Deep Analysis of EIT Dataset to Classify Apnea and Non-Apnea Cases in Neonatal Patients

Electrical impedance tomography (EIT) is a non-invasive imaging modality that can provide information about dynamic volume changes in the lung. This type of image does not represent structural lung information but provides changes in regions over time. EIT raw datasets or boundary voltages are comprised of two components, termed real and imaginary parts, due to the nature of cell membranes of the lung tissue. In this paper, we present the first use of EIT boundary voltage data obtained from infants for the automatic detection of apnea using machine learning, and investigate which components contain the main features of apnea events. We selected 15 premature neonates with an episode of apnea in their breathing pattern and applied a hybrid classification model that combines two established methods; a pre-trained transfer learning method with a convolutional neural network with 50 layers deep (ResNet50) architecture, and a support vector machine (SVM) classifier. ResNet50 training was undertaken using an ImageNet dataset. The learnt parameters were fed into the SVM classifier to identify apnea and non-apnea cases from neonates' EIT datasets. The performance of our classification approach on the real part, the imaginary part and the absolute value of EIT boundary voltage datasets were investigated. We discovered that the imaginary component contained a larger proportion of apnea features.