建立人工智能調整優化模型以利呼吸器脫離及教學

Abstract

呼吸器是重症加護病房中維持生命的重要醫療設備，成功脫離呼吸器成為加護病房重要的課題與目標之一。過去研究顯示，臨床往往會延遲拔管，因而增加呼吸器使用的併發症，或死亡率。長期使用或判斷錯誤導致重插管的併發症，也將延長加護病房住院時間與增加醫療花費。因此，如何訓練脫離與判斷拔管時機一直是重要的研究議題。然而，至今臨床並沒有一套完整且一致的呼吸器脫離標準，台灣的病人族群也可能與國外不同。因此，本研究想藉由人工智能建立一個呼吸器脫離預測模型，預測脫離時機，也藉此評估各個指標之重要性，驗證臨床操作經驗，並提供臨床操作實務參考，以期能優化參數調整，預測脫離率，輔助臨床脫離時機判斷，並運用於重症臨床呼吸器操作教學。本研究於2022年11月開始，收案對象為插管進入台大醫院成人內科加護病房之病患。收集每天常規記錄的呼吸治療參數與各項生命徵象，常規檢驗室報告。選用30個特徵資料，並將資料分為拔管前24小時、拔管前48小時及拔管前72小時，共三組，透過極限梯度提升分類器做模型訓練，並使用準確率、精確率、敏感度、特異性、F1分數、ROC-AUC等指標進行模型預測性能評估。在本研究中發現，傳統的脫離指標RSBI在拔管成功與失敗組中並沒有達到統計上的顯著差異。而成功組的呼吸器使用天數，顯著少於失敗組。在模型預測上可以發現使用拔管前72小時的平均資料做為模型訓練，有最佳的預測表現，預測脫離的準確率為68%，AUC可達0.75。

The Mechanical ventilator is a life-saving and important support system in intensive care units (ICU). The issue of successfully weaning mechanical ventilation remains a challenge in ICU. Delays in extubation can increase complications, mortality, ICU stays, and healthcare costs. Therefore, determining optimal timing for weaning and extubation remains a crucial research topic. However, there is no standard protocol exists in clinical practice, and patient demographics in Taiwan may differ from those abroad. This study aims to develop a ventilator weaning prediction model using artificial intelligence, assessing the importance of various indicators, validating clinical experience, and providing practical guidance for clinical operations. The goal is to optimize parameter adjustments, predict weaning success rates, assist in clinical decision-making on extubation timing, and aid in ICU clinical education. Starting in November 2022 at National Taiwan University Hospital, the study involved adult ICU patients, collecting respiratory therapy parameters, vital signs, and lab data. 30 features were selected and grouped into 3 sets: 24, 48, and 72 hours before extubation. A XGboost was trained on these data, evaluating model performance using metrics such as accuracy, precision, sensitivity, specificity, F1 score, and ROC-AUC. Results: RSBI did not show statistically significant differences between successful and failure extubation groups. The duration of ventilator use was significantly shorter in successful cases compared to failure group. The model achieved optimal predictive performance when trained on average data from the 72 hours before extubation, with an accuracy of 68% and an AUC of 0.75.