Using machine learning to calculate cardiac risk factor based on blood pressure measurements

Abstract

Methods and systems for estimating a left ventricular mass index (LVMI) value of a subject using ambulatory blood pressure of the subject and without need for transthoracic echocardiography & methods for treating LVH based thereon.

Description

USING MACHINE LEARNING TO CALCULATE CARDIAC RISK FACTOR BASED ON BLOOD PRESSURE MEASUREMENTS CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 668,400,filed July 8, 2024 and U.S. Provisional Application No. 63 / 777,346, filed March 25, 2025, the contents of which are hereby incorporated by reference.

[0002] Throughout this application, various publications are referenced, including referencedin parenthesis. The disclosures of all publications mentioned in this application in their entireties are hereby incorporated by reference into this application in order to provide additional description of the art to which this invention pertains and of the features in the art which can be employed with this invention. BACKGROUND OF THE INVENTION

[0003] Left ventricular hypertrophy (LVH) is a strong predictor of adverse cardiac outcomessuch as heart failure and arrhythmic events1–3, with concentric LVH being a LVH disease phenotype with an increased risk of cardiovascular outcomes4. LVH, determined through the measurement of elevated left ventricular mass index (LVMI)5, develops primarily in response to an increase in left ventricular wall stress caused by elevated blood pressure (BP)1,6.This increase in BP may occur in response to an arterial tree system with increased peripheral resistance or decreased artery compliance7.

[0004] It is well-known that pressure-only parameters (such as systolic pressure, diastolicpressure, and pulse pressure) are useful for assessing cardiac outcomes8–11. Nevertheless, studies have shown that for assessing LVMI and diagnosing LVH, the BP waveform itself (in addition to resting BP measurements) also has prognostic utility12–15. For example, one study showed that central systolic BP as assessed through pressure waveforms was superior to discrete clinic and ambulatory BP measurements for predicting LVH12. Measures derived from the BP waveform, such as wave reflection (amplitude and duration), may reflect development of LVH, and pulse wave analysis (PWA) of the BP waveform may help assess LV overload (and thereby 1 4901-7498-1200v.1corresponding increased afterload), which cannot be evaluated by conventional cuff pressure alone13.

[0005] Currently, measurement of LVMI primarily occurs with transthoracicechocardiography, using end-diastolic measurements of the left ventricle (via the Devereaux Method16as used in this study, among other methods with differing underlying geometric assumptions17,18). This current practice presents two important limitations. First, echocardiography takes place in clinics, hence limiting accessibility of LVMI measurements and LVH risk stratification for tens of millions of hypertensive patients and the broader population at risk of hypertension. Similarly, recent approaches which have used image-based deep learning techniques (based on retinal scan19, cardiac MRI20or echocardiography video analysis21) require collection of the images in a clinic setting (see Supplementary Tables 1 and 4 for a more detailed comparison). If diagnosed in a timely manner, LVH can be treated effectively with appropriate antihypertensive medications24. A second limitation of current practice is that information in time-series data are not used. Time-series electrocardiography (ECG) has been used to estimate LVMI25,26,but the devices (e.g.12-lead ECGs) require application in a clinic27. SUMMARY OF THE INVENTION

[0006] A method of estimating a left ventricular mass index (LVMI) value of a subjectcomprises steps of: obtaining pressure information associated with ambulatory blood pressure of the subject; obtaining tabular data associated with the subject; and processing the pressure information and the tabular data using a first machine learning algorithm trained using a first training set and providing a first ventricular mass index value as an output.

[0007] A method of diagnosing a left ventricular hypertrophy in a subjectcomprising determining a LVMI of a subject’s heart by a method described herein.

[0008] A method of treating a left ventricular hypertrophy (LVH) in a subject comprisingidentifying, or having identified, the subject as having LVH by a method described herein and administering to a subject so-identified an anti-hypertensive therapeutic.

[0009] A system for performing a method described herein, comprising:an ambulatory blood pressure measuring device configured to provide the pressure information associated with the subject; 2 4901-7498-1200v.1a computer system operably connected to the ambulatory blood pressure measuring device, wherein the ambulatory blood pressure measuring device is connected to the computer system via a wired or wireless connection to provide the pressure information to the computer system, wherein the computer system includes: a processor; and memory operably connected to the processor and including processor executable code, that, when executed by the processor, execute the steps set forth in a method described herein to provide the left ventricular mass index (LVMI) value.

[0010] A system for estimating a left ventricular mass index (LVMI) value of a subjectcomprising: an ambulatory blood pressure measuring device configured to provide pressure information associated with ambulatory blood pressure of the subject; a processor operably connected to the ambulatory blood pressure measuring device and configured to receive the pressure information; memory, operably connected to the processor and which includes tabular data associated with the subject, and processor executable code that, when executed by the processor, performs steps of: obtaining the pressure information; obtaining the tabular data; and processing the pressure information and the tabular data using a first machine learning algorithm trained using a first training set and providing a first ventricular mass index value as an output.

[0011] A system of diagnosing a left ventricular hypertrophy in a subject by determining aLVMI of a subject’s heart using a system described herein. 3 4901-7498-1200v.1BRIEF DESCRIPTION OF THE DRAWINGS

[0012] FIG. 1: Overview of method for predicting LVMI. This LVMI prediction and LVHclassification tool uses ambulatory BP measurements and / or waveforms, demographics and laboratory test results.

[0013] FIGS. 2A-2C: Overview of LVMI prediction and LVH classification architectures. a.Three sets of prediction / classification pipelines are shown, comprising models which ingest BP waveforms combined with tabular data of 37 variables, tabular data of 37 variables, or tabular data of 7 variables. BP waveforms were converted into a spectrogram before being fed into a 1- dimensional CNN-LSTM network. For tabular data of 37 variables, demographic data, laboratory test results (collected one time), and blood pressure readings were fed into a multi-layer perceptron (MLP) model. For tabular data of 7 variables, a subset of demographics data along with blood pressure readings were fed into another MLP. A first prediction / classification pipeline (top row, numbered 1), which predicts LVMI and classifies LVH and concentric LVH, concatenates CNN- LSTM embeddings with full tabular data MLP embeddings. A second prediction / classification pipeline (middle row, numbered 2), which predicts LVMI and classifies LVH and concentric LVH, uses full tabular data MLP embeddings. A third prediction pipeline (bottom row, numbered 3), which predicts LVMI, uses embeddings derived from a subset of tabular data. b. The transfer learning pipeline pre-trained a MLP model using a subset of FHS tabular data. The weights of the pre-trained MLP model were frozen and transferred to the third prediction pipeline, which fine- tuned the MLP on a subset of MHTN data, to predict LVMI. c. For multi-task learning, the MLP models for predicting LVMI in the first and second pipelines were replaced. These MLP models produce as outputs, in addition to LVMI, the end-diastolic echocardiographic measures IVSd, LVIDd, and PWTd (which are used to calculate LVMI using the Devereux equation).

[0014] FIGS 3A-3F: Classification of left ventricular hypertrophy (LVH) and concentric LVHusing stratified 10-fold cross-validation with different inputs. ROC curve for LVH classification with just the tabular data (A) and tabular plus time series (B) using the MHTN dataset. ROC curve for LVH classification with just the tabular data (C) and tabular plus time series (D) using the MHTN and IDH dataset. ROC curve for concentric LVH classification with just the tabular data (E) and tabular plus time series (F) using the MHTN and IDH dataset. In C, D, E, and F, we pooled IDH data with MHTN since the number of LVH cases were low. 4 4901-7498-1200v.1

[0015] FIG. 4: Comparison of multitask learning and transfer learning approaches forambulatory use. With the multitask learning approach, the study points to the feasibility of a continuously worn cuff for estimating LVMI or classifying LVH using discrete BP readings and waveforms, which is most accurate. The other feasible use case involves transfer learning, where a cuff is periodically worn for discrete BP readings; this approach produces slightly less accurate estimates of LVMI, but makes use of wearable monitors which are currently more widely available.

[0016] FIG. 5: Time series analysis: Blood pressure waveforms have different features andperiodicities, which are reflected in the frequency domain through a Fourier Transform. These spectrograms are input into the deep learning model.

[0017] FIG. 6: Left ventricular hypertrophy (LVH) phenotypes. LVH can be concentric, inwhich the left ventricular wall thickens, or eccentric, in which the left ventricular chamber is expanded.

[0018] FIG. 7 illustrates an exemplary block diagram of a system 100 for determining a leftventricular mass index (LVMI) value of a subject in accordance with some embodiments of the present application. DETAILED DESCRIPTION OF THE INVENTION

[0019] Methods and systems for estimating a left ventricular mass index (LVMI) value of asubject using ambulatory blood pressure of the subject and without need for transthoracic echocardiography & methods for identifying, diagnosing and / or treating LVH based thereon.

[0020] A method of estimating a left ventricular mass index (LVMI) value of a subjectcomprises steps of: obtaining pressure information associated with ambulatory blood pressure of the subject; obtaining tabular data associated with the subject; and processing the pressure information and the tabular data using a first machine learning algorithm trained using a first training set and providing a first ventricular mass index value as an output.

[0021] In some embodiments, the step of obtaining pressure information comprises:obtaining time series blood pressure waveform data associated with the subject; and converting the time series blood pressure waveform data into spectrogram data. 5 4901-7498-1200v.1

[0022] In some embodiments, the time series blood pressure waveform data corresponds to a10 second waveform at 100Hz. In some embodiments, the time series blood pressure waveform data corresponds to a 5-15 second waveform at 50Hz - 150Hz. In some embodiments, time series data can be collected at a series of discrete timepoints (predetermined or otherwise) over a selected period of time such as 0.5, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, or 24 hours; or 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30 or 31 days; or 1, 2, 3, 4, 5, 6, 7, 8, 9, or 10 weeks or any other preselected time period. Similarly, In some embodiments, the discrete timepoints can be 0.5, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, or 60 seconds; 0.5, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, or 60 minutes; or 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, or 24 hours.

