Cardiovascular screening method using an acoustic sensor

Abstract

A device, system, and method of analyzing high-fidelity heart sounds including applying acoustic sensors to a chest wall, capturing heart sounds from the chest wall, filtering the sounds to form a recording, and comparing the recording to other high-fidelity heart sound recordings. The acoustic sensors may include digital stethoscopes or contact microphones and may be arranged on the chest in a precordial configuration targeting a cardiac apex, other auscultation points, or a combination thereof. The recording may include 5 to 10 cardiac cycles, occur at 4,000 Hz, and in 5 second to 10 second segment intervals.

Description

Atty Docket: WIDOWSAVEROOIPCTCARDIOVASCULAR SCREENING METHOD USING AN ACOUSTIC SENSOR CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to United States Provisional Application Serial Number 63 / 791,341 filed on April 18, 2025, United States Provisional Application Serial Number 63 / 795,277 filed on April 26, 2025, and United States Provisional Application Serial Number 63 / 846.678 filed on July 18, 2025. All three applications are hereby incorporated by reference herein in their entireties.FIELD

[0002] Embodiments herein relate to a device and methods for using a digital stethoscope in a non-emergency environment to detect, record, and analyze acoustic properties of a human heart.BACKGROUND

[0003] Over time, fat and other viscous, inflammatory deposits along artery walls may cause blockages or slow blood flow such that heart failure, heart attacks, and other heart health emergencies may be more likely. A widowmaker heart attack, also known as a left anterior descending myocardial infarction, is a severe type of heart attack that occurs when the left anterior descending artery, the main artery that supplies blood to the heart, becomes completely blocked. This heart attack has a high mortality rate, especially if not treated promptly. The existence and extent of such blocking materials in the heart are hard to determine using current technology except for invasive procedures such as nuclear dye-based tests, CAT scans, etc.. These tests are often performed in outpatient centers or hospitals and interpreted by cardiologists in emergency situations when preventive health options are limited. These tests often compare the results of the testing with other people of known heart health, they do not compare measurements of the same person taken over time. This approach may miss subtle but clinically meaningful changes in an individual's heart patterns.

[0004] The acoustic signatures of blood flow through the components of the heart such as valves, arteries, and veins vary based on the individual and their heart’s age, size, strength, and other heart health factors. However, some general characteristics transcend any one person's physical being. For example, the consistent presence of only routine lub dub noises often implies that all components of the heart and the blood flow it collects and pumps are in fine health. If noises occur between the lubs and dubs, the implications could be that a valve is not workingAtty Docket: WIDOWSAVEROOIPCTproperly. Murmurs, swishing, and swirling of blood as it flows through arteries and valves may imply that an artery or valve is at least partially blocked.

[0005] A noninvasive and reliable test is needed that may be performed in a nonclinical setting by a person of limited or no health care training to detect if a patient has a heart health risk that merits careful review. A consistent method for avoiding false positive indications of heart health risk is needed. Further, a test is needed to detect a diminished, swishy, or swirly bloodflow in the heart to improve the likelihood that a patient may have an earlier, broader list of preventative options.SUMMARY

[0006] This summary is provided to comply with 37 C.F.R. § 1.73, requiring a summary of the invention briefly indicating the nature and substance of the invention. It is submitted with the understanding that it will not be used to interpret or limit the scope or meaning of the claims.

[0007] Embodiments of the invention include a device, system, and method of analyzing high-fidelity heart sounds including applying acoustic sensors to a chest wall, capturing heart sounds from the chest wall, filtering the sounds to form a recording, and comparing the recording to other high-fidelity heart sound recordings. The acoustic sensors may include digital stethoscopes or contact microphones and may be arranged on the chest in a precordial configuration targeting a cardiac apex, other auscultation points, or a combination thereof. In some embodiments, sensor-skin contact is firm and undisturbed. The recording may include 5 to 10 cardiac cycles, occur at 4,000 Hz, and in 5 second to 10 second segment intervals. Some embodiments use color-coded feedback for indicating heart health risk.

[0007] Embodiments of the invention include a device, system, and method of analyzing high fidelity heart sounds including applying acoustic sensors to a chest wall, capturing heart sounds from chest wall, filtering and isolating the sounds to form a recording, and comparing that recording to reference heart sounds. Embodiments of the invention also include a device, system, and method of analyzing high-fidelity heart sounds of a person including applying acoustic sensors to a chest wall of the person, capturing heart sounds from chest wall; filtering and isolating the sounds to form a recording, and comparing that recording to the person’s historical recordings.Atty Docket: WIDOWSAVEROOIPCTDRAWINGS

[0008] Aspects, features, benefits and advantages of the embodiments described herein will be apparent with regard to the following description, appended claims, and accompanying drawings where:

[0009] Figure 1 is a workflow diagram of an embodiment of a cardiovascular screening method.

[0010] Figure 2 is workflow diagram of components of an embodiment.

[0011] Figure 3 is a schematic view of an embodiment of acoustic sensors placed on a person.

[0012] Figure 4 is an embodiment of a device that includes four stethoscopes.

[0013] Figure 5 is a cross-section schematic view of the chest wall with an embodiment of a device.

[0014] Figure 6 is a diagram of an embodiment of a signal preprocessing.

[0015] Figure 7 is a workflow of an embodiment.

[0016] Figure 8 is a plot of similarity scores as a function of session for an embodiment.

