Identification of Mitral Valve Disease from Cardiac Potentials in Canine Patients with High-Density, E-Textile, Conducting Polymer BSPM Arrays

Valvular heart diseases (VHD) are a major contributor to cardiovascular mortality worldwide, affecting ~75 million people every year . The detection of VHD is currently achieved by means of non-invasive mechanical activity imaging modalities such as magnetic resonance imaging (MRI), computed tomography (CT) or echocardiography. These techniques require expensive equipment and technical user expertise which hinder their employment in general practice clinics. As a consequence, VHD can be only diagnosed at larger medical centres where these machines and the personnel required to operate them are available. In addition, these imaging modalities cannot be employed for the long-term monitoring of acute conditions outside healthcare centres, which forces patients with these conditions to remain in hospital for long periods of time.<br/><br/>Electrocardiography (ECG) is a well-established and inexpensive method for the identification of electrical heart conditions, or rhythms, which is readily available in most local clinics. Despite the fact that ECG is designed to detect electrical activity, it has been shown to be able to capture some of the symptoms caused by VHD. Furthermore, the implementation of deep learning (DL) algorithms which analyse conventional 12-lead ECG data has recently demonstrated the ability of ECG to fully characterise and classify between different types of VHD.<br/><br/>Even though DL algorithms are able to identify VHD, the maximum accuracy they can achieve is limited to ~80%. Furthermore, these complex classifiers require vast amounts of fitting data, long training times and offer little clinical insights into what features of ECG are involved in the identification of VHD. One of the primary factors which restricts the performance of these classifiers is the limited amount of spatial data the 12-lead ECG offers. The sampling of multiple cardiac sites across the torso by means of high-density electrodes in the form of body surface potential maps (BSPM) offers the opportunity to train classical machine learning algorithms on temporal and spatial ECG features, thus increasing both the explainability and the maximum achievable accuracy of VHD classifiers.<br/><br/>In this work, a novel high-density electrode array fabrication technique on e-textiles with conducting polymer coatings is presented through the implementation of a tailored vest for canine models. The data obtained from this vest is presented and specific temporal and spatial features are employed to train a classical machine learning algorithm able to identify mitral valve disease in a canine patient cohort. The results from the classifier are presented in a human-readable format and performance and training metrics are compared and shown to surpass those of state-of-the-art models in the literature. The results hereby obtained pave the way towards the use of BSPM as an affordable and scalable screening technique for VHD.