Dynamic prediction of late noninvasive ventilation failure in intensive care unit using a time adaptive machine model. (September 2021)

• Record Type: Journal Article
• Authors: Feng, Xue; Pan, Su; Yan, Molei; Shen, Yanfei; Liu, Xiaoqing; Cai, Guolong; Ning, Gangmin
• Is Part Of: Computer methods and programs in biomedicine. Volume 208(2021)
• Publication Date: 2021-09
• Subjects: Noninvasive ventilation -- Dynamic prediction -- Light gradient boosting machine -- Attention mechanism; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.106290
• Languages: English
• ISSNs: 0169-2607

Abstract

Abstract: Background: Noninvasive ventilation (NIV) failure is strongly associated with poor prognosis. Nowadays, plenty of mature studies have been proposed to predict early NIV failure (within 48 hours of NIV), however, the prediction for late NIV failure (after 48 hours of NIV) lacks sufficient research. Late NIV failure delays intubation resulting in the increasing mortality of the patients. Therefore, it is of great significance to expeditiously predict the late NIV failure. In order to dynamically predict late NIV failure, we proposed a Time Updated Light Gradient Boosting Machine (TULightGBM) model. Material and Methods: In this work, 5653 patients undergoing NIV over 48 hours were extracted from the database of Medical Information Mart for Intensive Care Ⅲ (MIMIC-Ⅲ) for model construction. The TULightGBM model consists of a series of sub-models which learn clinical information from updating data within 48 hours of NIV and integrates the outputs of the sub-models by the dynamic attention mechanism to predict late NIV failure. The performance of the proposed TULightGBM model was assessed by comparison with common models of logistic regression (LR), random forest (RF), LightGBM, eXtreme gradient boosting (XGBoost), artificial neural network (ANN), and long short-term memory (LSTM). Results: The TULightGBM model yielded prediction results at 8, 16, 24, 36, and 48 hours after the start of the NIV with dynamic AUC values of 0.8323, 0.8435, 0.8576, 0.8886, and 0.9123, respectively. Furthermore, the sensitivity, specificity, and accuracy of the TULightGBM model were 0.8207, 0.8164, and 0.8184, respectively. The proposed model achieved superior performance over other tested models. Conclusions: The TULightGBM model is able to dynamically predict the late NIV failure with high accuracy and offer potential decision support for clinical practice. Graphical Abstract: Image, graphical abstract