[0017] Figure 9 is a workflow diagram for an embodiment of a comparison of heart sounds.

[0018] Figure 10 is a confusion matrix summarizing classification performance of an embodiment.

[0019] Figure 11 is a plot of a true label as a function of predicted label for an embodiment.

[0020] Figure 12 is a plot of risk score as a function of actual label of an embodiment.

[0021] Figure 13 is a plot of probability density as a function of risk score of an embodiment.

[0022] Figure 14 is a side-by-side boxplot of a risk score distribution by class label of an embodiment.

[0023] Figure 15 is a curve comparing a true positive rate in comparison to a false positive rate.

[0024] Figure 16 is a calibration curve of the relationship between the fraction of positives to the mean predicted probability.

[0025] Figure 17a and 17b are confusion matrices that visualize the actual positive and negative identifications against the predicted positive and negative identifications.

[0026] Figure 18 is a chart that shows the sensitivity, specificity, and accuracy of an experimental result.Atty Docket: WIDOWSAVEROOIPCTDETAILED DESCRIPTION

[0027] This disclosure is not limited to the particular systems, devices and methods described, as these may vary. The terminology used in the description is for the purpose of describing the particular versions or embodiments only, and is not intended to limit the scope.

[0028] The embodiments of the invention described herein are intended for use by individuals without clinical training, such as consumers or wellness professionals, in a non-emergency and non-hospital environment. Embodiments herein may be used in scenarios including home health tracking, gym or fitness center wellness checks, and personal preventative screening. Unlike invasive tests such as angiography or nuclear scans that require interpretation by healthcare professionals, embodiments herein rely on placing sensitive acoustic sensors to capture heart sounds of sufficient quality for analysis and comparison. As used herein, “high-fidelity heart sounds” refers to heart sound recordings having a level of clarity and quality sufficient for further processing and assessment.

[0029] As used herein, the term “cardiac cycle” means the events that occur during one heartbeat, including but not limited to diastole, atrial systole, isovolumetric contraction, ventricular systole, isovolumetric relaxation, and ventricular diastole.

[0030] Embodiments herein relate to a workflow that captures and analyzes a person’s heart sound signals from the chest wall using an acoustic sensor, digital stethoscopes, or similar wearable devices; filters the sounds to form a recording; compares new recordings against reference recordings including personal baselines, personal historical recordings, and population heart sound database libraries; and provides simple, color-coded risk assessment and trend feedback. As used herein, “filtering” means the process of refining raw data by removing errors, reducing noise, and isolating relevant information for analysis. Several methods and devices may be used to accomplish filtering, the examples and description herein are not presented to provide a narrower definition of the activity than ordinary usage. The workflow outputs include indicators to encourage users to seek early follow-up when early undesirable patterns are observed. Such patterns may be indicators of potential heart problems such as arterial blockage or murmurs and the indicators may give users preventative health insights without needing a doctor’s office visit.

[0031] Heart sounds include several components with specific frequency profiles. The first heart sound (SI) ranges from 30 to 150 Hz, the second heart sound (S2) ranges from 200 to 280 Hz,Atty Docket: WIDOWSAVEROOIPCTand heart murmurs (pathological sounds) can extend up to 1 ,000 Hz. to ensure high-fidelity capture of all diagnostic and sub-diagnostic components, a bandwidth may be defined as 25 to 2,000 Hz. To meet Nyquist criteria and ensure high-quality signal resolution, the sampling rate is 4,000 Hz for some embodiments with a recording duration of 5 to 10 seconds per recording session. This duration may capture 5 to 10 cardiac cycles for signal processing and statistical analysis. To promote signal consistency across sessions, environmental assumptions are made. The user is at rest (seated or reclined), the environment is quiet (e.g., home, clinic room), sensorskin contact is firm and undisturbed, and low-level ambient noise (e.g., heating, ventilation, and air conditioning, HVAC) may be present but may be addressed during preprocessing using a bandpass filter.

[0032] The embodiments of the invention described herein may include: an acoustic interface to capture heart sounds H(t) using wearable sensors using a device or mobile app, a preprocessing unit to filter, segment, and normalize the signal using embedded or app-side deployment options, a processor which extracts features for scoring locally, remotely, or both, a similarity comparator to calculate the similarity score Sn and compare references locally or remotely, a trend and risk engine to classify risk and detect long-term changes that may be recorded in the cloud or using a mobile app, a visual feedback layer to present color badges, scores, and graphs using a user display or interface, a data management layer to log and manage session results using encrypted storage, and finally an alert, notification, or both to trigger soft alerts or follow-up prompts using an application (app) or message workflow.

