A machine learning model for real-time asynchronous breathing monitoring

Abstract The occurrence of asynchronous breathing (AB) during mechanical ventilation (MV) can have detrimental effect towards a patient’s recovery. Hence, it is essential to develop an algorithm to automate AB detection in real-time. In this study, a method for AB detection using machine learning, in particular, Convolutional Neural Network, (CNN), is presented and its performance in identifying AB when trained with different amount of training datasets and different types of training datasets is evaluated and compared between standard manual detection. A total of 486,200 breaths were analyzed in this study. It was found that the CNN algorithm achieved 69.4% sensitivity and 37.1% specificity when trained with 2000 AB cycles and 1000 normal breathing (NB) cycles; however, when it was trained with 5500 AB and 5500 NB, the CNN achieved 96.9% sensitivity and 63.7% specificity. The experimental results also indicate that the CNN was trained with modified images (region under the curve) CNN yielded sensitivity of 98.5% and specificity of 89.4% as opposed to sensitivity of 25.3% and 83.9% specificity when trained with line graph instead. Therefore the proposed method can potentially provide real-time assessment and information for the clinicians.