[0023] In some embodiments, the pressure information is provided usingphotoplethysmography. Ambulatory blood pressure measuring devices are known in the art. Cuff- based devices include Spacelabs Healthcare ABPM 90217A or 90207, Welch Allyn ABPM 7100, Oscar 2 by SunTech Medical, Mobil-O-Graph NG (by I.E.M.), Schiller BR-102 plus, AND TM- 2430 by A&D Medical, Bosch + Sohn (boso TM-2430). In some embodiments, the ambulatory blood pressure measuring devices is cuff-based. In some embodiments, the ambulatory blood pressure measuring devices is cuffless. Cuffless devices are known in the art, e.g. see US 20210219852 A1, hereby incorporated by reference.

[0024] In some embodiments, the processing step includes providing the spectrogram data asan input to a convolutional neural network (CNN) Long Short-Term Memory Network trained using at least the first training set and providing a first output associated with the first ventricular mass index value, wherein the first ventricular mass index value is based on the first output.

[0025] In some embodiments, the convolutional neural network (CNN) Long Short-TermMemory Network uses a 1-dimensional convolutional neural network.

[0026] In some embodiments, the processing step includes:providing the spectrogram data as an input into a convolutional neural network; 6 4901-7498-1200v.1providing an output of the convolutional neural network as an input to a Long Short-Term Memory Network, wherein the first ventricular mass index value is based on a concatenation of the output of the Long Short-Term Memory Network and an output of a multi-layer perceptron mode using the tabular information as an input.

[0027] In some embodiments, the convolutional neural network (CNN) is a 1-dimensionalconvolutional neural network.

[0028] In some embodiments, the tabular information includes a first group of informationincluding demographic data and test results associated with the subject, and, optionally, wherein the first group of information has been obtained or provided prior to obtaining pressure information.

[0029] In some embodiments, the processing step includes providing, as inputs, the first outputand the first group of information to a first multi-layer perceptron model and concatenating the inputs to provide a second output associated with the first ventricular mass index value and other parameters related to the first ventricular mass index value.

[0030] In some embodiments, the other parameters include a left ventricular hypertrophyparameter and / or a concentric left ventricular hypertrophy parameter.

[0031] In some embodiments, the tabular information includes a second group of tabularinformation and the processing step includes providing the second group of tabular information to a second multi-layer perceptron model trained using a second training set and providing a second output associated with the first ventricular mass index value.

[0032] In some embodiments, the second multi-layer perceptron model is trained based on atleast a portion of Framingham Heart Study data.

[0033] In some embodiments, the Framingham Heart Study data comprises BP measurements,IVS, PWTD, LVIDD, LVMI, CO and LVH data.

[0034] In some embodiments, the Framingham Heart Study data further comprises tonometry,height, weight, Echo result, sex and / or age data.

[0035] In some embodiments, the demographic data includes age, sex, and / or body mass indexof the subject.

[0036] In some embodiments, the demographic data includes race, ethnicity, and / or history offamilial hypertension of the subject. 7 4901-7498-1200v.1

[0037] In some embodiments, the test results include blood HDL cholesterol, LDL cholesterol,triglycerides, glucose, and / or HbA1c of the subject.

[0038] In some embodiments, the test results include urine sodium, urine potassium, urinechloride, urine creatinine, and / or microalbumin creatinine ratio of the subject.

[0039] In some embodiments, the pressure information includes systolic blood pressure,diastolic blood pressure, pulse pressure, and / or heart rate of the subject.

[0040] A method of diagnosing a left ventricular hypertrophy in a subjectcomprising determining a LVMI of a subject’s heart by a method described herein.

[0041] In some embodiments, the LVMI is determined without using any transthoracicechocardiography.

[0042] In some embodiments, concentric LVH is diagnosed.

[0043] A method of treating a left ventricular hypertrophy (LVH) in a subject comprisingidentifying, or having identified, the subject as having LVH by a method described herein and administering to a subject so-identified an anti-hypertensive therapeutic.

[0044] In some embodiments, concentric LVH is treated.

[0045] In some embodiments, the anti-hypertensive therapeutic comprises: a diuretic, a beta-blocker, an angiotensin-converting enzyme (ACE) inhibitor, a calcium channel blocker, an alpha- adrenergic blocker, a renin inhibitor, a direct acting peripheral vasodilator, or a centrally acting agent. In some embodiments, the anti-hypertensive therapeutic comprises a thiazide diuretic or a beta-blocker.

[0046] In some embodiments, the subject has essential hypertension, renal artery stenosis,athletic heart with physiological LVH, aortic valvar stenosis, coarctation of an aorta, hypertrophic cardiomyopathy without or with outflow tract obstruction (HOCM), subaortic stenosis (left ventricular outflow tract obstruction by muscle or membrane), aortic regurgitation, mitral regurgitation, dilated cardiomyopathy, ventricular septal defect, or an infiltrative cardiac processes (amyloidosis, Fabry disease, Danon disease).

[0047] In some embodiments, the subject is male and LVH is present when the LVMI is > 102g / m2, normalized by body surface area of the subject. In some embodiments, the subject is male and LVH is present when LVMI is one of 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119 or 120 or greater g / m2. 8 4901-7498-1200v.1

[0048] In some embodiments, the subject is female and LVH is present when the LVMI is >88 g / m2, normalized by body surface area of the subject. In some embodiments, the subject is female and LVH is present when LVMI is one of 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, or 114 or greater g / m2.

[0049] In some embodiments, concentric hypertrophy is determined as LVH with a relativewall thickness > 0.42 cm.

[0050] In some embodiments, the method identifies early stage LVH.

[0051] A system for performing a method described herein, comprising:an ambulatory blood pressure measuring device configured to provide the pressure information associated with the subject; a computer system operably connected to the ambulatory blood pressure measuring device, wherein the ambulatory blood pressure measuring device is connected to the computer system via a wired or wireless connection to provide the pressure information to the computer system, wherein the computer system includes: a processor; and memory operably connected to the processor and including processor executable code, that, when executed by the processor, execute the steps set forth in a method described herein to provide the left ventricular mass index (LVMI) value.

[0052] In some embodiments, the ambulatory blood pressure measuring device is an upper armcuff-based ambulatory blood pressure monitor.

[0053] In some embodiments, the ambulatory blood pressure measuring device is cufflessambulatory blood pressure monitor.

[0054] In some embodiments, the ambulatory blood pressure measuring device is waterproofor water-resistant.

[0055] In some embodiments, the system comprises a digital recorder that records ambulatoryblood pressure associated parameter(s).

[0056] In some embodiments, the system is programmed to take blood pressure readings atregular intervals throughout the day and / or night.

[0057] In some embodiments, the methods or systems discriminates LVH cases with an AUCof 0.95 or better and / or discriminates concentric LVH cases with an AUC of 0.93 or better.

[0058] In some embodiments, the tabular data comprises 10 or less variables.9 4901-7498-1200v.1

[0059] In some embodiments, the tabular data comprises 7 variables.

[0060] In some embodiments, the tabular data comprises 37 variables.

[0061] A system for estimating a left ventricular mass index (LVMI) value of a subjectcomprising: an ambulatory blood pressure measuring device configured to provide pressure information associated with ambulatory blood pressure of the subject; a processor operably connected to the ambulatory blood pressure measuring device and configured to receive the pressure information; memory, operably connected to the processor and which includes tabular data associated with the subject, and processor executable code that, when executed by the processor, performs steps of: obtaining the pressure information; obtaining the tabular data; and processing the pressure information and the tabular data using a first machine learning algorithm trained using a first training set and providing a first ventricular mass index value as an output.

[0062] In some embodiments, wherein the step of obtaining pressure information comprises:obtaining time series blood pressure waveform data from the ambulatory blood pressure measuring device; and converting the time series blood pressure waveform data into spectrogram data.

[0063] In some embodiments, the time series blood pressure waveform data corresponds to a10 second waveform at 100Hz.

[0064] In some embodiments, the ambulatory blood pressure measuring device usesphotoplethysmography.

[0065] In some embodiments, the processing step includes providing the spectrogram data asan input to a convolutional neural network (CNN) Long Short-Term Memory Network trained using at least the first training set and providing a first output associated with the first ventricular mass index value, wherein the first ventricular mass index value is based on the first output.

[0066] In some embodiments, the convolutional neural network (CNN) Long Short-TermMemory Network uses a 1-dimensional convolutional neural network.

[0067] In some embodiments, the processing step includes:providing the spectrogram data as an input into a convolutional neural network; and 10 4901-7498-1200v.1providing an output of the convolutional neural network as an input to a Long Short-Term Memory Network, wherein the first ventricular mass index value is based on a concatenation of the output of the Long Short-Term Memory Network and an output of a multi-layer perceptron mode using the tabular information as an input.

[0068] In some embodiments, the convolutional neural network (CNN) is a 1-dimensionalconvolutional neural network.

[0069] In some embodiments, the tabular information includes a first group of informationincluding demographic data and test results associated with the subject.

[0070] In some embodiments, the processing step includes providing, as inputs, the first outputand the first group of information to a first multi-layer perceptron model and concatenating the inputs to provide a second output associated with the first ventricular mass index value and other parameters related to the first ventricular mass index value.

[0071] In some embodiments, the other parameters include a left ventricular hypertrophyparameter and / or a concentric left ventricular hypertrophy parameter.

