Analysis of Respiratory Flow Signals using Poincaré Plot Descriptors for Machine Learning-Based Prediction of Ventilator Weaning Success

Acute respiratory distress syndrome (ARDS) is a severe pulmonary condition that often requires mechanical ventilation (MV) to ensure adequate gas exchange and minimize ventilator-induced lung injury. This study compares statistical descriptors derived from respiratory time series with nonlinear variability metrics obtained from Poincaré plots to predict the outcomes of the weaning process in mechanically ventilated patients. The database, which includes respiratory flow recordings from 243 patients who underwent a standardized 30-minute spontaneous breathing test (SBT), was used to validate the classification models. Patients were categorized into three clinical outcome groups: successful weaning (n = 132), failed weaning (n = 88), and reintubation within 48 hours after completion of the trial (n = 23). Features were selected using nonparametric statistical tests and correlation analysis, eliminating redundancy and retaining discriminative variables. Two machine learning (ML) classifiers, random forest and feedforward neural network, were designed to identify patients belonging to each of the three clinical outcome groups. Model performance was assessed using stratified hold-out cross-validation repeated over 150 iterations, with hyperparameters optimized using Bayesian methods. The random forest classifier using Poincaré descriptors achieved a mean accuracy of 90.4% and higher F1-scores across all groups, including those who required reintubation. These findings suggest that Poincaré-based variability metrics, in combination with ensemble learning, may enhance the accurate prediction of MV weaning outcomes in ARDS patients.