機器學習應用於MIMIC III資料庫，使用呼吸訓練時呼吸電阻、心律變異與臨床資訊預測拔管成敗

Abstract

背景: 侵入性機械通氣仍然是作為拯救重症患者生命的主要治療方法。然而，過久的侵入性機械通氣可能會增加呼吸器相關肺炎的風險，與增加醫院費用和死亡率。為評估是否適合脫離機械通氣，必須進行多因素評估，包括生理狀態、鎮靜劑使用、通氣機設置、插管的主要原因是否緩解，以及是否擁有通過自發性呼吸試驗的能力。不論是延長侵入性機械通氣的時間或因為拔管失敗都會產生相對應的風險，也因此，確定拔管的時機非常重要。到目前為止，還沒有關於拔管的統一評估準則。 目標: 本研究的主要目的是調查：（1）在自主呼吸訓練期間的電阻式呼吸訊號與心律變異度所提取的特徵，是否能夠用於訓練機器學習模型來預測拔管失敗；以及（2）綜合結構化數據和上述信號特徵的模型是否能夠優於僅使用結構化數據的模型。 方法: 本研究乃利用已公開的美國大型資料庫，分別被稱為多參數智能監測重症監護 (Multi-parameter Intelligent Monitoring in Intensive Care (MIMIC-III)) 和MIMIC-III匹配的波形數據庫。其中包含各種資訊，包括結構化和波形數據（電阻式呼吸訊號和心電圖）。納入已進行自主呼吸試驗並具有可用波形信號的患者。利用波形分析來提取電阻式呼吸訊號的特徵，和心律變異度分析來萃取心電圖的資訊。使用五種常見的機器學習模型進行拔管失敗的預測。此外，我們使用不同的數據類型組合來訓練模型，以評估它們對性能的影響。使用DeLong test的檢驗比較不同模型之間的AUROC，並且用Shapley value來解釋特徵對於模型預測的影響程度。 結果: 最終我們納入了 411 位滿足所有標準的拔管患者，有 46 位患者（11.2%）拔管失敗。我們發現，當使用所有的數據，包括電阻式呼吸訊號、心律變異度和結構化數據進行訓練時，XGBoost分類器優於其他分類器，並且比使用單獨的數據類型（僅結構化數據、僅電阻式呼吸訊號、僅心律變異度或綜合電阻式呼吸訊號和心律變異度）進行訓練的 XGBoost 分類器具有顯著更高的成效。特徵重要性分析顯示，電阻式呼吸訊號和心律變異度的特徵佔了 XGBoost 分類器前二十個重要特徵的一半以上。 結論：這些發現表明，將從電阻式呼吸訊號和心律變異度獲取的特徵納入機器學習模型，可以提高預測機械通氣下重症患者拔管失敗的成效。然而，需要進一步研究以驗證這些結果在不同臨床背景下的有效性。

Background: A multifaceted evaluation is necessary for weaning patients off mechanical ventilation (MV), and identifying the optimal time for extubation is crucial. The objective of this study was to explore whether impedance pneumography (IP) and heart rate variability (HRV) features could be utilized to train or improve a machine learning (ML) model for predicting post-extubation failure (PF) in mechanically ventilated patients in the intensive care unit. Methods: The MIMIC-III and MIMIC-III Waveform Database Matched Subset were the sources of data for this study, which consisted of clinical and waveform data, including IP signals and electrocardiogram (ECG) readings. Patients who underwent spontaneous breathing trials and had available waveform signals before extubation were eligible for inclusion. After a sequential assessment of waveform quality, the IP signal features were extracted, and HRV analysis was used to present the ECG features. Five common machine learning models were trained to predict PF. Moreover, the impact of different combinations of data types on the model's performance was evaluated. DeLong's test was utilized to compare the AUROC among individual models, and the SHapley Additive exPlanations method was applied to evaluate the models' expandability. Results: The final cohort included 411 extubated patients who met all the criteria, of whom 46 patients (11.2%) failed extubation. Our findings indicated that XGBoost classifier performed better than other classifiers when trained with the entire dataset (including IP signal, HRV, and clinical data) and produced a significantly higher AUROC than when it was trained using only one type of data (clinical data alone, IP signal alone, HRV alone, or both IP signal and HRV). Feature importance analysis revealed that IP waveform and HRV features constituted over half of the top 20 important features of the XGBoost classifier. Conclusion: The results indicate that integrating IP signal and HRV features can enhance the performance of ML models in predicting PF among critically ill patients on MV. Nevertheless, additional research is required to verify these outcomes in diverse clinical scenarios.