Abstract 4357265: Mobile Phone Auscultation Using Non-linear Dynamics Analysis to Detect Aortic Stenosis

Introduction: Aortic stenosis (AS) is the most prevalent valvular disorder in the United States with a lengthy presymptomatic period and variable course. Patients with symptomatic or severe AS have markedly elevated morbidity and mortality. Barriers to gold-standard diagnostics, like echocardiogram, have contributed to delayed diagnoses and worsening health disparities. There is a need for technologies that improve access to screening for structural heart diseases (SHD). Research Question: Assess the feasibility of using recordings collected via mobile phone auscultation (MPA) and analyzed with non-linear dynamics to develop an accurate model fitted to echocardiogram findings of aortic stenosis. Methods: We conducted a single-site feasibility study, enrolling subjects into three groups based on echocardiogram: AS, non-AS SHD, and controls. Unmodified mobile phones were used to collect 30-second recordings at the aortic and left mid-axillary site. Recordings were processed using non-linear dynamics, a physics-based approach, to identify recording features (independent variables) which were then reduced and mapped to gold-standard echocardiogram findings (dependent variable). Blinded training and test sets were utilized to avoid overfitting. Additional demographic and medical history was extracted from the electronic record. Model performance was reported as area-under-the-curve, sensitivity, and specificity. Results: 248 subjects were enrolled, 69 (27.8%) AS, 97 (39.1%) non-AS SHD, and 82 (33.1%) controls. Median age for all subjects was 73 years (IQR 64-78 years), 108 (43.6%) were female, and 246 (99.2%) were non-Hispanic White. All recordings were of low dimensionality (DCorr < 3) and chaotic (maximal Lyapunov exponent > 0). Aortic site recordings were fitted separately. Iterative analysis identified three candidate predictors and produced two well performing models with the following performance on test sets, respectively: AUC 0.92 and 0.91, sensitivity 85% and 92%, specificity 90% and 82%. Conclusion: We piloted a novel, and successful approach in establishing feasibility of a highly accurate model for the identification of aortic stenosis using a small set of MPA recordings analyzed using non-linear dynamics. This approach uses mathematics to develop models with significantly fewer patients compared to models developed using AI. Our use of unmodified mobile phones paves the way to explore vastly improving access to patient populations across the globe.