Valvular heart disease (VHD) remains an underappreciated and growing public health concern, currently affecting over 13% of the population 75 years and older and is a major contributor to preventable mortality in the aging population of the United States (1).

Aortic stenosis (AS) and mitral regurgitation (MR) represent two of the most common valvular lesions in the U.S. Both conditions are significantly underdiagnosed, increase in prevalence with age, and have prolonged asymptomatic periods with variable natural histories. The primary means of reducing morbidity and mortality depend on consistent monitoring and early intervention, when applicable (2, 3).

The diagnostic pathway of valvular disease involves a combination of symptoms and physical findings, which are both imperfect and nonspecific (4, 5). Auscultation with stethoscopes remains the primary tool to screen for VHD and therefore largely limits clinicians to disease that is audible. In addition to inter-provider variability in the auscultation and classification of audible cardiac disease (e.g., murmurs), confirmation and monitoring requires echocardiography, the accepted gold standard (6) which necessitates specialized equipment, and highly trained personnel. Additional barriers to diagnostic care include access, transportation, health literacy, and insurance coverage, which perpetuate a system of late and undiagnosed VHD.

Screening, diagnosis, and monitoring of VHD would greatly benefit from novel technologies that do not overly rely on the availability of advanced imaging or subspecialists. Basic mobile phones contain sensitive recording devices, improving computational power, and if paired with accurate modeling software, could potentially address enumerable existing barriers to screen or monitor patients with VHD. We conducted a pilot study using auscultation recordings collected via unmodified mobile phones and analyzed using nonlinear dynamics software to develop accurate models fitted to echocardiographic findings for aortic stenosis and mitral regurgitation.

This pilot established the feasibility of two novelties: utilizing basic mobile phones with unmodified original microphones for the collection of sound data, evaluating what may be considered a near-universally available recording device given the ubiquity of mobile devices. Second, the potential utility of a dynamical approach to analyzing that sound data for an overall improved model performance. The study met predetermined goals in creating three acceptable models for aortic stenosis using auscultation recordings collected by unmodified mobile phones and analyzed with nonlinear dynamics software. In aortic stenosis modeling, the collective results using either binary (presence or absence of a diagnosis of aortic stenosis) or continuous dependent variables (indexed aortic valve area), yielded a sensitivity, specificity, and accuracy as high as 92.3%, 89.7%, and 88.5%, respectively. Based strictly on performance metrics, the categorical diagnostic model (AS2, Table 2) performed best in terms of AUC (0.872) and sensitivity (0.923).

Given insufficient enrollment of moderate-to-severe MR, a post-hoc exploratory analysis incorporating mild MR was performed to evaluate signal detectability across a broader disease spectrum. The post-hoc MR model still did not meet predefined performance targets for sensitivity and AUC, reflecting both limited sample size and increased heterogeneity in disease characterization, and had a sensitivity and specificity of 66.7% and 75.9%, respectively. These results should be interpreted as hypothesis-generating only. Unlike aortic stenosis, mitral regurgitation severity was frequently determined using qualitative echocardiographic assessment based on jet appearance rather than a single continuous quantitative parameter. This likely introduced label variability, particularly among mild cases, which may have attenuated model performance.

To be clear, the performance of our modeling only reflects the feasibility of the approach that uses mobile phones and nonlinear features. Further, for the purposes of this study, feasibility was established for isolated aortic stenosis only. Model performance should not be interpreted as clinical applicable or accurate and further research will be necessary to understand model development in mixed valvular disease (aortic stenosis with concomitant aortic insufficient and/or mitral disease). We recognize that classical linear features may also be appropriate for modeling VHD and further appreciate that our work needs to be validated in larger cohorts with differing prevalence estimates. Yet we do not lose sight of the importance of exploring new modeling methods and believe our work contributes a great deal. Further, the models described in this study do not intend to reproduce echocardiogram findings, a biophysically different modality for the detection of conditions. Rather, these nonlinear dynamic models likely reflect turbulence, the fluid dynamics of the specified condition. This method is another means of detecting the same condition using different variables which are then mapped to predict echocardiogram findings (see Figure 2).

