【論文摘要】A Machine Learning Model for Predicting Successful Extubation in Intensive Care Units

【論文摘要】以機器學習模型來預測重症病人的成功拔管

鄭愛琴(AC Cheng) ； 劉忠峰(CF Liu) ； 陳欽明(CM Chen) ； 柯獻欽(SH Ko) ； 柯順耀(SY Ko) ； 紀宗呈(TC Chi)

《呼吸治療》 19卷附冊 (2020/10) Pp. 32-32

https://doi.org/10.6269/JRT.202010/SP_19.0020

摘要

Background: A considerable percentage of intensive care unit (ICU) patients require endotracheal intubation. However, clinicians have limited ability to predict extubation success. There is a need for more formidable tool to help judge the optimal time for extubation. Our objective is to develop machine learning prediction models for successful extubation in ICU ventilated patients. Methods: From January 1, 2019 to May 31, 2019, data from 674 patients with planned extubation in the ICU at the Chi-Mei Medical Center were used to train and test five machine learning algorithms (Logistic Regression [LR], Random Forest [RF], Support Vector Machine [SVM], K Nearest Neighbor [KNN], Light Gradient Boosting Machine [Light GBM]). Data collected and analyzed include personal characteristics (e.g., BMI, gender), validation of a severity of illness score (APACHE II), the total number of intubation attempts, the original disease diagnosis and comorbidities, spontaneous breathing trial (SBT) and vital signs before and after extubation. The feature variables were the 32 clinical risk factors; the output variable is predictive of extubation failure, defined as reintubation within 72 hours of extubation. 70% of the datasets were used for training, and 30% were used for validation. Results: Patients with planned extubation were divided into success and failure groups. The successful extubation ratio was 96.3%; age, APACHE II, BUN before extubation, ICU length, hospital mortality and medical costs, were significantly lower in the successfully extubated group than in the failed group (p ＜0.05). The overall performance of the LR model demonstrated an F1 score of 0.95, accuracy of 0.94, precision of 0.98, sensitivity of 1.00, and specificity of 0.93. The area under the receiver operating characteristic curve (AUC) was 0.97, which is better than four other algorithms and any of the following predictors: APACHE II (odds ratio [OR]:1.051; 95% Confidence Intervals [CI]: 1.001-1.041), age (OR: 1.028; 95% CI: 0.999-1.057), and blood urea nitrogen (BUN) (OR: 1.020; 95% CI: 1.008-1.033). Conclusions: The LR performed well in predicting extubation failure, which can be used to predict successful planned extubation.

關鍵字

intensive care unit ； predictor ； successful extubation ； machine learning ； logistic regression