| 摘要: | 過去研究發現，在急診的病人診斷錯誤發生率可高達15-55%。然而診斷錯誤是可以藉由改善方式避免，如醫護人員再教育，使用提醒清單等。其中系統性的解決方法如建立臨床診斷輔助系統 (Clinical decision support system, CDSS)提供看診時的提醒清單，較能降低因個人因素導致的診斷錯誤或延遲，進而提升病人照顧品質。藉由機器學習方法建立急診病患相關的預測模型，提供臨床醫師病患可能的預後或動向，早已不是新鮮事。然而得知病人的預後，雖然可達到警示醫師的效果，但對於當下醫師診療病患處置及作出鑑別診斷的過程並無幫助。若要達到減少診斷錯誤的發生，只預測動向與預後是不夠的，因此針對預測病患應接受的檢查與處置發展模型，輔助醫師做出臨床決策有其必要性。因此本研究中，希望能達到以下研究目的：
利用常規收集的病患生理數據與主訴建立模型預測病患需要的檢查，處置與用藥。本次研究中使用雙和醫院急診室成人病患之電子病歷，進行回溯性病歷分析。收案期間為2017年5月至2019年12月，排除到院前死亡，非緊急醫療需求，兒科與外科病患，最後進入分析的樣本數為176,806就診人次。收集的指標包含結構化數據（病人檢傷生命徵象）以及文字敘述（病人主訴與病史），利用機器學型與自然語言模型建立預測模型。預測病患此次就診 1) 是否會抽血檢查 2) 是否需要放射科檢查 3) 是否需要緊急侵入性處置 4) 是否需要使用特定藥物治療。本次研究可發現在機器學習方法中，以logistic regression 表現最佳，其AUROC 可高達0.908, 而BioClinical Bert的模型表現更可高達0.925. 因此證實使用例行性檢測之生理數據結合病人的主訴可成功預測病患在急診所需的處置。然而本次研究的樣本只收集雙和醫院急診病患，並無進行外部驗證，因此可能無法直接推行到其他地方。未來需進一步進行外部驗證，讓模型更佳成熟完美。
Previous research has found that the incidence of diagnostic errors in the emergency department can reach as high as 15-55%. However, these errors can be mitigated through improved approaches such as healthcare professional re-education and the use of reminder checklists. Establishing Clinical Decision Support Systems (CDSS) that provide reminder checklists during consultations has shown to be more effective in reducing diagnostic errors or delays caused by individual factors, thereby enhancing the quality of patient care.It is not a novel concept to develop predictive models using machine learning methods for emergency department patients. However, while knowledge of a patient's prognosis can alarm physicians, it does not assist in the immediate decision-making process for patient management. To reduce the occurrence of diagnostic errors, it is insufficient to only predict dispositions and prognoses. Therefore, it is necessary to develop models that predict related interventions and medications in order to assist physicians. Thus, the objectives of this study are as follows:To establish models using routinely collected patient physiological data and chief complaints to predict the examinations, interventions, and medications that patients may require.
This study is a retrospective medical chart analysis. All the adults who visited emergency department in Shuang-ho hospital during May 2017 to December 2019 were included. Exclusion criteria include pre-arrival deaths, non-urgent medical needs, pediatric and surgical patients, resulting in a final analysis sample of 176,806 patient visits. Indicators include structured data (patient vital signs) and narrative descriptions (chief complaints and medical histories). Machine learning method and natural language processing were used to establish the predictive models. The model is to predict whether the patient will require: 1) blood tests, 2) radiology examinations, 3) urgent invasive procedures, and 4) specific medication treatments.Among the machine learning methods applied, logistic regression out-performed with an AUROC of up to 0.908, while the BioClinical Bert model performed even better, with an AUROC of up to 0.925. Therefore, combining routine physiological data and patient chief complaints can successfully predict the interventions and examinations required for patients in the emergency department.However, it is important to note that this study only used data from single hospital without external validation. Therefore, generalizing the result in other hospital settings may not be possible. Further external validation is necessary in the future to make the model more mature and robust. |