[0072] In some embodiments, the tabular information includes a second group of tabularinformation and the processing step includes providing the second group of tabular information to a second multi-layer perceptron model trained using a second training set and providing a second output associated with the first ventricular mass index value.

[0073] In some embodiments, the second multi-layer perceptron model is trained based on atleast a portion of Framingham Heart Study data.

[0074] In some embodiments, the Framingham Heart Study data comprises BP measurements,IVS, PWTD, LVIDD, LVMI, CO and LVH data.

[0075] In some embodiments, the Framingham Heart Study data further comprises tonometry,height, weight, Echo result, sex and / or age data.

[0076] In some embodiments, the demographic data includes age, sex, and / or body mass indexof the subject.

[0077] In some embodiments, the demographic data includes race, ethnicity, and / or history offamilial hypertension of the subject.

[0078] In some embodiments, the test results include blood HDL cholesterol, LDL cholesterol,triglycerides, glucose, and / or HbA1c of the subject. 11 4901-7498-1200v.1

[0079] In some embodiments, the test results include urine sodium, urine potassium, urinechloride, urine creatinine, and / or microalbumin creatinine ratio of the subject.

[0080] In some embodiments, the pressure information includes systolic blood pressure,diastolic blood pressure, pulse pressure, and / or heart rate of the subject.

[0081] A system of diagnosing a left ventricular hypertrophy in a subject by determining aLVMI of a subject’s heart using a system described herein.

[0082] In some embodiments, the LVMI is determined without using any transthoracicechocardiography.

[0083] In some embodiments, concentric LVH is diagnosed.

[0084] FIG. 7 illustrates an exemplary block diagram of a system 100 for determining a leftventricular mass index (LVMI) value of a subject in accordance with some embodiments of the present application. In embodiments, an ambulatory blood pressure measuring device 12 my be provided to provide pressure information associated with ambulatory blood pressure of the subject. In some embodiments, the ambulatory blood pressure measuring device 12 may be an upper arm cuff-based ambulatory blood pressure monitor. In some embodiments, the ambulatory blood pressure measuring device 12 may be a cuffless ambulatory blood pressure monitor. In some embodiments, the ambulatory blood pressure measuring device 12 may take measurements at regular intervals for a predetermined period of time. In some embodiments, the ambulatory blood pressure measuring device 12 may use photoplethysmography to provide the pressure information. In some embodiments, the ambulatory blood pressure measuring device 12 may provide time series blood pressure waveform data.

[0085] In some embodiments, the pressure information may include time series blood pressurewaveform data from the ambulatory blood pressure measuring device. In some embodiments, the time series blood pressure waveform data may be generated based on the pressure information, for example, using the processor 10a. In some embodiments, the pressure information may include spectrogram data based on the time series blood pressure waveform data. In some embodiments, the spectrogram data may be provided based on conversion of the blood pressure waveform data by the processor 10a, for example.

[0086] In some embodiments, the ambulatory blood pressure measuring device 12 isconnected to a computer system 10 via either a wired or wireless connection to provide the pressure information thereto. In some embodiments, the computed system 10 may include a processor 10a 12 4901-7498-1200v.1operably connected to memory 10b. In embodiments, the computer system 10 mat be a personal computer, laptop computer, smart phone, or another mobile electronic device. In some embodiments, the computer system 10 may be a server or a node included as part of a communications network such as the internet. In some embodiments, the memory 10b may be provided adjacent to the processor 10a or may be positioned remotely and connected via a wire or wirelessly via a communication network. In some embodiments, the computer system 10 may include a transceiver to provide for receipt and transmission of data. In some embodiments, both the processor 10a and the memory 10 b may be positioned remotely from the ambulatory blood pressure measuring device 12.

[0087] In some embodiments, the memory 10b may include processor executable code thatwhen executed by the processor 10a, performs the steps of the method discussed above to provide a left ventricular mass index (LVMI) value. In some embodiments, the memory 12 may include tabular data related to the subject. In some embodiments, the tabular data may include demographic information and test results associated with the subject. In some embodiments, the tabular information may be retrieved from another memory operably connected to the computer system 10. In some embodiments, the processor may be used to implement machine learning algorithms to provide the left ventricular mass index (LVMI) value based at least on the pressure information and the tabular data. In some embodiments, the memory 12 may include processor executable code that when executed by the processor implements a machine learning module or modules that may be used to implement one or more machine learning algorithms to provide the left ventricular mass index (LVMI) value. In some embodiments, the machine learning algorithms may include one or more of a convolutional neural network, a Long Short-Term Memory Network model and one or more multi-layer perceptron models.

[0088] In some embodiments, the pressure information my be stored in memory 10b or inanother memory. In some embodiments, the left ventricular mass index (LVMI) value my be stored in the memory 10b or in another memory. In some embodiments, the left ventricular mass index (LVMI) value is used to diagnose left ventricular hypertrophy. In some embodiments, diagnosis information may be stored in memory 10b or in another memory. In some embodiments, the pressure information, tabular data and / or left ventricular mass index (LVMI) value may be displayed to a user via an electronic display operably connected to the computer system 10. 13 4901-7498-1200v.1

[0089] Left ventricular hypertrophy (LVH), clinically defined as an elevated left ventricularmass index (LVMI), is a strong predictor of adverse cardiac outcomes1,2. Currently, measurements of LVMI are performed using transthoracic echocardiography, accessible only in clinical settings. In this study, we use data from an ambulatory blood pressure monitor, combined with laboratory test results and demographics data collected once, to investigate whether LVMI and LVH may be predicted and classified accurately. We developed and tested machine learning models using clinical data. For LVMI prediction, a mean absolute error of 10.4 ± 1.5 with our test set, and 15.6 ± 1.2 with an external test set was achieved. For classification of LVH and concentric LVH, mean areas under the curve of ≥ 0.95 and ≥ 0.93 were achieved respectively. This study presents a method for accurately assessing cardiovascular risk using ambulatory measurements and highlights the use of waveform data for assessing hypertrophy.

[0090] In this disclosure, it was investigated whether LVMI can be estimated, and LVHclassified accurately, using ABP monitoring that can be performed at home, combined with one- time standard laboratory test results and demographics data. The variables used in our analysis also included waveforms. We employ two machine-learning models. First, we use multitask learning (a method in which multiple related output tasks are learned in parallel through a shared representation, with the goal of improving prediction accuracy and generalization28) with BP waveforms trained with echocardiography variables used in the calculation of LVMI. Second, we employ transfer learning (a machine learning method for applying knowledge derived from training a task in one domain to another similar domain, towards the goal of improving learning efficiency). Transfer learning has been used in biomedical signal processing29–31and time-series applications32, and is especially useful for cases with scarcity of data.

[0091] Left ventricular hypertrophy (LVH) is a strong predictor of adverse cardiac outcomessuch as heart failure and arrhythmic events1–3, with concentric LVH being a LVH disease phenotype with an increased risk of cardiovascular outcomes4. LVH, determined through the measurement of elevated ventricular mass index (LVMI)5, develops primarily in response to an increase in left ventricular wall stress caused by elevated blood pressure (BP)1,6. This increase in BP may occur in response to an arterial tree system with increased peripheral resistance or decreased artery compliance7.

[0092] It is well-known that pressure-only parameters (such as systolic pressure, diastolicpressure, and pulse pressure) are useful for assessing cardiac outcomes8–11. Nevertheless, studies 14 4901-7498-1200v.1have shown that for assessing LVMI and diagnosing LVH, the BP waveform itself (in addition to resting BP measurements) also has prognostic utility12–15. For example, one study showed that central systolic BP as assessed through pressure waveforms was superior to discrete clinic and ambulatory BP measurements for predicting LVH12. Measures derived from the BP waveform, such as wave reflection (amplitude and duration), may reflect development of LVH, and pulse wave analysis (PWA) of the BP waveform may help assess LV overload (and thereby corresponding increased afterload), which cannot be evaluated by conventional cuff pressure alone13.

[0093] Currently, measurement of LVMI primarily occurs with transthoracicechocardiography, using end-diastolic measurements of the left ventricle (via the Devereaux Method16as used in this study, among other methods with differing underlying geometric assumptions17,18). This current practice presents two important limitations. First, echocardiography takes place in clinics, hence limiting accessibility of LVMI measurements and LVH risk stratification for tens of millions of hypertensive patients and the broader population at risk of hypertension. Similarly, recent approaches which have used image-based deep learning techniques (based on retinal scan19, cardiac MRI20or echocardiography video analysis21) require collection of the images in a clinic setting (see Supplementary Tables 1 and 4 for a more detailed comparison). By comparison, ambulatory blood pressure measurements (ABP), which can be taken outside of a clinic, are closely correlated to organ damage and LVMI, potentially more so than clinic BP22,23. If diagnosed in a timely manner, LVH can be treated effectively with appropriate antihypertensive medications24. A second limitation of current practice is that information in time-series data are not used. Time-series electrocardiography (ECG) has been used to estimate LVMI25,26,but the devices (e.g. 12-lead ECGs) require application in a clinic27. By comparison, devices collecting ABP can collect BP waveforms over time. 15 4901-7498-1200v.197Tla37Cr-7P- ue9An5-l352an942oi9tulovnocasiN N C.noitacifissalc H V Lrofsehca .ornopprtapelacrnoepit rateyupalitmoluc m fo sin P osL irMap,yrmooCme:1melmrbaet-Ttryorh atsnegnmeollspipuM S T1.Sv0L0,21]k -4r8 99o 40w 7- 0t[en 610194.noitciderpIMVLrofsehcaorppalanoitatupmocfonosirapmoC:4elbaTyratnemelppu1.Sv0021-]8694907- 10[7019493597-7379 92425-A-PCT

[0097] In this study, we investigate whether LVMI can be estimated, and LVH classifiedaccurately, using ABP monitoring that can be performed at home, combined with one-time standard laboratory test results and demographics data. The variables used in our analysis also included waveforms. We employ two machine-learning models. First, we use multitask learning (a method in which multiple related output tasks are learned in parallel through a shared representation, with the goal of improving prediction accuracy and generalization28) with BP waveforms trained with echocardiography variables used in the calculation of LVMI. Second, we employ transfer learning (a machine learning method for applying knowledge derived from training a task in one domain to another similar domain, towards the goal of improving learning efficiency). Transfer learning has been used in biomedical signal processing29–31and time-series applications32, and is especially useful for cases with scarcity of data.