[0033] Figures 1 and 2 provide a workflow of an embodiment of a cardiovascular screening workflow using acoustic sensors. Figure 1 shows actions performed during operation of an embodiment of a cardiovascular screening workflow. Figure 2 provides an end-to-end workflow architecture for one embodiment. A device with acoustic sensors is placed on the user’s chest and the user’s heart sounds are recorded 101. A recorded acoustic heart signal 201 is preprocessed 202 by filtering, denoising, and segmenting the signal 201 into cardiac cycles. One or more initial sessions may be used to establish a baseline 77o(t) 103 for the user. In some embodiments, the baseline may be retained, refined, or updated over time using one or more earlier or stored recordings obtained from the user. The heart sound signals undergo feature extraction 204, in which the signals are mapped to a feature vector Fn104. The signals are compared and analyzed such as via dual-path similarity scoring 105, in which a new signal Hs(k)(t) is compared to aAtty Docket: WIDOWSAVEROOIPCTuser’s baseline, one or more earlier recordings of the user, reference libraries of healthy and unhealthy heart sounds, or a combination thereof. The resulting comparison values may be combined to create a score. Based on this score, a risk classification 106 may be assigned that correlates to a green, yellow, or red label. Signals may be analyzed over time and assigned risk classifications and risk score outputs 205. Each signal may also be used for trend tracking 107 and analysis of risk evolution including trend tracking 207. A connected device may provide display outputs via a user interface 108 showing updated scores, labels, or trend charts. The workflow may issue an alert 109 to the user if one or more thresholds or patterns are met. For later use by the user or medical professionals, the signals and outputs may be logged 110, stored locally or in remote storage, synchronized with one or more external devices or workflows 206, or exported to a connected device 207.

[0034] As shown in Figure 2, obtaining the acoustic heart signal 201, signal preprocessing 202, feature extraction 204, trend tracking 207, and risk score output 205 may be performed within the device 208a, 208b, 208c, 208d. Actions such as storage, synchronization, or communication with one or more external devices or workflows 206, 207 may be performed outside of the device 209a, 209b.

[0035] Figure 3, Figure 4, and Figure 5 provide an embodiment of the device and placement on a user’s body. Figure 3 provides an embodiment of the device and acoustic sensors on the chest of a user. Figure 4 provides an embodiment of the device and sensors. Figure 5 provides a crosssection of a user’s chest cavity with the embodied device and sensors placed on the chest. A device 303 with 1 to 4 acoustic sensors 304a, 304b, 304c, 304d are arranged on the chest of a user 301 in a precordial configuration, targeting the cardiac apex or other auscultation points to listen to the heart 302. The sensors 304a, 304b, 304c, 304d are placed in these configurations to capture heart sounds more accurately due to tissues, bones such as the ribs 504a, 504b and sternum 503. and pleural 502 and pericardial 501 cavities which contain lubricating fluid for the heart 302.

[0036] The heart sound signal, denoted as H(t), is acquired via digital stethoscopes or contact microphones that are placed and positioned on the chest. The acoustic sensors record 5 seconds to 10 seconds to capture multiple cardiac cycles in resting conditions. The acoustic signal captured during a recorded segment is represented by H(t) where it is the time-domain amplitudeAtty Docket: WIDOWSAVER001PCTof the heart signal at time t. This signal is continuous in nature but recorded as discrete sequences sampled at fixed rates.

[0037] Heart sound signals H(t) undergo preprocessing to isolate relevant acoustic features and reduce the impact of noise and environmental artifacts prior to analysis. The cleaned-up and segmented heart signals form the basis for constructing a personalized baseline Ho(t), which serves as a reference in all subsequent similarity comparisons.

[0038] Figure 6 provides an embodiment of a signal preprocessing process for heart sound analysis. To eliminate low-frequency baseline drift and high-frequency artifacts such as muscle tremors and background hum found in the heart signal 601, a bandpass filter 602 is applied. A 4th-order zero-phase Butterworth filter with a passband ranging from 25 to 400 Hz may be used in some embodiments. This preserves heart sounds (SI, S2) and a significant portion of murmur information while reducing irrelevant components. For environments with higher ambient noise such as HVAC workflows, wavelet denoising may be applied. Discrete wavelet transform (DWT), applying thresholding on detail coefficients, and reconstructing cleaned signal from modified coefficients may be performed. Wavelet-based approaches enhance performance in low-SNR settings and retain transient features important for murmur detection. In signal segmentation 603, the heart sound components are located by computing a short-time energy envelope:5 -A— y 4- m > wiwin -0Where w[m] is a symmetric window and N is typically 200 to 400 samples. The peaks in E\n\ corresponds to SI and S2 events allowing for segmentation into cardiac cycles.

[0039] Some embodiments may include cardiac cycle segmentation and fixed-length window segmentation. Cardiac cycle segmentation is the preferred segmentation strategy where the Segment = H(tk) to H(tk+i) where tk corresponds to detected SI onsets. Fixed-length window segmentation is where the Segment Length = 2 seconds = 8,000 samples at 4,000 Hz. Both strategies yield a sequence of windows for feature extraction or averaging.

[0040] The baseline signal Ho(t) acts as a personalized reference curve that identifies the user’s unique heart acoustics. This mitigates the influence of anatomical variations between individuals and provides a stable anchor for longitudinal comparisons. The baseline may be establishedAtty Docket: WIDOWSAVER001PCTusing different approaches. In one approach, a single-recording baseline is used, in which Ho(t) is based on the segmented and fdtered signal from an initial session. This approach may be suitable for immediate-use workflows, although it may be more sensitive to outliers. In another approach, the baseline is formed from multiple recordings or segments obtained from one or more sessions. In this approach,jyWA- 4 " E Alwhere Hs(k)(t) is the k-th segmented cardiac cycle, and N is the number of cycles. Prior to averaging, the segments are amplitude-normalized, and are time-aligned using dynamic time warping or fiducial alignment on SI peaks. This smooths random fluctuations and yields a personal signature. At the end of preprocessing, each user has a heart sound signal Hf(t) with isolated acoustic features, a sequence of segments Hs(tk), and a baseline reference Ho(t) for longitudinal and comparative analysis.

