| 摘要: | 研究動機：全球疫情爆發使醫療需求激增，尤其急診部門面臨極大壓力。人力不足導致急診部門壅塞，患者等待時間長、就醫品質下降，甚至可能引發診斷錯誤和延誤治療，危及患者安全。全球各地都面臨急診部門供需不平衡的問題，如何從眾多病患中辨識高風險患者並且優先處理治療，成為急診部門面臨的重要課題。目前現行的檢傷制度對於辨別明顯病況嚴重的患者有一定效果，但對於非預期危急的高風險患者則較無法有效篩檢，容易低估治療需求。此外，病患在短時間內非計畫性返診也是全球面臨的議題，這不僅耗費醫療資源，還增加了治療成本和對病患的負擔。導致短時間內非計畫性返診的潛在因素包括病情未得到適當治療、治療不完全或不合適、治療後不當遵循以及病情惡化。因此，解決這些問題為緩解急診壅塞、提升醫療品質和保障病患安全的關鍵。研究方法：本研究使用臺北醫學大學雙和醫院急診科2017年5月至2019年12月的臨床數據進行模型開發，總共245,721筆就醫紀錄。研究對象為至急診就醫的病患，排除了行政需求就醫、資料不完整、兒科和外科病患。預測病患動向的資料集排除了到院前心跳停止和無記載動向的病患資料，資料筆數為167,058筆。預測病患返診的資料集則選取了72小時內第一次來診的就醫紀錄，並排除了到院前心跳停止、入院或非醫師許可出院的病人，資料筆數為128,848筆。研究中使用的預測因子包括基本人口統計學資料、檢傷生命徵象和臨床病史的文字敘述。為了處理臨床文本敘述，研究將中、英文混雜的臨床文字翻譯成英文，並將數值資料轉化為特定的文字描述。預測模型使用了多個機器學習模型以及在生物醫學領域廣泛使用的BlueBERT和BioClinical BERT模型，透過預訓練的語言模型能夠更精確的表示急診文本資料中的關鍵資訊，有助於預測病患動向和非預期性返診。實驗結果：本研究模型驗證方法為10折交叉驗證，以確保模型在實際應用時的準確性和穩定性。在急診資料中，病患的入院率為23.8%、病患返診率為2.2%。本研究建立的具臨床敘述感知能力的預訓練語言模型在預測病患動向之AUROC為0.9014，預測返診之AUROC為0.6475。結論：本研究旨在解決急診擁擠和醫療資源不均的問題，並建立一個能夠預測病患動向及短時間內非計畫性返診的臨床決策輔助模型。研究使用具臨床敘述感知能力的預訓練語言模型，分析臨床資料以預測病患是否會返診。研究結果顯示模型在預測病患動向方面表現良好。未來的研究可擴大資料集，整合其他資料來提升預測效果，並考慮時間的動態性，為急診在臨床決策方面提供更關鍵的資訊和洞察。關鍵字：急診部門、病患動向、非計畫性返診、自然語言處理、基於變換器的雙向編碼器、臨床決策支援
Background: The global outbreak of the pandemic has led to a significant increase in healthcare demands, particularly in the emergency department, which is under immense pressure. Insufficient manpower has resulted in congestion in emergency departments, leading to long patient waiting times, decreased quality of care, and potential risks of misdiagnosis and delayed treatment, jeopardizing patient safety. Emergency departments worldwide are facing the challenge of an imbalance between supply and demand. Therefore, identifying high-risk patients from a large pool of patients and prioritizing their treatment has become a critical issue for emergency departments. Furthermore, unscheduled return visits by patients within a short period of time are a global issue, leading to resource consumption, increased treatment costs, and burdens on patients. Potential factors contributing to unscheduled return visits include inadequate treatment of the initial condition, incomplete or inappropriate treatment, non-compliance with post-treatment instructions, and disease progression. Addressing these issues is crucial to alleviate emergency department congestion, improve healthcare quality, and ensuring patient safety. Methods: This study utilized clinical data from the Emergency Department at Taipei Medical University Shuang Ho Hospital from May 2017 to December 2019, with a total of 245,721 medical records. The study focused on patients seeking emergency care, excluding administrative visits, incomplete data, and pediatric and surgical patients. The dataset for predicting patient outcomes excluded patients with pre-arrival cardiac arrest and those without recorded outcomes, resulting in 167,058 records. The dataset for predicting patient readmissions selected the first visit records within 72 hours, excluding patients with pre-arrival cardiac arrest, admissions, or unauthorized discharges, resulting in 128,848 records. The predictive factors used in the study included demographic data, vital signs, and clinical history described in textual format. To handle clinical narratives, the study translated the mixed Chinese and English clinical text into English and transformed numerical data into specific textual descriptions. The predicting models employed multiple machine learning models as well as pre-trained language models, such as BlueBERT and BioClinical BERT, widely used in the biomedical domain. By leveraging pre-trained language models, the study aimed to capture crucial information from emergency department text data more accurately, facilitating the prediction. Results: This study utilized 10-fold cross-validation to evaluate the model, which can ensure the accuracy and stability of the model in practical applications. In the emergency department data, the admission rate of patients was 23.8%, and the return visit rate in 72 hours was 2.2%. The clinical narrative-aware pre-trained language model established in this study can achieve an AUROC of 0.9014 for predicting patient disposition and an AUROC of 0.6475 for predicting unscheduled return visits. Conclusion: This study aims to address the issues of overcrowding in emergency departments and uneven distribution of healthcare resources by developing a clinical decision support model that can predict patient outcomes and unplanned readmissions within a short period of time. The research utilizes a pre-trained language model with clinical narrative understanding to analyze clinical data and predict patient readmissions. The results demonstrate a good performance of the model in predicting patient outcomes. Future research can expand the dataset, integrate additional data to enhance prediction accuracy and consider the dynamic nature of time to provide more critical information and insights for clinical decision-making in emergency departments. Keywords: Emergency department, Disposition, Unscheduled return visit, Natural language processing, Bidirectional Encoder Representations from Transformers, Decision support |