[0098] Results

[0099] We investigated deep-learning models for predicting LVMI and LVH usingmeasurements from non-invasive sensors. Measurements from non-invasive sensors could include BP values and waveform measurements as measured by a BP cuff, alongside demographic information and laboratory test results collected one time (Figure 1).

[0100] In order to identify models that could accurately predict LVMI and LVH usingmeasurements from non-invasive sensors, we investigated different deep learning models (Table 1).

[0101] Table 1. (See below): Summary of all the tested models. Rows highlighted in lightgreen meet the criteria for accurate approximations using ambulatory data, and the row in dark green is the transfer learning case which uses fewer variables for predictions. Rows with ‘Potential as cardiac risk prediction tool’ checkmark warrant further investigation towards being potentially suitable for ambulatory cardiac risk assessment. ‘amb’ is ambulatory data, ‘clin’ is clinic measurement. ‘R’ is the row number, ‘C’ is the column number. Cases with ‘tab’ under the ‘Variables’ column use all 37 variables.

[0102] The variables used in the FHS transfer learning analysis are in Supplementary Table 2(see below). The FHS variables were accessed from the following files: 1. Tonometry: HR, SBP, DBP, PP TON3_1D.csv 18 4901-7498-1200v.12. Height (BMI), weight (BMI): E_EXAM_EX01_3_0086D.csv 3. Echo: T_ECHO_EX01_3_0042D.csv 4. Sex, age: VR_DATES_2014_A_0912D.csv The tabular variables used in the full MHTN analysis were: 1. Demographics / anthropometric: race, sex, ethnicity, age, BMI, height, weight, history of familial hypertension, BP medication, diabetes 2. BP waveform: cardiac index, cardiac output, diastolic pressure, systolic pressure, heart rate, reflection coefficient, peripheral resistance, arterial compliance, mean arterial pressure, time of waveform recording, beat volume, pulse pressure, inflection pressure, inflection time, systole time, backward wave pressure, forward wave pressure 3. Blood / urine test: HDL cholesterol, LDL cholesterol, triglycerides, glucose, HbA1c, urine sodium, urine potassium, urine chloride, urine creatinine, microalbumin creatinine ratio.

[0103] Cases with ‘sub-tab’ use 7 variables (as seen in Supplementary Table 2). Cases with‘tab + ts’ use all 37 tabular variables and waveforms (time series). MHTN is Masked Hypertension Study, FHS is Framingham Heart Study, IDH is Improving the Detection of Hypertension. Results are presented as mean ± standard deviation, MAE is mean absolute error, AUC is area under the curve. Pipeline from Figure 2 refers to the figure number and pipeline number (e.g. “A2” refers to Figure 2A, pipeline 2). FHS data is available at, for example, the word wide web at framinghamheartstudy.org / fhs-for-researchers / data-available-overview / 19 4901-7498-1200v.1l97Tco a o it3n7Cd-7P- ro aictlci✓ ✓ ✓ d9Aa5-5itenr3924etp2oks9etesM M M M M MstM Meste%T0%10%10%10% % % %0%1010101 0101 ehtll+aN foteT NT NT NT NT NT NT NT s H H H H H H H H ynriM M M M M M M MS aar%T0%90%90%90%90%90%90%909 H F m mu r l+g aarkrni+onrkr gniSe:ni uewaeonr1le nlon ietelwt an kenele lenissnikrolsaatbi lareadloels er eswrutl rfus1 T Mag Ber ateuenBen N m Nar.tv0021]4n-o8 0Iit91ksMc4i7- 0[a d T V Ler10p0294✓ 8.00±5.90:CU A A,3giF2 2 1 1 2 1 A A A A A A )) )) )b b)b b bm mbmmaama(a( maa((g(gmrmr(gmrnidnioadfeof nea ve iofavadevareaPrPwPwePrPwP B B B B B B stststb b+b+ +atatatbatbatbat++ + +N N N N NT T N T T T H TH H H H M HM ()MM()M()M()%0%H% %H%H H 101DI 0101DI 01DI%01DI++ + +N N N N NT T N T T T H TH H H H M HM ()MM()M()M()%0%0H%0%0H%0H%H 9 9DI9 9DI9DI 09DIlalraureuneenniklr eeoniklrosawteesawt 1.B n Benv0no n02itacioi 1-crit tfna8 c94Hisecifi 7-Vsan Hsosa10LlcC V Llc1294

[0105] Supplementary Table 2: Summary statistics for the Framingham Heart Study (FHS)and MHTN datasets. All values are mean ± standard deviation or number of samples with percentages. BP waveform variables are taken from tonometry from FHS. *Left ventricular hypertrophy (LVH) was defined by: LVMI ≡ LVM / BSA >102g / m2for males and >88 g / m2for females. IVS is intraventricular septum, PWTd is posterior wall thickness at end-diastole, LVIDd is left ventricular internal dimension at end-diastole, CO is cardiac output.

[0106] We investigated different models for predicting LVMI, LVH, and concentric LVH. Forthe first prediction / classification pipeline (Figure 2a, top row), each 10-second waveform was converted into a spectrogram, and then processed through a CNN and then 3 LSTM layers. In 22 4901-7498-1200v.1addition, tabular data, which included a one-time set of laboratory test results, demographics, and blood pressure, were fed into a multi-layer perceptron (MLP). Each dense layer included batch normalization, dropout, and a rectified linear unit (ReLU) activation function (additional information on model parameters and size are provided in Fig.6). The pipeline then concatenated the embeddings from CNN / LSTM (i.e. time-series waveform pipeline) and tabular data, and predicted LVMI and classified LVH and concentric LVH. A second prediction / classification pipeline (Figure 2a, middle row) used embeddings from only the tabular data to predict LVMI and classify LVH and concentric LVH. A third prediction pipeline (Figure 2a, bottom row) used embeddings from only a subset of tabular data to predict LVMI and classify LVH and concentric LVH.

[0107] Next, for transfer learning, we pre-trained an MLP model on FHS data containing asubset of variables that overlapped with MHTN variables (Figure 2b). This pre-trained MLP consisted of 5 dense, batch normalization and dropout layers, to predict echocardiography-derived LVMI. This model was fine-tuned on the MHTN data by removing the last 2 layers and unfreezing the last 6 layers for training. The output of these layers, i.e. the subset of tabular data embeddings, were fed through 2 fully connected layers of 10 units to predict LVMI (Figure 2a, bottom row).

[0108] We also performed multitask learning, where a MLP model was built such that the fullyconnected layers estimated LVMI as well as end-diastolic echocardiographic measures (PWTd, LVIDd and IVSd) (Figure 2c). This MLP model was used in the first two pipelines to predict LVMI.

[0109] Finally, for LVH and concentric LVH classifications, the same model architecture wasused with different activation and loss functions (sigmoid activation with binary cross entropy loss).

[0110] The results are summarized in Table 1. Comparisons of performances from differentpipelines are presented for each of three prediction / classification tasks: predicting LVMI, classifying LVH, and classifying concentric LVH. For predicting LVMI, mean absolute errors (MAEs) using 10-fold cross-validation were used as assessment criteria. For classifying LVH and concentric LVH, areas under the curve (AUC) were used as assessment criteria.

[0111] We examined the results further as follows. First, for predicting LVMI, we comparethe MAEs from pipelines with or without time series data from BP waveforms, and employing 23 4901-7498-1200v.1either a neural network or one with multitask learning (as compared against baseline linear regression) (Table 2).

[0112] Table 2: LVMI prediction performance across models using only tabular and tabularwith time series data. The table shows MAEs and standard deviations of LVMI prediction using tabular plus time series data with multitask learning architecture compared against baseline linear regression and neural network models. MAEs were calculated using 10-fold cross-validation. For the multitask learning case with tabular plus time series data, external dataset IDH MAE was included as well. Baseline linear Baseline neural + Multitask learningsingjust tabular, and then tabular plus time series data (BP waveforms), respectively; for linear regression, the addition of time series waveform data did not improve the score significantly (p ≥ 0.90). Using a neural network architecture improved the MAE to 15.4 ± 2.7 and 13.3 ± 2.8 for just tabular and then tabular plus time series data respectively. Both tabular as well as tabular plus time series neural network models improved MAE compared to the linear regression models (p ≤ 0.01), but the addition of time series data did not improve upon just using tabular data. In the case of adding in multitask learning, the MAE with tabular data was 14.8 ± 3.4 using just tabular data, and 10.4 ± 1.5 using both data streams. Multitask learning did not significantly improve the baseline neural network within their own data streams. However, importantly, the best-performing model was multitask learning with tabular plus time series data, which improved the MAE compared to the baseline neural network with just tabular data (p ≤ 0.01), and which exhibited an MAE of 15.6 ± 1.2 on the external IDH dataset. 24 4901-7498-1200v.1

[0114] Second, we highlight another set of results for predicting LVMI, by comparing MAEsfrom pipelines ingesting tabular data (either the full set, or a subset), and employing either a neural network or (in the case of a subset of tabular data) one with transfer learning (Table 3).