As outlined in our methods, the models were derived through interrogation of different features (NLD variables, standard acoustics, and subject data). While some non-NLD features (e.g., gender, age) were found to be statistically significant in KS analysis, only NLD variables survived the modeling. Thus, all the features used in the combinatorial optimization process of model development came from nonlinear dynamics extraction. Although auscultation-based screening tools are often sensitive to body habitus and acoustic attenuation, the nonlinear dynamics features used here primarily reflect low-frequency hemodynamic turbulence, which propagates efficiently through tissue and is less dependent on surface acoustic transmission. Consistent with this, subject-level demographic and anthropometric variables were not retained in final models. Nonetheless, validation in larger and more diverse populations remains necessary to confirm generalizability. This makes our modeling process unique when compared to what is represented elsewhere in literature. Based on the calculations used, all the models likely represent specific types of turbulence that reflect stress caused by organ dysfunction and compensation. However, this study was not designed to verify the physiological nature of the features, only their modeling utility.

No recordings, collected by either study personnel or study subjects, were rejected secondary to sound quality, interference, or similar reason. Further, the analysis of the separately collected recordings supported the notion that they observations of the same system, and that self-recordings (e.g., remote monitoring) is feasible as well.

A noninvasive means of screening, diagnosing, and monitoring VHD has significant potential implications. The largest contemporary mortality review of patients with untreated aortic stenosis found that only 60% of patients with severe disease underwent treatment within 4 years of diagnosis. Further, over 40% of patients with AS had moderate to severe disease, among whom the 4-year all-cause mortality ranged from 33.5% to 44.9% (14). There is an urgent need for earlier detection among patients with AS, in addition to identifying ways to improve follow up and expedite intervention (14, 15).

Identifying and monitoring patients with mitral regurgitation earlier in the disease course is similarly important. Appropriate timing for valve repair/replacement leads to better outcomes (3). This remains a challenge across the United States, but rural populations are disproportionately affected, usually because of challenges accessing surgical centers (16).

Compared to our work presented here, prior studies exploring non-invasive methods to detect VHD relied heavily on machine learning algorithms and/or sophisticated hardware.

Chorba et al. assessed the performance of a deep learning model to detect murmurs using recordings from digital stethoscopes. Compared to echocardiograms, the algorithm had a sensitivity of 93.2% and specificity of 86.0% for AS and a sensitivity of 66.2% and a specificity 94.6% for MR. Yet, the authors excluded mild AS cases from modeling and had to exclude an additional 13% of recordings due to poor quality. Had held-out recordings been included, the algorithm's sensitivity for moderate-severe AS and MR murmur detection could have been as low as 84% and 57%, respectively (17). Prince et al. similarly developed deep learning algorithms to both detect and classify different forms of SHD using digital stethoscopes (18). For all audible SHD, the algorithm performed well (sensitivity 87.7% and specificity 82.0%), and for aortic stenosis specifically, the reported sensitivity and specificity was 100% and 73.5%, respectively.

A notable limitation of this work using digital stethoscopes is limiting analyses to audible murmurs. Of the 409 persons with echo-confirmed SHD in the study by Prince et al., one-third had inaudible disease and were unaccounted for in their algorithm performance (18). Additionally, both Prince and Chorba limited their analyses to moderate to severe AS, excluding mild AS from model performance. Models that are restricted to moderate-to-severe, audible heart disease will inevitably fall short of the population's screening needs. The methods used in this pilot were agnostic to audible nature of the enrolled subject's condition. At no point did we “listen” to the patient or even check if a murmur had been previously documented.