[0041] After signal preprocessing and segmentation, heart sound recordings may be transformed into structured numerical forms that allow for comparison and classification. Informative features can be extracted from the time, frequency, and time -frequency domains. These features are aggregated into a unified feature vector Fn for each session n, forming the basis of similarity scoring and risk tracking.

[0042] Time-domain analysis captures characteristics of signal shape and energy. Such characteristics include the root mean square (RMS) energy which indicates signal intensity and power, the zero-crossing rate which measures frequency of sign changes in the signal and correlates with signal activity, the skewness and Kurtosis which capture the asymmetry and peak of the signal distribution over time, and the energy envelope which is derived from the Hilbert transform or squared magnitude, this reflects the beat structure.

[0043] Frequency domain features reveal patterns in the heart sound’s spectral composition using mel-frequency cepstral coefficients (MFCCs). MFCCs capture perceptual spectral characteristics using a mel scale filter bank. The features include the spectral centroid whichAtty Docket: WIDOWSAVEROOIPCTindicates the center of mass of the frequency spectrum, the spectral bandwidth which describes how spread the spectrum is around the centroid, the spectral roll-off which is the frequency below which a fixed percentage of total spectral energy lies, and finally the spectral entropy which quantifies the disorder or unpredictability in the spectrum.Cenuoid

[0044] For non-stationary components such as murmurs, a time-frequency analysis is performed. The short-term Fourier transform (STFT) may compute the time-evolving frequency content via sliding windows.STFTiX / ) V .T

[0045] Discrete wavelet transform (DWT) performs multiresolution analysis to detect features such as clicks, swishes, or murmurs. These two methods produce spectrograms or wavelet coefficient maps that may be further summarized into feature statistics. Clinical acoustic features are domain-specific features that capture medically relevant signal structures such as SI and S2 duration and amplitude, the presence and bandwidth of murmurs, and the temporal placement of murmurs (whether they are systolic or diastolic). These features provide interpretability and are aligned with auscultation criteria. For feature normalization and to ensure fair comparisons across sessions and users, amplitude normalization adjusts signals to unit energy or peak values, session-independent scaling applies min-max or z-score scaling using parameters derived from the reference dataset, and time alignment synchronizes features based on SI markers or dynamic time warping.

[0046] The output for each session n, the preprocessed signal Hn(t), is mapped to a feature vector where each fi represents a time, frequency, or clinical acoustic descriptor. The resulting vector is used in the similarity engine to compute risk-relevant scores.

[0047] Figure 7 provides an embodiment of a dual-path similarity scoring process for comparing and analyzing a user’s heart sounds. The preprocessed heart sounds 701 undergo feature extraction 708 and are compared and analyzed in a two-prong approach. In prong one, a personalAtty Docket: WIDOWSAVER001PCTlongitudinal comparison is conducted against the user’s personal baseline 702 and undergoes similarity scoring 704a. The user’s baseline (HO) may be established from one or more earlier heart sound recordings, and future recordings (Hn) are compared against HO to detect deviations that may indicate a change in the individual’s cardiovascular profile. In prong two, a comparison is conducted in parallel against a reference population database 703 and undergoes similarity scoring 704b. Each fresh recording of the user’s heart sound is compared against curated databases of healthy (Dh) and unhealthy (Du) heart sounds, representing a spectrum of acoustic profiles with identified heart characteristics. This external reference anchors the risk scoring framework. By combining both comparison results in a weighted scoring model to create a dualaxis score 705, the workflow reduces the likelihood of false positives while maintaining sensitivity to concerning trends. Users receive visual results — green, yellow, or red — along with a trend indicator that reflects long-term shifts in cardiovascular acoustics.

[0048] The embodiments herein use similarity scoring to compare a user’s fresh heart sound recording Hn(t) to two references. The new heart sound recording is compared to the user’s personal baseline H0(t), which may be formed from one or more earlier recordings or segments from the user, and to the population databases of healthy Dh and unhealthy Du heart sounds. This dual-axis strategy increases correct identification by anchoring detection to both the user’s own personal trends and to those of other individuals. The outcome is a continuous similarity score Sn G [0,1], which is used to classify cardiovascular wellness into intuitive, color-coded categories. For similarity score computation, the extracted feature vector for the nth session is selected to be Fn. The similarity between current features and the personal baseline is denoted by Sbaseline = sim(Fn, F0). The average similarity to healthy reference samples is denoted by Sh = avg sim(Fn, Dh). The average similarity to unhealthy reference samples is denoted by Su = avg sim(Fn, Du). To evaluate a library -based similarity:5V8?? 4-

[0049] A weighted fusion score is computed as Sil = Wbaseline x Sbaseline + Wlib X Slib where Wbaseline = 0.30 and wiib = 0.70. These weights reflect empirical tuning to reduce false positives.

[0050] Several similarity metrics are supported in the embodied workflow. The distance and similarity measure is cosine similarity:Atty Docket: WIDOWSAVER001PCT

[0051] The second measurement, the Euclidean distance (transformed to similarity):

[0052] The final measurement is the wavelet domain or dynamic time warping similarity is sensitive to small deformations in temporal alignment.