[0115] Table 3: LVMI prediction performance across models using data with 7 or 37 variables.The table shows MAEs and standard deviations of LVMI prediction using a subset of 7 variables with transfer learning compared against the baseline neural network model with 7 variables or a complete set of 37 variables. MAEs were calculated using 10-fold cross-validation. Baseline neural network + Transfer learning earning improved themean absolute errors of our LVMI prediction significantly (p ≤ 0.05). Interestingly, this transfer learning model performance by ingesting only a subset of tabular data (MAE of 14.7 ± 3.6) performed similarly to the baseline neural network ingesting the complete set of tabular data (MAE of 15.4 ± 2.7).

[0117] In a third highlight of results, we compared the AUC values of pipelines for classifyingLVH and concentric LVH (Figure 3). Concentric LVH reflects an increased wall thickness, and has been shown to increase risk of mortality and cardiovascular outcomes compared to other LVH subtypes4. The results are shown as receiver operating characteristic (ROC) plots with stratified 10-fold cross-validation for predicting LVH and concentric LVH using tabular and tabular-time series data (Figure 3). Overall, baseline neural network models using tabular data alone, as well as tabular data alongside time series data, exhibited AUC of 0.93 ± 0.09 or better. We also calculated additional performance metrics (precision, recall, F1 score, and balanced accuracy) (Supplementary Table 5); in general, precision scores were higher than recall scores.

[0118] Supplementary Table 5. Obtained metrics from 10-fold cross-validation for LVH andconcentric LVH classification. Labels correspond to Figure 3: Metrics for LVH classification with just the tabular data (A) and tabular plus time series (B) using the MHTN dataset. Metrics for LVH 25 4901-7498-1200v.1classification with just the tabular data (C) and tabular plus time series (D) using the MHTN and IDH dataset. Metrics for concentric LVH classification with just the tabular data (E) and tabular plus time series (F) using the MHTN and IDH dataset. All values are mean ± standard deviation.

[0119] Discussion

[0120] We have described the development and validation of methods to predict LVMI andclassify LVH using data that can be collected outside of the clinic, towards more accessible methods for early stratification of cardiovascular risk. As the method differs from prior LVH and LVMI detection models through its use of ambulatory data, it presents different possibilities for how and when the wearable monitors would be worn by a user. The best-performing pipeline for predicting LVMI involves a multitask learning model which uses BP waveforms, which requires cuff-based BP waveform monitors (Figure 4, left). While ambulatory blood pressure waveform monitors are not as widely available as regular BP cuffs that provide discrete readings, cuffless blood pressure devices also present a promising avenue for overcoming challenges related to measuring blood pressure in outpatient settings33, and could potentially be useful for this method. This study also describes a simpler use case involving transfer learning which does not require BP- related waveforms, where only a regular BP cuff would be needed for discrete BP readings, alongside a subset of demographics variables and BP-related variables (Figure 4, right). The generalizability of these models across populations with different demographic characteristics is further highlighted in this study, as the models performed reasonably across external datasets (e.g. IDH dataset).

[0121] The performances of the models present several observations. For predicting LVMI,the neural networks improved MAEs compared to baseline linear regression. The approximately 26 4901-7498-1200v.113% mean percentage error falls within the range of echocardiography intra- and inter-technician variabilities for left ventricular mass measurements, found to be from 10 - 30%34–36. (Even with these variabilities, echocardiography remains as current practice for diagnosing left ventricular dysfunction and other cardiac conditions37.) Our transfer learning results also point to the possibility of reducing the numbers of variables needed for accurate prediction of cardiac risk, via the use of a pre-trained model (using 7 tabular variables common to both the FHS and MHTN datasets, performing similarly to our baseline neural network architecture, which used 37 tabular variables; see Supplementary Table 2 for a list of the 7 tabular variables and Supplementary Data for a list of the 37 tabular variables). For our external dataset analyses, we used our multitask learning model to predict LVMI for participants in the IDH study. While IDH LVMI prediction errors are higher than the MHTN test set, the model still performed better than a baseline linear regression, marking a step towards a reliable LVMI measurement across different datasets. Therefore, LVMI measurements derived from ambulatory devices can potentially increase the accessibility of LVH risk assessment using the currently accepted clinical methods, and moreover, enable more timely measurements of LVMI for the larger population as LVMI can change on the scale of months38.

[0122] For LVH classification, the accuracies obtained by our deep learning models suggestfurther investigating their viability for stratifying early cardiac risk through analysis of ambulatory data, especially for high risk individuals, such as those with masked hypertension. Studies have shown that individuals with masked hypertension have an increased risk of LVH5,39; the ability to diagnose LVH in masked hypertensives (to be confirmed with echocardiography) would enable healthcare providers to prescribe hypertrophy management earlier than would be caught at a clinic. Our classification models were able to discriminate LVH cases with an AUC of 0.95 or higher (and AUC of 0.93 or higher for concentric LVH). By comparison, previously published LVH detection models used 12-lead ECGs for their analyses, with AUCs of 0.8240, 0.8726, 0.8941 and 0.9642 (refer to Supplementary Table 1 for a more detailed comparison). While the last method achieved a comparable AUC, our method is more appropriate for ambulatory use cases.

[0123] Importantly, our results demonstrate the utility of analyzing waveform data for cardiacrisk assessment. We have specifically demonstrated that adding waveform information to our model in conjunction with multitask learning reduces the error significantly compared to baseline models. These results support previous analyses that found correlations between pressure 27 4901-7498-1200v.1waveform variables and hypertrophy12,15,43. BP waveforms, generally measured through tonometry, PPG or oscillometry, have been used to investigate the mechanical properties of the arterial tree and studies have shown that BP waveforms have different phenotypes corresponding to different cardiac risk profiles15,44–46. Waveform-derived parameters, such as reflection coefficients, provide additional information about left ventricular hypertrophy risk15. For example, diastolic dysfunction is an early feature of hypertensive LVH1, and the BP waveforms of patients with hypertensive LVH have been shown to phenotypically differ, due to wave reflections, from healthy waveforms47. While clinicians currently do not use arterial waveforms in their assessment of cardiovascular disease, our results point to the potential utility of incorporating waveform data as part of cardiac risk determination, with the availability of increasingly sophisticated data analysis techniques.

[0124] Methods

[0125] Datasets: Masked Hypertension Study, Improving the Detection of HypertensionStudy, and Framingham Heart Study

[0126] Data from an ancillary study conducted on participants from Phase 2 of the MaskedHypertension Study (MHTN) where brachial pressure waveforms were measured with a Mobil-O- Graph device (I.E.M.; Industrielle Entwicklung Medizintechnik und Vertriebsgesellschaft mbH, Stolberg, Germany) are used to train and evaluate the LVMI and LVH prediction models. More information about the study protocol and screening procedure can be found in Schwartz et al49.The ancillary study involved the participation of 234 subjects, from which 143 subjects data were used in this study. Each patient’s demographic information, fasting blood sample, 2-dimensional echocardiography-derived variables, ambulatory BP and brachial pressure waveforms (every hour for 40 hours, n=143 individuals) were used in our analysis. As part of the analysis, we removed all patients who did not receive an echocardiogram (3.6% of total). Specific demographics, anthropometric, and echocardiography data used in this study from the ancillary MHTN dataset are provided in Table 4.

[0127] Table 4: Summary statistics for the Masked Hypertension Study (MHTN) and theImproving the Detection of Hypertension (IDH) datasets. All values are mean ± standard deviation or number of samples. BP waveform variables are from brachial pressure waveform 28 4901-7498-1200v.1measurements. *LVH was calculated using BSA adjustment: LVMI >102 g / m2for males and >88 g / m2for females. MHTN (n= 2736 IDH (n= 1342 f 14 f29 4901-7498-1200v.1Concentric LVH, n (%) 27 (19) 6 (17.1) vingthe Detection of Hypertension (IDH) study dataset (n = 400 individuals, 35 of whom were enrolled from the parent IDH study to measure waveforms) are used as an external test set for the LVMI and LVH prediction models. Brachial pressure waveforms (along with BP values) were recorded using the Mobil-O-Graph device every 30 minutes. More information about the study protocol and screening procedure of the parent study can be found in Schwartz et al50. Comparable data to the MHTN dataset was acquired for the IDH dataset. Both the parent studies and the ancillary / pilot studies were approved by institutional review boards, and all participants provided informed consent. Specific demographics, anthropometric, and echocardiography data used in this study from the ancillary IDH dataset are provided in Table 4.

[0129] Data from the 3rd generation of the Framingham Heart Study (FHS) are used fortransfer learning to train a base multilayer perceptron model, to predict LVMI and LVH . In this instance, cardiovascular variables derived from the tonometry waveform from this dataset were used as a stand-in for the variables derived from pressure waveforms during the knowledge transfer stage. Framingham Generation 3 data were accessed through the NHLBI Biologic Specimen and Data Repository (n=3269 patients after dropping rows with missing values). Note this manuscript does not necessarily reflect the opinions or views of FHS or the NHLBI. For the subset list of FHS variables used in the analysis, refer to Supplementary Table 2.

[0130] Summary statistics for the ancillary MHTN and IDH datasets are presented in Table 1.The IDH dataset had a similar distribution of echocardiography variables to the MHTN dataset. The study population had a mean LVMI of 80.7 ± 19.7 g / m2, and 21% percent of the population had LVH. FHS summary statistics are provided in Supplementary Table 2. To be clear, no new data was collected, and instead, pre-existing datasets were leveraged for the development and validation of models found in this study.

