老人與非老人急診住院預測與評分表發展之研究

Abstract

無論是流感、大流行特殊傳染病疾病傳播事件，如: 新冠疫情（COVID-19）的傳播，屢屢會使逐年攀升的醫療需求更為加重，若能初步在急診檢傷階段精確地識別不同優先順序的患者進行看診，有益幫助患者縮短待診時間。本研究與台北馬偕紀念醫院急診部合作，以電子病歷系統（Electronic Health Record，EHR）的資料為分析對象，資料範圍在2019年至2022上半年，總計372,820筆就診紀錄。並拆分上、下半年與非老人、老人，共14個資料集相比較各時間區段的變化。欄位包含診斷關鍵的非結構化資料——主訴（Chief Complaint）和結構化資料（如：年齡、檢傷級數、到院方式、血壓等），試圖透過統計描述、自然語言處理技術（Natural Language Processing，NLP）之BERT模型和集成學習演算法（Ensemble Learning）中的XGBoost，推測疫情前、中、後不同時期之影響與住院患者的主要變數因子。後續藉由過往文獻或臨床研究、演算法變數篩選，初步探索建立非疫情期間（Non-Pandemic Periods）與疫情期間（COVID-19 Pandemic Periods）的住院風險評估表，以供急診實務上使用。主要結果發現本研究新增的三項變數（腦傷程度、呼吸頻率、供氧狀態）可以提高兩種模型的預測能力（BERT的AUROC: 0.8643-0.9815 提升至 0.9075-0.9879 ；XGBoost的AUROC: 0.7847-0.8603 提升至 0.7862-0.8605），多數時候「非老人」的預測結果是好於「老人組」。最佳的住院預測結果在「2021下半年」，推估疫情中後期患者的就診行為較為雷同，因此模型學習到相似的模式。由於BERT展現對文本資料的強大分析優勢，進而將BERT輸出層的住院機率值整合至XGBoost當成其中輸入變數。實驗結果發現能夠大幅度提高XGBoost的預測能力（AUROC: 0.7881-0.8605 提升至 0.9122-0.9858）。 住院評分表建立的探索，同樣也增加「主訴預測住院機率值」來達到區別非住院組及住院組兩群的驗證效果。由此可知合作醫院的主訴語料品質非常好，亦結合穩定性良好的BERT語言模型，達成準確度高的預測性。Whether it is influenza or pandemic-specific infectious disease transmission events, such as the COVID-19 pandemic, they repeatedly exacerbate the rising demand for healthcare. If, at the initial stage of emergency triage, injuries can be accurately identified to prioritize patients for treatment, it would help shorten the waiting time for medical care. This study, in collaboration with the Department of Emergency Medicine at MacKay Memorial Hospital in Taipei, analyzed data from the Electronic Health Record (EHR) system, covering the first half of the period from 2019 to 2022, with a total of 372,820 records. The data was split into the first and second halves of the year and further divided into non-elderly and elderly groups, resulting in a total of 14 datasets to compare changes in each time zone. The fields contain unstructured data—Chief Complaint, which is key to diagnosis, and structured data (e.g., Age, Triage level, Mode of arrival, Blood pressure, etc.). The study attempted to support clinical diagnosis through statistical description, the BERT model of Natural Language Processing (NLP), and XGBoost of Ensemble Learning. Subsequently, by referencing past literature, clinical studies, and algorithmic variable screening, we preliminarily explored establishing a Hospitalization Scoring System for Non-Pandemic Periods and COVID-19 Pandemic Periods for practical use in emergency medicine.The main results found that the three new variables (Glasgow Coma Scale, Respiratory Rate, and Oxygen Saturation) in this study improved the predictive ability of both models. The AUROC for BERT increased from 0.8643-0.9815 to 0.9075-0.9879, and for XGBoost, it increased from 0.7847-0.8603 to 0.7862-0.8605. Most of the time, the "non-elderly" group had better predictive results than the "elderly" group. The best hospitalization prediction result was in the "second half of 2021", which is speculated to be because patient behavior during the middle and later stages of the epidemic was relatively similar, allowing the model to learn similar patterns. Since BERT demonstrates a strong advantage in analyzing text data, the hospitalization probability values from the BERT output layer were integrated into XGBoost as one of the input variables. The experimental results showed a significant improvement in the predictive power of XGBoost (AUROC: 0.7881-0.8605 increased to 0.9122-0.9858). In the establishment of the Hospitalization Scoring System, the"Chief Complaints hospitalization probability value" was also added to achieve the verification effect of distinguishing the non-hospitalization group and the hospitalization group. It can be seen from this that the quality of the Chief Complaints corpus of the cooperative hospital is excellent, and combined with the stable BERT language model, it can achieve high-accuracy prediction.