[0053] Each new heart recording Hnis routed to a comparison engine that evaluates its similarity to a healthy heart-sound database and an unhealthy heart-sound database. The engine outputs a similarity score Sn(0 < Sn< 1) indicating how closely the new heart sound matches healthy sounds relative to the reference unhealthy sounds of other individuals. To reduce the effect of anomalies or noise, a moving average filter over a set number of the previous sessions is used to smooth Sn. If using dynamic time warping or wavelet-based comparisons, one may normalize score scales before fusion. Early sessions may favor baseline similarity more heavily (50 / 50), then taper to 30 / 70 as more population data becomes available.

[0054] Embodiments herein provide longitudinal insight beyond single-session similarity scoring by tracking the progression of the user’s cardiovascular risk over time to complications. This is achieved by maintaining a time series of computed similarity scores {Si, S2, ..., Sn}, from which trends and volatility can be analyzed. A trend tracking module flags deterioration patterns in the heart sound recordings that may not cross a threshold in any single session, but indicate cardiovascular risk through consistent change. Let S = {Si, S2, ..., Sn} be the chronological sequence of similarity scores from all recorded sessions. Each Si E [0,1] is stored with a timestamp and corresponding risk label (green, yellow, or red).

[0055] Embodiments herein may contain three main techniques to analyze the evolution of cardiovascular risk. Linear regression applies a least-squares regression line to S. The slope [3 indicates trend direction. If the slope is greater than zero, the heart sound recordings are aligning more closely with the healthy references and that risk is improving. If the slope is less than zero, the heart sound recordings are becoming more unhealthy and the risk is worsening. The | | magnitude reflects rate of change. Moving average or exponential smoothing averages or bothAtty Docket: WIDOWSAVER001PCTare used to smooth short-term noise and highlight persistent shifts and exponential smoothing applies higher weights to recent scores.- | s . s,

[0056] Volatility index is variance or standard deviation over a fixed window. High volatility suggests inconsistent acoustic patterns; this may be due to arrhythmias or signal quality issues.<ys5™ < ~~ 5 . "">z

[0057] A risk label is assigned using predefined thresholds. The similarity score for healthier heart sounds and a low risk level > 0.70 has a visual indicator that is green. For moderate risk levels that may be used as early warning for a potential downturn in heart health, the similarity score is 0.50 < Sn < 0.70 and has a visual indicator that is yellow. Finally, the high-risk similarity score indicating the likelihood of unhealthy heart sounds is < 0.50 and has a visual indicator that is red.

[0058] Figure 8 provides the trend of similarity scores as a function of sessions for an embodiment. As shown, the X-axis indicates the session number 802, and the Y-axis indicates the similarity score Sn801. The risk band thresholds are represented by green (0.70) 804 and red (0.50) 805 lines. The points are color-coded by risk levels green (> 0.70) 806a, 806b, yellow (0.50 to0.70) 807a, 807b, 807c, and red (<0.50) 808a, 808b. The trend line 803 overlay shows the slope and moving averages. Flag markers may also be used to indicate sessions triggering alerts (e.g., red triangle for urgent flag).

[0059] The workflow embodiments herein classify cardiovascular risk into intuitive categories and define alert logic to prompt follow-up behaviors when patterns of concern emerge. The classification workflow is non-diagnostic, focusing on wellness guidance. It leverages the computed similarity score Sn. its recent history, and trend behavior to determine and display a corresponding color-coded risk status. The green indicator may be interpreted as “your heart sounds healthy,” the yellow indicator may be interpreted as “monitor your heart - changes detected,” and the red indicator may be interpreted as “your heart sounds at risk - consider follow-up.” This output simplifies interpretation and removes complex metrics. The visual feedback output layer displays the classification and trend data via color badges such as red, yellow, or green icons (LED indicators, mobile UI label), numerical score displays for similarityAtty Docket: WIDOWSAVEROOIPCTscores with “normal,” “monitor,” “at risk” labels, trend charts with line plots, including threshold and visual alerts with interactivity and session risk logs that show tabs of past sessions and their outcomes with capabilities to export the data.

[0060] To reduce false positives and support decision-making, alerts are only triggered under multi-criteria conditions. The alerts are generated using a combination of absolute score thresholds and relative trend changes. Potential alert conditions in some embodiments may include a persistent risk alert when there are two or more consecutive red sessions suggesting retesting or consultation. A gradual worsening trend may be alerted when the slope drops slightly but consistently over a number of sessions suggesting a flag long-term worsening pattern (-0.05 over 3+ sessions). Sudden drops in the change of the slope in a single step may trigger an acute risk deterioration alert that encourages retesting and self-check (-0.20 in 1 step). An unstable signal or soft flag indicates a high variance in the slope which warns the user to continue tracking. Another alert may occur when the score rebounds but the trend remains negative, informing the user to continue tracking. Alerts may be categorized as informational, warning, or urgent, based on severity. The alerts in some embodiments of the workflow may appear as banners, pop-ups, or notifications depending on the workflow interface (e.g., app, device screen, smartwatch).

[0061] Upon generating an alert, embodiments herein may take a variety of actions and make recommendations. The workflow may log the event with timestamp and type, display a contextual recommendation. Customizable logic allows for alerts to escalate in severity depending on frequency, trend velocity, or persistence of abnormal readings.

[0062] Embodiments discussed herein use tools and methods to preserve privacy and ensure data integrity. Data is timestamped, anonymized, and encrypted during local storage or cloud transmission. Application programming interfaces (APIs) may be used to interface with health apps or provider dashboards. Similarity scores and abstracted risk indicators may be retained for non-diagnostic use.

