Abstract 4365172: Aortic Valve Type and Stenosis Can Be Identified from Wearable-Recorded Seismocardiograms Using Machine-Learned Classifiers

Abstract

Introduction: Patients with aortic valve disease, such as bicuspid aortic valve (BAV), require regular echocardiography or cardiovascular (CV) MRI to monitor for complications such as valve stenosis (AS) and aortic dilation. However, repeated imaging can be burdensome and incur substantial cost. Seismocardiogram (SCG) chest acceleration measurements recorded by inexpensive wearable devices can give indicators of valve-mediated hemodynamic changes, and as such may have supplemental value for such patients. This study investigated using SCG recordings coupled with a novel machine-learned (ML) classifier for SCG signals to identify patient valve type and presence/absence of aortic valve stenosis (AS). Hypothesis: We hypothesize that accurate classification of aortic valve type and AS can be made from SCG recordings with ML analysis compared to those from standard-of-care imaging (ground truth: cardiac MRI or echo). Methods: Healthy controls (no known CV disease) and aortic valve disease patients with tricuspid (TAV), BAV, or post-repair mechanical valve who received echo or MRI (clinical CV protocol) were enrolled for same-day 2-minute wearable SCG measurement (fig. A). Standard clinical assessment of valve/flow function was used (fig. B). Informed consent was given with IRB oversight. Clinical imaging used 4D flow MRI (1.5T,1-3mm3/30-40ms) or 2D Doppler echo (1.7-3.3MHz,12-40FPS). From clinical read of valve type/function, subjects were grouped in four classes: AS (any degree), BAV no-AS, TAV no-AS, mechanical. A hybrid network with convolutional neural network and multi-layer perceptron was trained (80/20 train/test) to classify patient valve status from SCG wavelet coefficients and demographics (age/sex/height/weight). Performance was evaluated by 20-fold cross-validation. Results: Enrolment was 129 subjects (97 MRI/32 echo): 46 controls (45.9±17.4y/20F) and 83 patients (22.4±15.8y/20F; 67 BAV/6 TAV/10 mech.). Classification area-under-curve (AUC) was high for all classes (AUC≥0.79). Across all ML validations, correct classification was achieved for ≥75% of subjects. Conclusion: This evaluation of a machine-learned classifier for SCG indicate potential utility in screening for valve-mediated hemodynamic changes, which reverberate through the chest and cause altered vibrations. The low cost and ease of acquisition for SCG would make it an appealing complement to imaging as the current standard for aortic valve abnormality screening and management.