Mechanical Ventilation Mode Classification: A Dual-Input Convolutional Neural Network Approach with Class Activation Mapping

Mechanical ventilation (MV) is critical for patients with respiratory failure, but patient-ventilator asynchrony (PVA) due to poor patient-ventilator interaction can lead to increased mortality rates. Although numerous machine learning (ML) models were developed to assist with the detection of PVA, they do not often include information on MV mode, increasing detection error rates. Specifically, PVA phenotypes manifest differently in different MV modes, and knowing the specific PVA type is important for modifying sub-optimal ventilator settings. This study presents a dual-input convolutional neural network (CNN) model for MV mode classification utilising 1D and 2D input data structures. The models in this study were developed to perform MV mode classification between breaths of pressure-controlled (PC) or volume-controlled (VC) mode. Data from 17 MV patients were used for training, and outcome models were tested on an independent dataset of 10 patients. The testing dataset consists of 448,772 breaths, and the 2D model obtained an overall accuracy of 92.14% as opposed to 61.81% for the 1D model. Class activation mapping (CAM) was also incorporated to better understand model decision-making process and provide insight on which waveform features influence MV classification. Overall, MV mode classification using the developed dual-input CNN models shows the potential to improve PVA identification and asynchronous waveform reconstruction by automatically providing prior information on the MV mode used.