[0063] Figure 9 provides a workflow diagram for an embodiment of a comparison of heart sounds. The embodiments herein incorporate several steps to compare a person’s heart sounds to their historical recordings and to reference recordings of healthy and unhealthy individuals. The preprocessed heart signals recorded on the user are used for feature extraction 908. The informational features are extracted from the time domain (RMS, energy envelope), frequencyAtty Docket: WIDOWSAVEROOIPCTdomain (peak frequency, spectral entropy), and the time-frequency domain (STFT, DWT), with additional normalization and alignment. Features are plotted into feature vectors where initialized 901 to identify the weight for baseline (wi 0.3) and population (W2 0.7) similarity with the thresholds being green: >0.70, yellow: 0.50 to 0.70, red: <0.50. The similarity scores are computed 902 and compared to the baseline and population data. The baseline scores and population scores are compared separately and averaged. Scores are combined together to create a user’s similarity score (Sn) that is used to identify and assign risk indicators. The risk indicators assigned 903 to the Snare determined by the similarity thresholds; green: >0.70, yellow: 0.50 to 0.70, red: <0.50. The logs and trends are updated 904 with the new recordings, similarity scores, and risk indicators which are timestamped and used to create a moving trend. Robustness checks and alarms 905 are performed. If the risk label assigned to the heart sound is red for two or more consecutive readings, a soft trigger alert is issued. If the trend declines over a certain threshold over a singular reading, a soft trigger alert is issued. The output generated 906 in this workflow is the risk label assigned to the heart sounds, the similarity scores, the trend graph of the user’s heart sounds over time, and a timestamped record to aid in heart health.

[0064] Performance was validated in reducing false positive rates for acoustic cardiac screening using the CADScor system, a commercial device available from Acarix company of Malmo, Sweden as the primary benchmarking tool. The implementation followed a framework that is used in embodiments of the workflow to compare a person’s acoustic heart signals to their historical recordings and the recordings of references. The results demonstrate a substantial improvement in specificity and a notable decrease in false positive rate compared to the CADScor system, confirming the efficacy and clinical relevance of the signal-level fusion approach.

[0065] An aim was to benchmark its false positive (FP) rate against that of existing commercial devices, notably the CADScor system. Prior internal evaluations in milestones 2-4 had indicated strong sensitivity and significantly reduced FP rates; this validation against reference heart sounds as an external proof-of-concept.

[0066] The PhysioNet / CinC 2016 dataset, a dataset commercially available from the MIT Laboratory for Computational Physiology of Cambridge, Massachusetts was selected as the primary dataset due to its extensive annotations and diversity of healthy and unhealthy heartAtty Docket: WIDOWSAVEROOIPCTsounds. After noise filtering and segmentation, recordings were processed into clean cardiac cycles using envelope-based beat detection.

[0067] To suppress inter-subject variability and highlight pathological deviations, we implemented a beat-level subtraction technique: template whitening. For each recording we derive a subject-specific template Ho(t) by averaging the 5-10 cleanest beats. Subsequent beats Hn(t) are converted to residuals AHn(t) = Hn(t) - Ho(t). Feature vectors from AH feed a dual-axis fusion: Risk = 0.30 x Similarity_to_Baseline + 0.70 x Library_Match.

[0068] The workflow computes two similarity scores for each beat: one against the user’s recorded heart sound baseline and another against a reference pathological library of healthy and unhealthy heart sounds. These are weighted to produce a final risk score per beat, then aggregated to assess overall risk for the users recording.

[0069] The model was evaluated on the full PhysioNet dataset and validated using 5-fold cross-validation. A small hold-out set was preserved to demonstrate generalizability. The key performance metrics are summarized below.Metric Classification Metrics CADScor (MIB 174 Report)False Positive Rate 18.7% [95% CI: 16.2 to 21.3%] -47%Specificity 81.3% [95% CI: 78.7 to 83.8%] -53%Sensitivity 88.5% [95% CI: 86.0 to 90.4%] Not reportedAccuracy 84.2% Not reportedTable I. Classification Metrics vs. CADScor Benchmark

[0070] Table I provides a comparative analysis of core classification metrics between the embodiments of a workflow and the CADScor system, as reported in the MIB 174 evaluation. Notably, the false positive rate (FPR) of embodiments is significantly lower at 18.7percent (95percent CI: 16.2 to 21.3 percent), compared to an estimated 47 percent for CADScor. This reduction — over 60 percent improvement — addresses one of the most critical limitations of existing auscultation-based screening tools: high false alarm rates erode clinical trust with users and lead to unnecessary referrals to high-cost specialists.

[0071] Results also demonstrate a strong specificity of 81.3percent (95percent CI: 78.7 to 83.8 percent), considerably higher than CADScor's approximate 53 percent, further confirming its advantage in correctly identifying healthy individuals by comparing their heart sounds to the reference database and minimizing misdiagnosis. The workflow achieves this without sacrificingAtty Docket: WIDOWSAVEROOIPCTsensitivity, maintaining a high sensitivity of 88.5 percent (95 percent CT: 86.0 to 90.4 percent), even though no such sensitivity estimate is reported for CADScor. This balance contributes to an overall accuracy of 84.2 percent on the validation set. The ability to deliver high sensitivity while reducing false positives positions embodiments of the workflow discussed herein as a clinically viable and scalable screening tool.

