急診族群預後模型的建構與驗證

Abstract

急診分流須在極短時間內進行準確的風險分層，以最佳化醫療資源配置並改善病人預後，尤其在重症病人大量湧入時更顯重要。休克指數（shock index, SI）及年齡－休克指數（age-shock index, ASI）為臨床上常用之簡易血流動力學指標，廣泛應用於急診分流；然而，其預後預測能力受限於未納入氧合狀態，而氧合不良為不良臨床結局的重要決定因子。本論文旨在以缺氧－年齡－休克指數（hypoxia-age-shock index, HASI）為核心，建構、驗證並延伸一套機器學習輔助之急診分流預後評估架構，適用於不同急診照護族群。 本整合性研究包含三項於急診情境中進行之回溯性世代研究。第一項研究探討 SI 與 ASI 於老年 COVID-19 病患中的預後效用，評估其與加護病房（ICU）收治、氣管內插管及死亡率之關聯性。第二項研究則在前述基礎上，提出將血氧飽和度（SpO₂）納入 ASI 的新型綜合指標 HASI，並於急診分流時比較 HASI、SI 與 ASI 對 COVID-19 病患不良臨床結局之預測能力。第三項研究將 HASI 的應用延伸至另一高風險急診族群，接受初級經皮冠狀動脈介入治療之 ST 段上升型心肌梗塞（STEMI）成人病患，並進一步結合以急診分流變項訓練之機器學習模型，以提升預測效能。模型鑑別力以受試者操作特徵曲線下面積（area under the curve, AUC）評估，並透過 SHapley Additive exPlanations（SHAP）進行模型可解釋性分析。 於老年 COVID-19 病患中，SI 與 ASI 對 ICU 收治、氣管內插管及死亡率皆展現可接受之鑑別能力，其中 ASI 在死亡率預測上顯著優於 SI。將缺氧狀態納入後之 HASI，其預測效能優於 SI 與 ASI，特別是在 ICU 收治與氣管內插管之預測上表現更為突出，並同時提升死亡率預測之敏感度。於 STEMI 病患族群中，HASI 亦一致優於 SI 與 ASI 於院內死亡及緊急插管之預測，證實其預測效用不侷限於感染性疾病族群。機器學習模型，尤其為樹狀結構演算法，在死亡率預測上顯著提升整體表現，惟對插管事件之預測改善幅度有限。可解釋性分析顯示，SpO₂ 為死亡率預測中最具影響力之變項，突顯氧合狀態在早期風險分層中的關鍵角色。 本論文呈現急診分流預後評估工具由傳統血流動力學指標，逐步發展為納入氧合調整之綜合指數，並進一步結合機器學習模型的演進歷程。HASI 為一項簡便、具高度可解釋性且穩健之工具，能有效於急診早期辨識高風險病患，適用於多元急診族群。雖然機器學習方法可進一步提升死亡率預測能力，HASI 仍保有臨床實用性與透明度，具備納入常規急診分流流程之可行性，有助於即時臨床決策及精準化醫療資源配置。

Emergency department (ED) triage requires rapid, accurate risk stratification to optimize resource allocation and improve patient outcomes, particularly during surges of critically ill patients. The shock index (SI) and age-shock index (ASI) are simple hemodynamic tools widely used at triage; however, their prognostic performance is limited by the absence of oxygenation status, a key determinant of adverse outcomes. This dissertation aimed to develop, validate, and extend a machine learning–assisted triage framework centered on the hypoxia-age-shock index (HASI) for emergency care populations. This integrated research program consisted of three retrospective cohort studies conducted in emergency settings. The first study evaluated the prognostic utility of SI and ASI in geriatric patients with COVID-19, assessing their associations with intensive care unit (ICU) admission, endotracheal intubation, and mortality. Building upon these findings, the second study introduced HASI, a novel composite index incorporating oxygen saturation (SpO₂) into ASI, and compared its predictive performance with SI and ASI for adverse outcomes in COVID-19 patients at triage. The third study extended the application of HASI to a different high-risk emergency population, adult patients with ST-segment elevation myocardial infarction (STEMI) undergoing primary percutaneous coronary intervention, and further enhanced prediction using machine learning (ML) models trained on triage variables. Discriminatory performance was evaluated using the area under the receiver operating characteristic curve (AUC), and model interpretability was explored using SHapley Additive exPlanations (SHAP). In geriatric patients with COVID-19, both SI and ASI demonstrated acceptable discrimination for ICU admission, intubation, and mortality, with ASI significantly outperforming SI in mortality prediction. The incorporation of hypoxia into the index (HASI) resulted in superior predictive performance compared with both SI and ASI, particularly for ICU admission and endotracheal intubation, while also improving sensitivity for mortality prediction. When applied to patients with STEMI, HASI consistently outperformed SI and ASI in predicting in-hospital mortality and emergency intubation, confirming its generalizability beyond infectious disease contexts. ML models, especially tree-based algorithms, markedly enhanced mortality prediction; however, gains for intubation prediction were modest. Explainability analyses highlighted SpO₂ as the most influential predictor of mortality, underscoring the critical role of oxygenation in early risk stratification. This dissertation demonstrates a progressive evolution from traditional hemodynamic indices to an oxygenation-adjusted composite index and ultimately to machine learning–enhanced prognostic models for ED triage. HASI represents a simple, interpretable, and robust tool for early identification of high-risk patients across diverse emergency populations. While ML approaches further improve mortality prediction, HASI remains clinically practical and transparent, supporting its integration into routine emergency triage to facilitate timely decision-making and precision-guided resource allocation.