Multi-Modal Feature Fusion-Based Machine Learning to Detect Abnormal Mechanical Ventilation

Abstract

Mechanical ventilation (MV) is a critical life-supportive technique for saving patients with acute respiratory failure. Abnormal ventilation happens frequently due to patient-ventilator dyssynchrony (PVD), condensation in the circuit, increased airway resistance, and so on. The previous studies that only rely on time-domain features fail to provide high identification accuracy. In this study, we develop a machine learning method to detect abnormal ventilation from ventilator waveforms. This method includes not only multi-modal features, but also time-domain, time-frequency, and entropic features in machine learning. We apply three classical machine learning models (random forest, support vector machine, and k-nearest neighboring) to detect five types of abnormal ventilation, including three types of PVD (missed triggering, double triggering, and prolonged cycling), circuitry condensation, and the flow expiration limit caused by high airway resistance. The results show that the optimal F1 scores for detecting prolonged cycling, double triggering, missed triggering, circuitry condensation, and expiratory flow limit are 97.56%, 92.26%, 96.46%, 89.18%, and 96.05%, respectively, which are superior to the results using purely time-domain features. In conclusion, the fusion of multi-modal features is beneficial to the identification of abnormal ventilation. It is promising to promote the application of machine learning models to detect abnormal ventilation in real clinical settings.

Keywords

• Patient-ventilator dyssynchrony
• condensation
• expiratory flow limit
• multi-modal features
• random forest