[0072] Figure 10 provides a classification performance using a matrix. The confusion matrix shows a decision threshold of 0.6 111, summarizing the model's classification performance on the validation dataset. The matrix shows the comparison between the true label 112 of the case and the predicted label 113 made by the model. The model correctly identified 16 out of 29 normal cases 114 and 30 out of 64 abnormal cases 117. On the other hand, 13 normal cases were misclassified as abnormal 115, contributing to the overall false positive (FP) rate. However, compared to benchmarks such as CADScor (which has an FP rate of ~47percent), these results demonstrate improvement. The observed pattern supports the hypothesis that the dual-axis similarity approach, especially when combined with template whitening, enhances the model’s ability to differentiate pathological from non-pathological signals while reducing spurious alarms.

[0073] Figure 11 provides the sensitivity v. false positive rate relationship for an embodiment. The trade-off between sensitivity (recall) 119 and false positive rate (FPR) 118 across different classification thresholds is represented through a plot curve 120. Increasing sensitivity generally results in higher false positive rate, underscoring the inherent balance between detecting abnormal cases and avoiding false alarms. The curve 120 shows a smooth and monotonic trend, indicating stable threshold behavior and reliable calibration of the prediction scores. The model demonstrates sensitivity values exceeding 0.75 with a relatively moderate FPR (-0.25), suggesting a practical operating point for deployment where clinical safety (high recall) is preserved without overwhelming practitioners with false positives. This performance supports the use of a tunable threshold, allowing the algorithm to be adapted for either conservative (low FPR) or aggressive (high sensitivity) screening environments based on healthcare context or risk appetite.

[0074] Figure 12 provides the risk score distribution by ground truth classes of an embodiment. The distribution of risk scores 121 assigned by the model is shown across two actual label 122 classes: normal (label 0) 124 and abnormal (label 1) 123. The distinct gap between the twoAtty Docket: WIDOWSAVEROOIPCTgroups suggests that the risk scoring mechanism — based on template similarity or latent embeddings — is effective in capturing pathological characteristics. Normal cases cluster tightly around a higher score (~0.38-0.40), while abnormal cases are consistently assigned much lower values (~0.26), indicating that the model not only classifies but also quantifies abnormality in a way that aligns well with expert-labeled ground truth. The low variance and absence of overlap between classes also reinforce the model’s robustness in real-world screening scenarios.

[0075] Figure 13 provides the probability density distribution of risk scores by label of an embodiment. The probability density 131 distribution of risk scores 132 for normal 133 and abnormal 134 cases are compared. The plot shows that normal samples 133 are generally assigned higher risk scores, whereas abnormal cases 134 cluster toward the lower end. Although the two curves exhibit partially overlapping tails, the central tendencies remain distinguishable — pointing to the model’s capacity to learn class-discriminative representations. The presence of mild overlap around the 0.57 to 0.62 range balances the trade-off between false positives and false negatives near the decision threshold. The overall spread and distributional shift between the two groups reinforce the reliability of the score as a soft indicator of abnormality.

[0076] Figure 14 provides the risk score 141 distribution by class label 142 for an embodiment using a side-by-side boxplot. Unlike Figure 13, which showed strong score separation in a prior experiment, this view highlights a case where score distributions for normal 143 and abnormal 144 cases largely overlap, including nearly identical medians 145a, 145b. While the interquartile ranges 146a, 146b (IQR) suggest comparable spread across classes, the minimal separation between the central tendencies may indicate less discriminative model behavior for this particular validation fold or data subset. This may be due to class imbalance, underfitting, or subtle signal differences in certain recordings. Despite this overlap, the model still retains utility for risk stratification in borderline cases, especially when combined with other clinical variables. However, it also emphasizes the importance of score threshold tuning, ensemble validation, and cross-cohort benchmarking before deployment.

[0077] This case study validates the effectiveness of the dual-axis similarity algorithm on the PhysioNet Challenge dataset when comparing and identifying a person’s recorded heart acoustic signature, demonstrating its capability to significantly reduce false positive rates compared to established benchmarks such as the CADScor system. The combination of template whitening,Atty Docket: WIDOWSAVEROOIPCTbeat-level similarity analysis, and risk-based scoring contributes to a more interpretable and generalizable classification pipeline.

[0078] An ablation study was performed on an embodiment of the device and workflow algorithm. In connection with the study, test-set inputs, reproducibility controls, and supporting functions were established. Dataset paths were defined, and checks were performed to confirm that label files, reference files, and WAV-formatted audio files were identifiable and accessible. This allowed the dataset to be evaluated for data quality, correctness, and consistency. The reference files were read and reviewed, and file parsing was validated. A data structure was created containing file identifiers, locations of WAV files, actual labels, and subject identifiers. A fold manifest was generated to assign training, validation, and test splits for each fold. A prediction results structure was also created to store model outputs for later analysis. Directories and candidate files were identified from one or more input sources or data repositories.

[0079] Candidate model files were identified from one or more input sources or data repositories. Consistency checks were performed on the initial dataset. An operating-point evaluation was then conducted using previously generated prediction results to compute a confusion matrix and other performance measures at selected threshold values. Audio features were extracted and stored for training and evaluation. A logistic regression model was trained for each partition of a cross-validation procedure, and the prediction results were updated with computed risk scores. Performance measures were calculated across multiple thresholds and at an operating point of 0.85. Additional evaluation included generation of figures, per-partition confusion matrices, a calibration curve, and Brier scores. Per-partition performance measures were also generated with 95 percent confidence intervals calculated using the Wilson score method, using the corresponding test partition associated with the prediction results. The ablation study further evaluated operation without residual-based features on the test partition only. In a subsequent configuration, features were omitted across all partitions, and a model with imputation was trained on one subset of partitions and evaluated on a different partition.

