基於深度學習方法之急診心臟病患住院預測研究

Abstract

急診壅塞問題將增加病患等待時間和造成醫療資源配置的困難，故能於病患檢傷階段進行住院預測可將醫療資源配置給急需急診醫療資源之患者。本研究以「台北馬偕教學醫院」之2011年至2018年共計八個年度，1,065,480筆急診病患於檢傷階段可得之主訴、基本資料與診斷資料為研究資料，用以預測住院可能性。本研究是以XGBoost來進行結構化變數的篩選，然後以BERT來進行主訴和補充主訴的住院預測，最後再以深度學習的方法來對於主訴和補充主訴進行住院預測。本研究涵蓋一系列的深度學習的流程，包含了結構化資料與主訴資料前處理、主訴之否定詞處理、不平衡資料集處理、心臟疾病判別、結構化變數轉成補充主訴以及XGBoost、BERT、BiLSTM和CNN模型建立及評估。研究結果發現住院預XGBoost結果最高AUC為0.8182、BERT結果最高AUC為0.8859、BiLSTM結果最高AUC為0.9447、CNN結果最高AUC為0.9268。研究推論深度學習模型在住院預測方面有較好的預測結果。再加入心臟疾病後住院預測BERT結果最高AUC為0.8972、BiLSTM結果最高AUC為0.9361、CNN結果最高AUC為0.9341。研究推論心臟疾病對於住院預測是可以提升預測力。研究方法與發現提供急診住院預測參考並希冀提升急診室資源有效配置。

The congestion issue in the emergency room will augment patients' waiting time and pose challenges in the allocation of medical resources. Therefore, by conducting inpatient prediction during the triage stage, healthcare resources can be allocated to patients in urgent need of emergency medical attention. This study utilized data from Taipei Mackay Memorial Hospital spanning eight years, from 2011 to 2018, encompassing a total of 1,065,480 records of emergency patients' chief complaints, basic information, and diagnostic data for research purposes, aiming topredict the likelihood of hospitalization.This research employed XGBoost for the selection of structured variables, followed by BERT for predicting hospitalization based on chief complaints and supplementary complaints. Finally, deep learning techniques were utilized to predict hospitalization regarding both chief complaints and supplementary complaints. The study encompassed a series of deep learning processes, including preprocessing of structured and chief complaint data, handling of negation terms in chief complaints, addressing imbalanced datasets, discerning cardiac ailments, converting structured variables into supplementary complaints, and constructing and evaluating XGBoost, BERT, BiLSTM, and CNN models.The research findings indicated that the highest AUC for hospitalization prediction was achieved by XGBoost with a score of 0.8182, followed by BERT with a score of 0.8859, BiLSTM with a score of 0.9447, and CNN with a score of 0.9268. It was inferred that deep learning models exhibited superior predictive outcomes in hospitalization prediction. When incorporating cardiac ailments into the hospitalization prediction, the highest AUC was observed with BERT at 0.8972, followed by BiLSTM at 0.9361, and CNN at 0.9341. It was concluded that the inclusion of cardiac ailments enhanced the predictive power in hospitalization prediction. The research methodology and findings provide valuable insights for reference in emergency room hospitalization prediction, with the aim of effectively optimizing emergency department resources.