More recently, Zhou et al. performed a prospective study using AI stethoscopes to develop machine learning algorithms for the detection of left VHD by processing raw sounds, and extracting features across different domains including the nonlinear domain (19). Performance was best for the collective detection of left-sided VHD (sensitivity 70.0%, AUC 75.5%) with decreased performance for aortic stenosis specific models (sensitivity 66.7%, AUC 63.3%). The authors used raw sound, therefore did not limit analyses to audible murmurs and included subjects with mild disease, the combination of which could explain the relatively poor model performance compared to previously published work. None of the Test Set models had performance characteristics that exceeded 80%. Yet, the models were overly reliant on linear features extracted from the frequency and cepstrum domains. This work and previous work by study collaborators suggests that auscultation data is chaotic and nonlinear features are better predictors of human conditions. The main limitation of the acoustic features is that they generally assume stable periodicity and linear phasal relationships that do not typically exist in chaotic systems which places unnecessary burden on model fitting.

Dhingra et al. used artificial intelligence to develop an ensemble deep learning algorithm using electrocardiographic (ECG) images to screen for structural heart disease. The AUC for the detection of aortic stenosis was 0.804 (95% CI 0.784–0.824) and 0.792 (95% CI 0.776–0.807) for mitral regurgitation (20). Similarly, Poterucha et al. developed a deep learning model (EchoNext) for the detection of SHD using 12-lead ECG's and reported an algorithm performance with strong discriminatory capacity for aortic stenosis (AUC 0.864) (21). These SHD deep learning models using ECGs performed well overall and obviate the concerns for audibility. But like many digital stethoscope models, Dhingra and Poterucha excluded mild AS and optimized the model for positive predictive value at the expense of sensitivity, limiting their future utility as a screening tool.

An additional challenge in scaling VHD detection with AI algorithms using either digital stethoscopes or ECGs is a requirement for additional, potentially sophisticated and expensive, hardware. The NLD models presented here have impressive accuracy and less data burden and can be paired with any technological hardware that can record sound. This study used unmodified mobile phones for their high recording speeds, wide detection of frequencies (including infrasonic), excellent sensitivity, and most importantly, ubiquity, with estimates approaching 8 billion mobile phones in the world (22). Sound and data collection via mobile phone significantly improves the accessibility and acceptability of the proposed application. There is a wide range of potential applications, from in-home self-screening and diagnosis to in-office front line screening and referrals. Leveraging mobile phones could democratize access to cardiovascular screening. Additionally, our models for AS and MR included mild disease and retained high performance characteristics compared to other published studies.

Nonetheless, the deep learning algorithms above show great promise and are indeed a significant improvement on current standards of care. Other studies combine spectrogram analysis of murmurs with ECG interpretation to improve AI algorithm performance for the detection of SHD (23). NLD software could further advance this important work.

In this study, we simply sought to prove that NLD variables could be mapped to echocardiogram data for specific conditions. While the models developed as part of this work could be used for the detection of structural heart diseases like AS and MR, larger studies are necessary to confirm these findings.

This study was limited to a single site with a relatively homogenous population. The demographic characteristics of our target populations (AS and MR) are similar to what is described elsewhere (24). Notwithstanding, external validity in different populations and contexts needs to be established. Further, the sample size was small as this first step was to establish feasibility. The Train-Test models should be viewed as a control against over-fitting and not conclusive evidence of reproducibility. Yet, we contend that the small number of subjects with outcome of interest understates the feasibility of the technology and future performance metrics could be greater. We did not meet target enrollment to develop a model for mitral regurgitation and while our post-hoc adapted MR model that included mild disease had promising performance, additional studies will be needed to establish feasibility. Inter-operator measurements in echocardiography and reading are an additional limitation but mitigated by our study protocol's addition of a second sonographer and cardiologist overread.

Our study establishes the feasibility of using auscultation recordings collected from unmodified mobile phones and analyzed using nonlinear dynamics to identify features subsequently mapped to predict gold standard echocardiogram findings among subjects with aortic stenosis. NLD models have the potential to improve the screening, diagnosis, and monitoring of VHD and overcome existing technical, cost, and logistical barriers that remain for existing standards of care and emerging deep learning algorithms.