The Feasibility of a Machine Learning Approach in Predicting Successful Ventilator Mode Shifting for Adult Patients in the Medical Intensive Care Unit

Abstract

Background and Objectives: Traditional assessment of the readiness for the weaning from the mechanical ventilator (MV) needs respiratory parameters in a spontaneous breath. Exempted from the MV disconnecting and manual measurements of weaning parameters, a prediction model based on parameters from MV and electronic medical records (EMRs) may help the assessment before spontaneous breath trials. The study aimed to develop prediction models using machine learning techniques with parameters from the ventilator and EMRs for predicting successful ventilator mode shifting in the medical intensive care unit. Materials and Methods: A retrospective analysis of 1483 adult patients with mechanical ventilators for acute respiratory failure in three medical intensive care units between April 2015 and October 2017 was conducted by machine learning techniques to establish the predicting models. The input candidate parameters included ventilator setting and measurements, patients’ demographics, arterial blood gas, laboratory results, and vital signs. Several classification algorithms were evaluated to fit the models, including Lasso Regression, Ridge Regression, Elastic Net, Random Forest, Extreme Gradient Boosting (XGBoost), Support Vector Machine, and Artificial Neural Network according to the area under the Receiver Operating Characteristic curves (AUROC). Results: Two models were built to predict the success shifting from full to partial support ventilation (WPMV model) or from partial support to the T-piece trial (sSBT model). In total, 3 MV and 13 nonpulmonary features were selected for the WPMV model with the XGBoost algorithm. The sSBT model was built with 8 MV and 4 nonpulmonary features with the Random Forest algorithm. The AUROC of the WPMV model and sSBT model were 0.76 and 0.79, respectively. Conclusions: The weaning predictions using machine learning and parameters from MV and EMRs have acceptable performance. Without manual measurements, a decision-making system would be feasible for the continuous prediction of mode shifting when the novel models process real-time data from MV and EMRs.