[0131] Time series BP waveform processing and tabular data preparation workflows

[0132] In model evaluations which used ambulatory waveforms from the Mobil-O-Graphmonitor, each segment was recorded for 10 seconds at 100 Hz (1000 samples), where each patient had multiple BP waveforms. In these experiments, vectorized waveforms were converted into 30 4901-7498-1200v.1spectrograms (see Fig.5 for an example of this analysis). The spectrograms were then fed into a 1-dimensional convolutional neural network (CNN), in which the kernel moves in the temporal direction. This encoding was then run through a long short-term memory (LSTM)51–53model for sequence analysis, with the output used for downstream processing. For more information on model layers and size, refer to the neural network for time series data model found

[0133] Tabular data included one-time features such as demographics data, BP measurementdata, and laboratory test results. Categorical variables were one-hot encoded. Continuous variables were standardized by de-meaning and normalizing to unit variance. Date and time variables were encoded cyclically through a sine transformation. The tabular data was fed through a multi-layer perceptron (MLP) using dense layers, batch normalization, and dropout, with the output used for downstream processing. For more information on model layers and size, refer to the neural network for tabular data model found in Supplementary Table 3.

[0134] Supplementary Table 3: Layer parameters for deep learning models.31 4901-7498-1200v.1

[0135] Machine learning model evaluations workflow overview (Tabular data alone and incombination with time series BP waveforms) 32 4901-7498-1200v.1

[0136] To measure LVMI prediction and LVH classification performance using tabular dataalone, and in combination with time series data, experiments were designed using different datasets and processing workflows, all set up and carried out using Python and the TensorFlow package54 (see Figure 2). In experiments where tabular data alone is used, ambulatory BP readings provide BP information; while in experiments where tabular data and time series data are used, the above processed ambulatory waveform time series data is additionally utilized for BP information. So for experiments on combined tabular and waveform data, outputs from the above time series BP waveform processing and tabular data preparation workflows are first concatenated so that each row of data is a unique processed waveform alongside discrete BP data found in the tabular data as BP variables. On the other hand, experiments with only tabular data utilize only the discrete BP data as BP variables.

[0137] In both cases, the either concatenated or tabular data alone are fed through fullyconnected layers with ReLU activation for final prediction. For LVMI prediction, echocardiography-derived LVMI was used as the ground truth. This describes the baseline neural network used for LVMI prediction, upon which multitask learning and transfer learning components were built into. For classifying LVH, the same fully connected model with ReLU activations (non-linear activation functions used in neural networks) were used up to the final layer, where a sigmoid activation was used in the final dense layer to produce a probability estimate for LVH. For more information on model layers and size, refer to the output model for single task learning model found in Supplementary Table 3. In the case of LVMI prediction, the loss function used was mean absolute error (MAE). In the case of LVH classification, the loss function was binary cross entropy. The algorithms were trained using the Adam optimizer at a learning rate of 0.01 for 30 epochs.

[0138] Multitask learning

[0139] Multitask learning was employed in multiple experiments involving both tabular dataalone and combined tabular and time series data for LVMI prediction. Since LVMI is an index that is calculated using measured dimensions during diastole, we hypothesized that training our algorithms to output LVMI along with the variables that are used in its calculation could help prediction. LVM was estimated using the Devereux equation16. To calculate LVMI, LVM was divided by estimated body surface area55, as is the standard way for how LVMI is defined. While 33 4901-7498-1200v.1the baseline model was trained with only echocardiography-derived LVMI as the output, in the multitask learning-based approach, the models were trained with LVMI, interventricular septum thickness in diastole (IVSd), left ventricular internal diameter at diastole (LVIDd), and posterior wall thickness in diastole (PWTd) measures as the outputs of interest. Notably these are the variables used in the calculation of LVMI. In this case, the outputs were 4 separate dense layers of 1 unit each for each variable. The composite loss function in the case of this multitask model was constructed with equal task weights, to obtain a standard measure of algorithmic model performance using general multitask learning. Similarly because of this, task domination and error profiles for each task were not explored. Gradient conflict resolution techniques were not specifically employed since by design, the model has aligned output goals.

[0140] Transfer learning

[0141] Transfer learning was used in experiments involving LVMI prediction using a subsetof tabular data. In this instance, a subset of variables from the 3rd generation of the FHS dataset were used to train a base multilayer perceptron model to predict LVMI (see Figure 2 for model architecture, and Supplementary Table 2 for a list of variables used). Variables derived from the tonometry waveform were used as a stand-in for variables derived from pressure waveforms. Following the knowledge transfer stage, the last 6 layers of the transferred base model were unfrozen and further fine-tuned during the training stage using the Masked Hypertension Study training dataset.

[0142] Clinical Designations

[0143] Since we were looking to predict early signs of LVH, we defined the delineation forLVH as anything greater than the upper limit of the normal range of LVMI, where LVMI is LVM normalized by body surface area. Using this, LVH was defined as an LVMI greater than 102 g / m2for males and 88 g / m2for females56. Concentric hypertrophy was defined as LVH with a relative wall thickness > 0.42 cm56. See Fig.6 for details on LVH phenotypes and a visual presentation of these delineations, and Supplementary Table 6 for a tabular representation of these designations.

[0144] Supplementary Table 6. Clinical designations of LVH and concentric LVH. LVH wasdefined as an LVMI greater than 102 g / m2for males and 88 g / m2for females. Concentric hypertrophy was defined as LVH with a relative wall thickness > 0.42 cm. 34 4901-7498-1200v.1