[0080] Figure 15 is receiver operating characteristic curve, an ROC curve, that visualizes the performance of this binary classification regime. The graph plots the true positive rate (sensitivity) against the false positive rate across various thresholds. The area under the curve (AUC) quantifies the model’s overall discriminative performance. An AUC of 1.0 representsAtty Docket: WIDOWSAVEROOIPCTperfect discrimination, while an AUC of 0.5 represents random classification performance. The AUC for Figure 15 is 0.932, indicating strong classification performance.

[0081] Figure 16 is a calibration curve that visualizes the devices response against the predicted probabilities. This curve shows the relationship between the fraction of positives to the mean predicted probability. The calibration curve generally follows the line that indicates perfect responses showing that the device tends to make positive predictions close to the mean predicted probability.

[0082] Figure 17a and 17b are confusion matrices that visualize actual positive and negative identifications against predicted positive and negative identifications. Figure 17a is a confusion matrix for a configuration in which residual-based features were included in the analysis (“On_residual”). Figure 17b is a confusion matrix for a configuration in which raw features were used without the residual-based feature configuration (“OFF_raw”). In both matrices, the alignment between actual positive cases and predicted positive cases reflects the ability of embodiments of the invention to identify positive cases. Figure 17a further shows alignment between actual negative cases and predicted negative cases, indicating improved discrimination of negative cases under the residual-based configuration. By contrast, Figure 17b shows a configuration in which predictions were predominantly or exclusively positive, indicating that the configuration was less effective at distinguishing negative cases from positive cases. This comparison demonstrates that inclusion of the residual-based feature configuration may improve classification balance and reduce overprediction of positive cases.

[0083] Figure 18 is a chart showing sensitivity, specificity, and accuracy for fold 4, together with 95 percent confidence intervals across folds, at a threshold of 0.85. The On_residual configuration showed lower sensitivity than the OFF_raw configuration, which exhibited very high sensitivity. However, the On_residual configuration showed high specificity, whereas the OFF_raw configuration showed zero specificity. In this context, zero specificity indicates that negative cases were not correctly identified under the OFF_raw configuration, reflecting poor discrimination of negative cases and a high false positive tendency. The On_residual configuration also showed higher overall accuracy than the OFF_raw configuration.Comprehensive evaluation across multiple figures, including confusion matrices, ROC behavior, and risk score distributions, supports the discriminative capability of the algorithm acrossAtty Docket: WIDOWSAVEROOIPCTvalidation folds. In particular, reduction of false alarms while maintaining useful sensitivity supports the practical value of the disclosed screening approach.

[0084] These findings substantiate the technical foundation of embodiments of a workflow and reinforce its potential utility in low-cost, non-invasive acoustic heart signal screening — especially for deployment in wellness or resource-constrained healthcare settings. By offering improved specificity and interpretability, this approach may allow patients to obtain information about their heart health individually or at the primary care level, thus helping to alleviate unnecessary referrals and reduce healthcare burdens.

Claims

Atty Docket: WIDOWSAVEROOIPCTCLAIMS1. A method of analyzing high-fidelity heart sounds, comprising:applying acoustic sensors to a chest wall;capturing heart sounds from the chest wall;filtering the sounds to form a recording; andcomparing the recording to other high-fidelity heart sound recordings.

2. The method of claim 1, wherein the heart sound recording is at 4,000 Hz and in 5 second to 10 second segments.

3. The method of claim 1, wherein the acoustic sensors comprise digital stethoscopes, contact microphones, or both.

4. The method of claim 1, wherein the acoustic sensors are arranged on the chest in a precordial configuration targeting a cardiac apex, other auscultation points, or a combination thereof.

5. The method of claim 1, wherein sensor- skin contact is firm and undisturbed.

6. The method of claim 1, further comprising using color-coded feedback for indicating heart health risk.

7. A method of analyzing high-fidelity heart sounds, comprising:applying acoustic sensors to a chest wall;capturing heart sounds from chest wall;filtering and isolating the sounds to form a recording; andcomparing that recording to reference heart sounds.

8. The method of claim 7, wherein the recording comprises 5 to 10 cardiac cycles.

9. The method of claim 7, wherein the recording is at 4,000 Hz and in 5 second to 10 second segment intervals.

10. The method of claim 7, wherein the acoustic sensors comprise digital stethoscopes, contact microphones, or both.

11. The method of claim 7, wherein the acoustic sensors are arranged on the chest in a precordial configuration targeting a cardiac apex, other auscultation points, or a combination thereof.

12. The method of claim 7, wherein sensor-skin contact is firm and undisturbed.

13. A method of analyzing high-fidelity heart sounds of a person, comprising;applying acoustic sensors to a chest wall of the person;Atty Docket: WIDOWSAVEROOIPCTcapturing heart sounds from chest wall;filtering and isolating the sounds to form a recording; andcomparing that recording to the person’s historical recordings.

14. The method of claim 13, wherein a duration of the recording comprises 5 to 10 cardiac cycles.

15. The method of claim 13, wherein a heart sound recording is at 4,000 Hz and in 5 second to 10 second segment intervals.

Images