| 摘要: | Background: 當前全球醫療體系普遍面臨急診壅塞與醫療人力流失的雙重挑戰，這使得急診流程之效率與準確性成為醫療品質管理的關鍵。患者初次抵達醫院時，檢傷分類的準確性對醫療資源分配與臨床決策具決定性影響。台灣自民國99年起，導入以加拿大五級檢傷量表（Canadian Triage and Acuity Scale, CTAS）為基礎的台灣急診檢傷分類系統（Taiwan Triage and Acuity Scale, TTAS），將急迫性劃分為一至五級，數字越低代表越高的臨床緊急程度。隨著大數據分析科技的快速發展，在先前的幾個研究中，導入急診的臨床文字敘述能夠為臨床決策提供更豐富的資料。我們收集與分析急診自由文本資料，旨在開發更精準的急診處置預測模型。
Purpose: 本研究旨在建立一套預測模型，以輔助急診檢傷分類與醫療決策。我們整合多種患者數據，包含入院時間、方式、TTAS檢傷指數以及各項生命徵象(如血壓、年齡、體溫、意識狀態等) ，並進一步運用自然語言處理（NLP）技術把主訴、護理紀錄和過去病史自由文本形式的非結構化資料轉換成可供於機器學習使用的結構化數據。透過這些數據，我們預測病患接受醫療處置後是否需住院，以期提升急診流程之效率與預測準確性。
Methods: 本研究資料來自台灣聯新國際醫院急診2012至2022年之電子健康紀錄，共781,342筆資料。自由文本部分經由翻譯、NIH MetaMap分析處理，完成文本命名實體識別NER，並映射至耶魯大學發表之《急診分診關鍵概念》文獻中所列之關鍵概念(key concepts)，搭配本地資料進行篩選，再以NegEX演算法進行否定語意偵測。本研究假設經由結合患者生命資訊與格式化的症狀概念，可以用於訓練機器學習模型以提升對患者結局的預測能力。
Result: 具有連續性質的身體特徵（如年齡、血壓和體溫等）被格式化為浮點數值；而非連續性質的資訊（如到院方式、星期幾及時間段）則以 One-Hot Encoding 處理。從自由文本萃取的概念則轉換為三元狀態（正向提及為+1、否定提及為-1、未提及為0）。僅使用15個患者初始結構化資料訓練Random Forest，並經Grid Search做參數最佳化後，模型於測試集之 AUC為0.8058(95% CI: 0.8026-0.8085)。以此為基礎再納入598筆來自耶魯文獻之對照概念後，AUC提升至0.8552(95% CI: 0.8526-0.8577)；最後加上篩選後602筆本地專屬概念特徵後，模型AUC再提升至0.8829(95% CI: 0.8804-0.8852)。進一步分析各文本貢獻，以護理紀錄為主的模型表現最佳（AUC = 0.8570），過去病史則表現相對較低（AUC = 0.8133）。
Conclusion: 本研究證實，整合自然語言處理技術所提取之語意概念特徵，無論是對照國際醫學文獻所萃取的關鍵概念，或是本地化篩選後的語意特徵，皆能有效提升住院預測模型之效能。這些結果不僅強化語意特徵在急診預測任務中的應用價值，也顯示 NLP 技術與標準化醫療術語結合應用的可行性。此外，本研究提出結合否定偵測與三元編碼策略，使模型更細緻反映臨床語境的語意極性，為急診檢傷分類與醫療決策輔助提供創新方向。未來可延伸至跨院模型應用、時間序列資訊建模及視覺化臨床解釋工具開發，進一步強化模型的臨床實用性與推廣價值。
;Background:
Emergency departments (EDs) worldwide are increasingly challenged by overcrowding and workforce shortages, making the efficiency and accuracy of emergency triage processes critical for healthcare quality management. The accuracy of triage decisions upon a patient′s arrival directly affects medical resource allocation and clinical decision-making. Since 2010, Taiwan has adopted the Taiwan Triage and Acuity Scale (TTAS), a system based on the Canadian Triage and Acuity Scale (CTAS), which classifies patients into five levels of urgency, with lower numbers indicating higher clinical urgency. Several previous studies have shown that clinical free text descriptions of emergency department visits can provide richer data for clinical decision making. We collected and analyzed emergency free text data to develop a more accurate prediction model for emergency department triage.
Purpose:
This study aims to develop a predictive model to support emergency triage and medical decision-making. We integrated various patient data, including arrival time, mode of arrival, TTAS, and vital signs (e.g., blood pressure, age, body temperature, and consciousness level). Additionally, we applied natural language processing (NLP) techniques to convert unstructured free-text data—such as chief complaints, nursing records, and past medical history—into structured features suitable for machine learning. Using these combined data, we sought to predict whether a patient would require hospital admission after initial emergency care, thereby improving triage efficiency and prediction accuracy.
Methods:
We collected a total of 781,342 emergency department records from 2012 to 2022 at Landseed International Hospital in Taiwan. Free-text nursing records in Chinese were translated into English and processed using NIH MetaMap for named entity recognition (NER). The extracted entities were then mapped to key concepts listed in Yale University′s Emergency Triage Key Concepts literature. Additional concepts not included in Yale’s Emergency Triage Key Concepts were filtered from local data and categorized separately for Landseed’s data. Negation detection was performed using NegEx. We hypothesized that combining vital signs with standardized symptom concepts would enhance predictive modeling of patient outcomes.
Results:
Out of 1215 extracted features, 9 were continuous physiological features (e.g., age, blood pressure, heart rate), formatted as floating values. The remaining 1206 were categorical data (N=6, e.g., arrival mode, day of the week, time period) and extracted key concepts (N=1200, e.g., chest pain, breathless, fever), which were encoded using one-hot encoding. Concepts extracted from free text were represented in a three-state format: positive mention (+1), negated mention (−1), or unmentioned (0). Using only a basic set of 15 structured features (including TTAS), a Random Forest model optimized via Grid Search, the model achieved an AUC of 0.8058 (95% CI: 0.8026–0.8085) on the test set. Incorporating additional 598 features derived from Yale’s Emergency Triage Key Concepts increased the AUC to 0.8552 (95% CI: 0.8526–0.8577). Finally, adding 602 refined local concepts further improved the AUC to 0.8829 (95% CI: 0.8804–0.8852). When analyzing the contribution of individual text sources in addition to the basic 15 structured features (including TTAS), models incorporating nursing records achieved the highest performance (AUC = 0.8570), whereas those based on chief complaints and past medical history performed slightly lower (AUC = 0.8529 and 0.8133 respectively).
Conclusion:
This study demonstrates that integrating semantic concepts extracted from free-text records through NLP significantly enhances the performance of hospital admission prediction models. Both internationally recognized and locally curated concepts contributed to improved accuracy in our studied models. Moreover, our implementation of negation detection in a three-level concept representation enabled the model to better capture the sentiment polarity within clinical narratives. These findings support the feasibility and value of combining NLP with standardized medical terminologies in emergency care. Future directions include cross-institutional model validation, temporal information modeling, and the development of visual explanation tools to further enhance clinical applicability and adoption. |
