Abstract
Over 5 billion people worldwide, many of whom inhabit remote, medically underserved areas, own cell phones. Without easy access to screening, diagnostic, and monitoring services, these individuals face adverse health outcomes. Solutions for this pervasive dilemma should ideally be accurate, financially viable, and scalable to the population level. This study aims to demonstrate the accuracy of a proprietary mathematical method designed to interpret acoustic data captured by any unmodified cell phone. Our methods use raw audio recordings to generate cardiovascular health data metrics similar to current advanced imaging technologies. Our team approached patients who presented to the University of Louisville Emergency Department with stroke or cardiac symptoms who underwent advanced diagnostic testing (Computed Tomography Angiogram (CTA) of the head and neck, and Transthoracic Echocardiography (TTE)) between December 2020 and June 2021. This study was approved by the UofL Institutional Review Board. For subjects who consented, we obtained three or four 15-30 second recordings at predefined physical landmarks (left carotid, right carotid, aortic valve, mitral valve) via an unmodified cellular phone microphone held to the skin. We implemented a proprietary mathematical model using deterministic mathematics and nonlinear dynamics of unaltered acoustic data. This model generates cardiovascular health data metrics: stroke volume (SV), ejection fraction (EF), and valvular function for each patient. We then took the known SV and EF data from advanced imaging and determined the model parameters needed to accurately predict these values using the unmodified phone recordings. Once enough patients have been recruited to the study, the sound recordings will be used to predict these values independently from the advanced imaging data. This study aimed to recruit 100 individuals for data collection. For this interim analysis, we report data on 60 recordings from 60 subjects. The recordings were divided into two sets: 59 recordings analyzing ejection fraction (EF) and 38 recordings analyzing stroke volume (SV). For all 60 six-second recordings, 16,738 calculations were performed to generate fitted solutions of EF or SV. Low EF was predetermined as EF of 39% or less and low SV was predetermined as SV of 35 mL or less. Based on the data generated from the phone recordings, the proprietary model could identify low EF with 100% sensitivity and specificity and low SV with 100% sensitivity and at least 97.4% specificity. Using basic acoustic recordings, the proprietary model generated cardiovascular health data metrics that match gold standard testing with reasonable precision. Our deterministic and non-linear dynamic methods can be calibrated to individual patients and do not require massive amounts of data usually used in Artificial Intelligence (AI) or Machine Learning (ML) models. Given the ubiquity of common cell phones and the relative ease of transmitting short acoustic recordings, this model represents an opportunity to expand the frontier of medicine to people who live in remote and underserved areas.View Large Image Figure ViewerDownload Hi-res image Download (PPT)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have