[0145] Statistics: Descriptive statistics are presented as mean ± standard deviation forcontinuous variables and percentages for categorical variables in Table 1 and Supplementary Table 2. Mean absolute error (MAE) ± standard deviation was assessed using 10-fold cross-validation for LVMI predictions in Tables 3 and 4 (with waveforms from the same individual in the same set, to prevent data leakage). MAE is used to provide an interpretation of the average error magnitude across experiments. Binary classification accuracy was used for stratified 10-fold cross-validation for LVH classification in Figure 3. Linear regression baseline models were created using Python’s sklearn library. One-way analyses of variance (ANOVAs) were used for comparing different model means for Table 3 and 4 values using Prism 9. Receiver operating characteristic (ROC) curves were plotted for LVH and concentric LVH detection with 10-fold cross-validation in Figure 3. Binary cross-entropy loss for LVH classification was handled independently. 35 4901-7498-1200v.1REFERENCES 1. Kahan, T. The importance of left ventricular hypertrophy in human hypertension. J. Hypertens. Suppl. 16, S23-9 (1998).2. Frohlich, E. D. & Tarazi, R. C. Is arterial pressure the sole factor responsible for hypertensive cardiac hypertrophy? Am. J. Cardiol. 44, 959–963 (1979).3. Gosse, P. Left ventricular hypertrophy as a predictor of cardiovascular risk. J. Hypertens. Suppl. 23, S27-33 (2005).4. Bang, C. N. et al. Four-group classification of left ventricular hypertrophy based on ventricular concentricity and dilatation identifies a low-risk subset of eccentric hypertrophy in hypertensive patients. Circ. Cardiovasc. Imaging 7, 422–429 (2014).5. Cuspidi, C. et al. Incident left ventricular hypertrophy in masked hypertension. Hypertension 74, 56–62 (2019).6. Devereux, R. B. et al. Left ventricular hypertrophy in hypertension. Prevalence and relationship to pathophysiologic variables. Hypertension 9, II53-60 (1987).7. de Simone, G. et al. Left ventricular mass and incident hypertension in individuals with initial optimal blood pressure: the Strong Heart Study. J. Hypertens. 26, 1868–1874 (2008).8. Narayan, O. et al. Central aortic reservoir-wave analysis improves prediction of cardiovascular events in elderly hypertensives. Hypertension 65, 629–635 (2015).9. Davies, J. E. et al. Excess pressure integral predicts cardiovascular events independent of other risk factors in the conduit artery functional evaluation substudy of Anglo- Scandinavian Cardiac Outcomes Trial. Hypertension 64, 60–68 (2014).10. Behnam, V. et al. Windkessel measures derived from pressure waveforms only: the framingham heart study. J. Am. Heart Assoc. 8, e012300 (2019).11. McEniery, C. M. et al. Increased stroke volume and aortic stiffness contribute to 36 4901-7498-1200v.1isolated systolic hypertension in young adults. Hypertension 46, 221–226 (2005).12. Lekakis, J. P. et al. Cardiac hypertrophy in hypertension: relation to 24-h blood pressure profile and arterial stiffness. Int. J. Cardiol. 97, 29–33 (2004).13. Hashimoto, J., Nichols, W. W., O’Rourke, M. F. & Imai, Y. Association between wasted pressure effort and left ventricular hypertrophy in hypertension: influence of arterial wave reflection. Am. J. Hypertens. 21, 329–333 (2008).14. Roman, M. J., Ganau, A., Saba, P. S. & Devereux, R. B. Left ventricular hypertrophy, arterial compliance, and aging. Adv. Exp. Med. Biol. 432, 13–22 (1997).15. Tice, F. D. et al. Vascular hypertrophy is an early finding in essential hypertension and is related to arterial pressure waveform contour. Am. Heart J. 132, 621–627 (1996). 16. Devereux, R. B. & Reichek, N. Echocardiographic determination of left ventricular mass in man. Anatomic validation of the method. Circulation 55, 613–618(1977). 17. Foppa, M., Duncan, B. B. & Rohde, L. E. P. Echocardiography-based left ventricular mass estimation. How should we define hypertrophy? Cardiovasc. Ultrasound 3,17 (2005). 18. Seo, H.-Y. et al. Discrepancies in Left Ventricular Mass Calculation Based on Echocardiography and Cardiovascular Magnetic Resonance Measurements in Patients with Left Ventricular Hypertrophy. J. Am. Soc. Echocardiogr. 28, 1194–1203, e2 (2015).19. Diaz-Pinto, A. et al. Predicting myocardial infarction through retinal scans and minimal personal information. Nat. Mach. Intell. 4, 55–61 (2022).20. Khurshid, S. et al. Deep Learning to Predict Cardiac Magnetic Resonance- 37 4901-7498-1200v.1Derived Left Ventricular Mass and Hypertrophy From 12-Lead ECGs. Circ. Cardiovasc. Imaging 14, e012281 (2021).21. Duffy, G. et al. High-Throughput Precision Phenotyping of Left Ventricular Hypertrophy With Cardiovascular Deep Learning. JAMA Cardiol. 7, 386–395 (2022).22. Mancia, G. et al. Ambulatory blood pressure is superior to clinic blood pressure in predicting treatment-induced regression of left ventricular hypertrophy. SAMPLE Study Group. Study on Ambulatory Monitoring of Blood Pressure and Lisinopril Evaluation. Circulation 95, 1464–1470 (1997).23. Tucker, B. et al. Left ventricular hypertrophy and ambulatory blood pressure monitoring in chronic renal failure. Nephrol. Dial. Transplant. 12, 724–728 (1997).24. Devereux, R. B. et al. Prognostic significance of left ventricular mass change during treatment of hypertension. JAMA 292, 2350–2356 (2004).25. Ahmed, S. N., Jhaj, R., Sadasivam, B. & Joshi, R. Prediction of Left Ventricular Mass Index Using Electrocardiography in Essential Hypertension - A Multiple Linear Regression Model. Med Devices (Auckl) 13, 163–172 (2020).26. Tison, G. H., Zhang, J., Delling, F. N. & Deo, R. C. Automated and interpretable patient ECG profiles for disease detection, tracking, and discovery. Circ. Cardiovasc. Qual. Outcomes 12, e005289 (2019).27. Satija, U., Ramkumar, B. & Manikandan, M. S. A review of signal processing techniques for electrocardiogram signal quality assessment. IEEE Rev. Biomed. Eng. 11, 36–52 (2018). 28. Caruana, R. Multitask Learning. Machine Learning 28, 41–75 (1997).29. Bird, J. J., Kobylarz, J., Faria, D. R., Ekart, A. & Ribeiro, E. P. Cross-Domain 38 4901-7498-1200v.1MLP and CNN Transfer Learning for Biological Signal Processing: EEG and EMG. IEEE Access 8, 54789–54801 (2020).30. Bahador, N. & Kortelainen, J. Deep learning-based classification of multichannel bio-signals using directedness transfer learning. Biomed. Signal Process. Control 72,103300 (2022). 31. Bizzego, A., Gabrieli, G. & Esposito, G. Deep neural networks and transfer learning on a multivariate physiological signal dataset. Bioengineering (Basel) 8, 35 (2021).32. Otović, E. et al. Intra-domain and cross-domain transfer learning for time series data—How transferable are the features? Knowledge-Based Systems 239, 107976 (2022).33. Bradley, C. K. et al. Cuffless blood pressure devices. Am. J. Hypertens. 35, 380–387 (2022). 34. Himelman, R. B., Cassidy, M. M., Landzberg, J. S. & Schiller, N. B. Reproducibility of quantitative two-dimensional echocardiography. Am. Heart J. 115, 425–431 (1988). 35. Sahn, D. J., DeMaria, A., Kisslo, J. & Weyman, A. Recommendations regarding quantitation in M-mode echocardiography: results of a survey of echocardiographic measurements. Circulation 58, 1072–1083 (1978).36. Kuecherer, H. F., Kee, L. L., Modin, G., Cheitlin, M. D. & Schiller, N. B. Echocardiography in serial evaluation of left ventricular systolic and diastolic function: importance of image acquisition, quantitation, and physiologic variability in clinical and investigational applications. J. Am. Soc. Echocardiogr. 4, 203–214 (1991).37. Ciampi, Q. & Villari, B. Role of echocardiography in diagnosis and risk stratification in heart failure with left ventricular systolic dysfunction. Cardiovasc. 39 4901-7498-1200v.1Ultrasound 5, 34 (2007).38. Franz, I. W., Tönnesmann, U. & Müller, J. F. Time course of complete normalization of left ventricular hypertrophy during long-term antihypertensive therapy with angiotensin converting enzyme inhibitors. Am. J. Hypertens. 11, 631–639 (1998).39. Cuspidi, C. et al. Left ventricular hypertrophy in isolated and dual masked hypertension. J Clin Hypertens (Greenwich) 22, 673–677 (2020).40. Sparapani, R. et al. Detection of left ventricular hypertrophy using bayesian additive regression trees: the MESA. J. Am. Heart Assoc. 8, e009959 (2019).41. Liu, C.-M. et al. Artificial Intelligence-Enabled Model for Early Detection of Left Ventricular Hypertrophy and Mortality Prediction in Young to Middle-Aged Adults. Circ. Cardiovasc. Qual. Outcomes 15, e008360 (2022).42. Liu, C.-W. et al. Left ventricular hypertrophy detection using electrocardiographic signal. Sci. Rep. 13, 2556 (2023).43. Post, W. S. & Levy, D. New developments in the epidemiology of left ventricular hypertrophy. Curr. Opin. Cardiol. 9, 534–541 (1994).44. Picone, D. S. et al. Discovery of new blood pressure phenotypes and relation to accuracy of cuff devices used in daily clinical practice. Hypertension 71, 1239–1247 (2018).45. Roman, M. J. & Devereux, R. B. Association of central and peripheral blood pressures with intermediate cardiovascular phenotypes. Hypertension 63, 1148–1153(2014). 46. Anyfanti, P. et al. Noninvasive assessment of myocardial perfusion in different blood pressure phenotypes and its association with arterial stiffness indices. Am. J. Hypertens. 32, 557–563 (2019).40 4901-7498-1200v.147. Ikonomidis, I. et al. Incremental value of arterial wave reflections in the determination of left ventricular diastolic dysfunction in untreated patients with essential hypertension. J. Hum. Hypertens. 22, 687–698 (2008).48. Ferrara, L. A. et al. Indexation criteria of ventricular mass and predictive role of blood pressure and body composition. Am. J. Hypertens. 18, 1282–1287 (2005).49. Schwartz, J. E., Feig, P. U. & Izzo, J. L. Pulse wave velocities derived from cuff ambulatory pulse wave analysis. Hypertension 74, 111–116 (2019).50. Schwartz, J. E. et al. Reliability of office, home, and ambulatory blood pressure measurements and correlation with left ventricular mass. J. Am. Coll. Cardiol. 76, 2911–2922 (2020). 51. Hochreiter, S. & Schmidhuber, J. Long Short-Term Memory. Neural Comput. 9,1735–1780 (1997). 52. Connor, J. & Atlas, L. Recurrent neural networks and time series prediction. in IJCNN-91-Seattle International Joint Conference on Neural Networks 301–306 (IEEE, 1991). doi:10.1109 / IJCNN.1991.155194. 53. Wang, M. et al. A deep learning network based on CNN and sliding window LSTM for spike sorting. Comput. Biol. Med. 159, 106879 (2023).54. Abadi, M. et al. {TensorFlow}: A System for {Large-Scale} Machine Learning. in 265–283 (USENIX Association, 2016). 55. Bois, D. D. & Bois, E. F. D. A height-weight formula to estimate the surface area of man. Exp. Biol. Med. 13, 77–78 (1916).56. Lang, R. M. et al. Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and 41 4901-7498-1200v.1the European Association of Cardiovascular Imaging. J. Am. Soc. Echocardiogr. 28, 1-39.e14 (2015). 42 4901-7498-1200v.1

Claims

CLAIMS 1. A method of estimating a left ventricular mass index (LVMI) value of a subject comprises steps of: obtaining pressure information associated with ambulatory blood pressure of the subject; obtaining tabular data associated with the subject; and processing the pressure information and the tabular data using a first machine learning algorithm trained using a first training set and providing a first ventricular mass index value as an output.

2. The method of claim 1, wherein the step of obtaining pressure information comprises: obtaining time series blood pressure waveform data associated with the subject; and converting the time series blood pressure waveform data into spectrogram data.

3. The method of claim 2, wherein the time series blood pressure waveform data corresponds to a 10 second waveform at 100Hz.

4. The method of claim 1, 2 or 3, wherein the pressure information is provided using photoplethysmography.

5. The method of claim 2, 3 or 4, wherein the processing step includes providing the spectrogram data as an input to a convolutional neural network (CNN) Long Short-Term Memory Network trained using at least the first training set and providing a first output associated with the first ventricular mass index value, wherein the first ventricular mass index value is based on the first output. 43 4901-7498-1200v.

16. The method of claim 5, wherein the convolutional neural network (CNN) Long Short-Term Memory Network uses a 1-dimensional convolutional neural network.

7. The method of any of claims 2-6, wherein the processing step includes: providing the spectrogram data as an input into a convolutional neural network; providing an output of the convolutional neural network as an input to a Long Short-Term Memory Network, wherein the first ventricular mass index value is based on a concatenation of the output of the Long Short-Term Memory Network and an output of a multi-layer perceptron mode using the tabular information as an input.

8. The method of any of claims 5-7, wherein the convolutional neural network (CNN) is a 1-dimensional convolutional neural network.

9. The method of any of claims 1-8, wherein the tabular information includes a first group of information including demographic data and test results associated with the subject, and, optionally, wherein the first group of information has been obtained or provided prior to obtaining pressure information.

