基於自然語言技術之急診病患檢傷階段再住院預測研究

Abstract

在急診早期檢傷階段中，病患主訴語(chief complaints)為後續檢傷與醫療初步判定之重要依據。本研究以台北馬偕醫院2011到2018的八年度急診病患就診資料，將檢傷階段可取得之主訴語、年齡、檢傷分級、到院模式…等資料經過非結構後與結構化資料預處理、關鍵字分析、自然語言處理模型，機器學習程序，以進行急診住院、再住院、再入院、死亡預測的實證分析研究。研究在機器學習的程序上，首先採用類神經網路之Word2vec詞嵌入語言模型，由主訴篩選住院相關之重要語意關聯詞，研究進而透過BERT模型進行後續住院、再住院、再入院、死亡預測的研究。研究採用Word2vec與BERT自然語言處理模型進行預測研究，預期可以協助醫院及早準備重症病患相關醫療資源。實驗結果顯示(1) BERT模型預測效力優於Word2vec模型；(2)採用主訴語可以有好的急診住院、再住院、再入院、死亡預測力，於住院、再住院、再入院、死亡預測方面，BERT模型單純採用主訴進行預測， AUC分別為0.9446、0.9877、0.9883、1.0000；(3)考慮結構化變數以產生本研究提出之延伸主訴(Expanded CCs)概念將可提升急診再住院預測效果，BERT模型於住院、再住院、再入院、死亡預測方面的AUC分別為0.9611、0.9949、0.9947、1.0000；(4) 在死亡預測方面，不論是否不平衡處理，單純採用主訴的情況下，Word2vec在維度50的0.8750 AUC優於維度200的0.8394 AUC；若考慮檢傷階段的重要結構化變數，Word2vec在維度50的0.7730 AUC優於維度200的0.7325 AUC，採用單純主訴的各項評估值優於考慮重要結構化變數之Expanded CCs的各項評估值。本研究提出之架構與延伸主訴概念可提供急診預測相關研究的參考。In the early stage of emergency department triage, the patient's chief complaints are important for subsequent injury assessment and preliminary medical diagnosis. This study analyzed the eight-year emergency department data from Taipei Mackay Memorial Hospital from 2011 to 2018, including data such as chief complaints, age, injury severity score, and mode of arrival, which were processed through non-structured and structured data pre-processing, keyword analysis, natural language processing models, and machine learning algorithms to perform empirical analysis of emergency hospitalization, readmission, reentry, and death prediction. The study first used the Word2vec word embedding language model of neural networks to select important semantic keywords related to hospitalization from chief complaints, and then used the BERT model to perform subsequent hospitalization, readmission, reentry, and death prediction research. The study expected that the use of Word2vec and BERT natural language processing models in prediction research could help hospitals prepare early for critically ill patients. The experimental results showed that (1) the BERT model had better prediction performance than the Word2vec model; (2) the use of chief complaints can have good prediction performance for emergency hospitalization, readmission, reentry, and death, and the AUC of the BERT model in hospitalization, readmission, reentry, and death prediction were 0.9446, 0.9877, 0.9883, and 1.0000, respectively; (3) considering structured variables can improve the prediction performance of emergency readmission, and the BERT model's AUC in hospitalization, readmission, reentry, and death prediction were 0.9611, 0.9949, 0.9947, and 1.0000, respectively, when using the Expanded CCs concept proposed in this study; (4) in death prediction, regardless of whether it is imbalanced processing, the AUC of Word2vec in dimension 50 (0.8750 AUC) is better than that in dimension 200 (0.8394 AUC) when only considering chief complaints. If important structured variables in the injury assessment stage are considered, the AUC of Word2vecin dimension 50 (0.7730 AUC) is better than that in dimension 200 (0.7325 AUC), and the evaluation values of using only chief complaints are better than those of considering Expanded CCs with important structured variables. The framework and Expanded CCs concept proposed in this study can provide a reference for emergency prediction research.