Detection of differences of cardiorespiratory metrics between non-invasive respiratory support modes using machine learning methods

Abstract

Many preterm infants require invasive mechanical ventilation in the first days of life. Once ventilation is removed, non-invasive respiratory support assists the transition to spontaneous breathing. Three common support modes are Nasal Continuous Positive Airway Pressure (CPAP), Non-Invasive Neurally Adjusted Ventilatory Assist (NIV-NAVA), and Nasal Intermittent Positive Pressure Ventilation (NIPPV). Cardiorespiratory behavior under these supports is not well understood. We hypothesized that if behavior changed with the support, then machine learning should be able to distinguish them, and the most discriminatory features would indicate how the behaviors differ. To test this, we recorded cardiorespiratory signals from infants during CPAP, NIV-NAVA, and NIPPV immediately following extubation. Forty-five metrics related to the amplitude and frequency of these signals were computed at each time and six summary statistics computed for each metric. The feature set was reduced by principal component analysis and used to train a Random Forest classifier. We found that (1) Classification of the three modes performed poorly (accuracy = 58.6%); (2) Classification of CPAP from NIV-NAVA or NIPPV performed well (accuracy = 81.5%). Thus, cardiorespiratory behavior during CPAP is significantly different from NIV-NAVA or NIPPV. The features with the greatest discriminatory power were related to the amplitude of respiratory movements and the correlation between the ribcage and abdomen movements. These results serve to direct further investigation with clinical relevance. Thus, infants with shallow breathing may benefit from the support with largest amplitude respiratory movements, while infants with asynchronous chest and abdominal movements may benefit from the modality with the most correlated movements.