10. The method of claim 9, wherein the processing step includes providing, as inputs, the first output and the first group of information to a first multi-layer perceptron model and concatenating the inputs to provide a second output associated with the first ventricular mass index value and other parameters related to the first ventricular mass index value.

11. The method of claim 10, wherein the other parameters include a left ventricular hypertrophy parameter and / or a concentric left ventricular hypertrophy parameter.

12. The method of claim 9, 10 or 11, wherein the tabular information includes a second group of tabular information and the processing step includes providing the second group 44 4901-7498-1200v.1of tabular information to a second multi-layer perceptron model trained using a second training set and providing a second output associated with the first ventricular mass index value.

13. The method of claim 12, wherein the second multi-layer perceptron model is trained based on at least a portion of Framingham Heart Study data.

14. The method of claim 13, wherein the Framingham Heart Study data comprises BP measurements, IVS, PWTD, LVIDD, LVMI, CO and LVH data.

15. The method of claim 13 or 14, wherein the Framingham Heart Study data further comprises tonometry, height, weight, Echo result, sex and / or age data.

16. The method of claim 9, 10, 11, 12 or 13, wherein the demographic data includes age, sex, and / or body mass index of the subject.

17. The method of claim 14, wherein the demographic data includes race, ethnicity, and / or history of familial hypertension of the subject.

18. The method of any of claims 9-14, wherein the test results include blood HDL cholesterol, LDL cholesterol, triglycerides, glucose, and / or HbA1c of the subject.

19. The method of any of claims 9-15, wherein the test results include urine sodium, urine potassium, urine chloride, urine creatinine, and / or microalbumin creatinine ratio of the subject.

20. The method of any of claims 1-16, wherein the pressure information includes systolic blood pressure, diastolic blood pressure, pulse pressure, and / or heart rate of the subject.

21. A method of diagnosing a left ventricular hypertrophy in a subject comprising determining a LVMI of a subject’s heart by the method of any of claims 1-20. 45 4901-7498-1200v.

122. The method of claim 21, wherein the LVMI is determined without using any transthoracic echocardiography.

23. The method of claim 21 or 22, wherein concentric LVH is diagnosed.

24. A method of treating a left ventricular hypertrophy (LVH) in a subject comprising identifying, or having identified, the subject as having LVH by the method of claim 21, 22 or 23 and administering to a subject so-identified an anti-hypertensive therapeutic.

25. The method of claim 24, wherein concentric LVH is treated.

26. The method of claim 24 or 25, wherein the anti-hypertensive therapeutic comprises: a diuretic, a beta-blocker, an angiotensin-converting enzyme (ACE) inhibitor, a calcium channel blocker, an alpha-adrenergic blocker, a renin inhibitor, a direct acting peripheral vasodilator, or a centrally acting agent.

27. The method of claim 26, wherein the anti-hypertensive therapeutic comprises a thiazide diuretic or a beta-blocker.

28. The method of any of claims 20-27, wherein the subject has essential hypertension, renal artery stenosis, athletic heart with physiological LVH, aortic valvar stenosis, coarctation of an aorta, hypertrophic cardiomyopathy without or with outflow tract obstruction (HOCM), subaortic stenosis (left ventricular outflow tract obstruction by muscle or membrane), aortic regurgitation, mitral regurgitation, dilated cardiomyopathy, ventricular septal defect, or an infiltrative cardiac processes (amyloidosis, Fabry disease, Danon disease).

29. The method of any of claims 1-28, wherein the subject is male and the LVMI is > 102 g / m2, normalized by body surface area of the subject.

30. The method of any of claims 1-28, wherein the subject is female and the LVMI is > 88 g / m2, normalized by body surface area of the subject. 46 4901-7498-1200v.

131. The method of any of claims 1-30, wherein concentric hypertrophy is determined as LVH with a relative wall thickness > 0.42 cm.

32. The method of claim 29 or 30, which identifies early stage LVH.

33. A system for performing the method of any of claims 1-23, comprising: an ambulatory blood pressure measuring device configured to provide the pressure information associated with the subject; a computer system operably connected to the ambulatory blood pressure measuring device, wherein the ambulatory blood pressure measuring device is connected to the computer system via a wired or wireless connection to provide the pressure information to the computer system, wherein the computer system includes: a processor; and memory operably connected to the processor and including processor executable code, that, when executed by the processor, execute the steps set forth in claims 1-23 to provide the left ventricular mass index (LVMI) value.

34. The system of claim 33, wherein the ambulatory blood pressure measuring device is an upper arm cuff-based ambulatory blood pressure monitor.

35. The system of claim 33, wherein the ambulatory blood pressure measuring device is cuffless ambulatory blood pressure monitor.

36. The system of any of claims 33-35, wherein the ambulatory blood pressure measuring device is waterproof or water-resistant.

37. The system of any of claims 33-36, comprising a digital recorder that records ambulatory blood pressure associated parameter(s). 47 4901-7498-1200v.

138. The system of any of claim 37, programmed to take blood pressure readings at regular intervals throughout the day and / or night.

39. The method of any of claims 1-32 or system of any of claims 33-38 which discriminates LVH cases with an AUC of 0.95 or better and / or discriminates concentric LVH cases with an AUC of 0.93 or better.

40. The method of any of claims 1-32, wherein the tabular data comprises 10 or less variables.

41. The method of any of claims 1-32 or 40, wherein the tabular data comprises 7 variables.

42. The method of any of claims 1-32, wherein the tabular data comprises 37 variables.

43. A system for estimating a left ventricular mass index (LVMI) value of a subject comprising: an ambulatory blood pressure measuring device configured to provide pressure information associated with ambulatory blood pressure of the subject; a processor operably connected to the ambulatory blood pressure measuring device and configured to receive the pressure information; memory, operably connected to the processor and which includes tabular data associated with the subject, and processor executable code that, when executed by the processor, performs steps of: obtaining the pressure information; obtaining the tabular data; and 48 4901-7498-1200v.1processing the pressure information and the tabular data using a first machine learning algorithm trained using a first training set and providing a first ventricular mass index value as an output.

44. The system of claim 43, wherein the step of obtaining pressure information comprises: obtaining time series blood pressure waveform data from the ambulatory blood pressure measuring device; and converting the time series blood pressure waveform data into spectrogram data.

45. The system of claim 44, wherein the time series blood pressure waveform data corresponds to a 10 second waveform at 100Hz.

46. The system of claim 43, 44 or 45, wherein the ambulatory blood pressure measuring device uses photoplethysmography.

47. The system of claim 44, 45 or 46, wherein the processing step includes providing the spectrogram data as an input to a convolutional neural network (CNN) Long Short-Term Memory Network trained using at least the first training set and providing a first output associated with the first ventricular mass index value, wherein the first ventricular mass index value is based on the first output.

48. The system of claim 47, wherein the convolutional neural network (CNN) Long Short-Term Memory Network uses a 1-dimensional convolutional neural network.

49. The system of any of claims 44-48, wherein the processing step includes: providing the spectrogram data as an input into a convolutional neural network; and providing an output of the convolutional neural network as an input to a Long Short-Term Memory Network, 49 4901-7498-1200v.1wherein the first ventricular mass index value is based on a concatenation of the output of the Long Short-Term Memory Network and an output of a multi-layer perceptron mode using the tabular information as an input.

50. The system of any of claims 47-49, wherein the convolutional neural network (CNN) is a 1-dimensional convolutional neural network.

51. The system of any of claims 43-50, wherein the tabular information includes a first group of information including demographic data and test results associated with the subject.

52. The system of claim 51, wherein the processing step includes providing, as inputs, the first output and the first group of information to a first multi-layer perceptron model and concatenating the inputs to provide a second output associated with the first ventricular mass index value and other parameters related to the first ventricular mass index value.

53. The system of claim 52, wherein the other parameters include a left ventricular hypertrophy parameter and / or a concentric left ventricular hypertrophy parameter.

54. The system of claim 51, 52 or 53, wherein the tabular information includes a second group of tabular information and the processing step includes providing the second group of tabular information to a second multi-layer perceptron model trained using a second training set and providing a second output associated with the first ventricular mass index value.

55. The system of claim 54, wherein the second multi-layer perceptron model is trained based on at least a portion of Framingham Heart Study data.

56. The system of claim 55, wherein the Framingham Heart Study data comprises BP measurements, IVS, PWTD, LVIDD, LVMI, CO and LVH data.

57. The system of claim 55 or 56, wherein the Framingham Heart Study data further comprises tonometry, height, weight, Echo result, sex and / or age data. 50 4901-7498-1200v.

158. The system of claim 51, 52, 53, 54 or 55, wherein the demographic data includes age, sex, and / or body mass index of the subject.

59. The system of claim 56, wherein the demographic data includes race, ethnicity, and / or history of familial hypertension of the subject.

60. The system of any of claims 51-56, wherein the test results include blood HDL cholesterol, LDL cholesterol, triglycerides, glucose, and / or HbA1c of the subject.

61. The system of any of claims 51-57, wherein the test results include urine sodium, urine potassium, urine chloride, urine creatinine, and / or microalbumin creatinine ratio of the subject.

62. The system of any of claims 43-58, wherein the pressure information includes systolic blood pressure, diastolic blood pressure, pulse pressure, and / or heart rate of the subject.

63. A system of diagnosing a left ventricular hypertrophy in a subject by determining a LVMI of a subject’s heart using the system of any of claims 43-62.

64. The system of claim 63, wherein the LVMI is determined without using any transthoracic echocardiography.

65. The system of claim 63 or 64, wherein concentric LVH is diagnosed. 51 4901-7498-1200v.1

Images