Modally reduced representation learning of multi-channel ECG signals through simultaneous alignment and reconstruction

Abstract

Methods and systems described herein provide for modally reduced representation learning techniques for ECG signals. Such a method includes obtaining, for each person in a batch of persons, electrocardiogram (ECG) data for each of N-channels sensed using one or more ECG recorder devices, wherein for each person, in the batch of persons, the ECG data sensed for each of N-channels are sensed concurrently and correspond to a same window of time. The method also includes training, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons, wherein each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair, and wherein reconstruction losses and alignment losses are jointly optimized during the training of each autoencoder of the N autoencoders.

Description

TITLEMODALLY REDUCED REPRESENTATION LEARNING OF MULTI-CHANNEL ECG SIGNALS THROUGH SIMULTANEOUS ALIGNMENT AND RECONSTRUCTIONINVENTORSMasood MortazaviNabil IbtehazCLAIM OF PRIORITY

[0001] This application claims priority to U.S. Provisional Patent Application No. 63 / 579,854, filed August 31 , 2023, and U.S. Provisional Patent Application No. 63 / 554,804, filed February 16, 2024, both titled MODALLY REDUCED REPRESENTATION LEARNING OF MULTI-LEAD ECG SIGNALS THROUGH SIMULTANEOUS ALIGNMENT AND RECONSTRUCTION, and both incorporated by reference herein in their entirety.FIELD

[0002] Embodiments described herein generally provide for modally reduced representation learning methods for electrocardiogram (ECG) signals, and related systems. Such embodiments can be used, for example, to improve the usefulness of ECG signals obtained using consumer grade ECG recorder devices, which are readily available and are less expensive than and easier to use than first type of ECG recorder devices.BACKGROUND

[0003] An electrical stimulus is generated by a sinoatrial node (sinus node) of the heart. This sinus node generates an electrical stimulus regularly to activate the atria. The electrical stimulus travels down through the conduction pathways to stimulate upper and lower chambers of the heart. Normally at rest, as the electrical impulsemoves through the heart to cause the heart to rhythmically contract as a heartbeat. Typically, the heart contracts about 60 to 100 times a minute.

[0004] Electrocardiogram (ECG) signals, profiling the electrical activities of the heart, are used for a plethora of diagnostic and other applications. However, ECG acquisition devices require multiple leads (aka channels) of signals to capture a complete view of a heart (aka cardiac system), which limits their application in consumer grade devices, such as smartwatches and other wearables.

[0005] The importance of ECG signals cannot be exaggerated due to its widespread use in diagnosis of cardiovascular diseases, which are the leading causes of death worldwide and claims a life every 33 seconds in the United States alone. ECG signals capture and allow observation and analysis of the electrical activities of a heart during cardiac cycles. An ECG signal can also be referred to herein more succinctly herein as an ECG.

[0006] Conventionally, utilizing ECG signals to their fullest potential required use of a conventional 12-channel (aka 12-lead) or 14-channel (aka 14-lead) medical grade ECG recorder device. Such a conventional 12- or 14-channel ECG recorder device utilizes 10 skin electrodes placed on a person’s arms, legs and around their chest to sense the 12- or 14-channel ECG. The 10 skin electrodes are named right arm (RA), left arm (LA), right leg (RL), left leg (LL), V1 , V2, V3, V4, V5, V6 electrodes. The V1 , V2, V3, V4, V5, V6 electrodes are placed around the chest, and more specifically, the V1 electrode is place at the fourth intercostal space on the right sternum, the V2 electrode is placed at the fourth intercostal space at the left sternum, the V3 electrode is placed midway between placement of the V2 and V4 electrodes, the V4 electrode is placed at the fifth intercostal space at the midclavicular line, the V5 electrode is placed anterior an axillary line on the same horizontal level as the V4 electrode, and the V6 electrode is place mid-axillary line on the same horizontal level as the V4 and V5 electrodes. The limb electrodes RA, LA, RL, LL are placed on limbs, and more specifically, the RA electrode is placed anywhere between the right shoulder and right elbow, the RL electrode is placed anywhere below the right torso and above the right ankle, the LA electrode is placed anywhere between the left shoulder and the left elbow, and the LL electrode is placed anywhere below the left torso and above the left ankle.

[0007] Electrical activities of the heart can be picked up on the skin via the above described electrodes. The measurement of a voltage requires 2 poles, i.e. , a negative pole and a positive pole. A medical grade ECG recorder device uses the negative pole as a zero reference. Thus, the position of the positive pole can be thought of as the “point of view,” and the line between the 2 poles can be thought of as the “line of sight”. Leads (aka channels) I, II, and III are bipolar in that they measure electrical potential between 2 of the 3 limb electrodes. Lead I (aka channel I) represents the voltage between the RA electrode, which acts as the negative pole, and the LA electrode, which acts as the positive pole, and thus looks at the heart from the left. Lead II (aka channel II) measures electrical potential between the RA electrode, which acts as the negative pole, and the LL electrode, which acts as the positive pole, providing for an inferior left view. Lead III (aka channel III) measures electrical potential between the LA electrode, which acts as the negative pole, and the LL electrode, which acts as the positive pole, providing for an inferior right view. Leads aVR, aVL, and aVF, which are also known as “augmented limb leads,” are unipolar, and more specifically, they use one limb electrode as the positive pole, and take the average of inputs from the other two limb electrodes as the zero reference. Hence, lead aVR lead (aka channel) looks at the upper right side of the heart, aVL lead (aka channel) looks at the upper left side of the heart, and aVF lead (aka channel) looks at the inferior wall of the heart. The chest leads, which are also known as the precordial leads, view the heart in a horizontal plane and are unipolar leads. The corresponding chest electrodes serve as the positive poles, and the reference negative value is the same for all chest leads and is calculated as the average of inputs from the three limb electrodes. Depolarization of cardiac tissue toward a lead produces a positive deflection and depolarization of cardiac tissue away from a lead produces gives a negative deflection. The reverse is the case for depolarization, i.e., repolarization of cardiac tissue toward a lead produces a negative deflection and repolarization of cardiac tissue away from a lead produces gives a positive deflection. Accordingly, it should be appreciated that leads (aka channels) that view the heart from different angles may have deflections pointing in different directions.

[0008] A benefit of the above described medical grade 12-lead (aka 12-channel) ECG recorder device is that it provides for medical grade ECG signals that provide for much more accurate analysis of a heart than simpler consumer grade ECG recorderdevices, such as wearable devices (e.g., smartwatches and chest patches) and portable consumer grade ECG recorders that are much more ambulatory but only provide one or two leads (aka channels). However, because a medical grade 12-lead ECG recorder device requires 10 skin electrodes attached to limbs and the chest, their use is typically limited to medical facilities, and they are not ambulatory. Hence, there is growing interest in developing competent ways of analyzing a person’s heart using a limited set of channels that can be sensed using a consumer grade ECG recorder device, wherein the consumer grade ECG recorder device may only be capable of sensing one or two channels.SUMMARY

[0009] One aspect of the present disclosure relates to a computer implemented method that includes obtaining, for each person in a batch of persons, electrocardiogram (ECG) data for each of N-channels sensed using one or more ECG recorder devices, wherein for each person, in the batch of persons, the ECG data sensed for each of N-channels are sensed concurrently and correspond to a same window of time, and wherein N is an integer that is at least 3. The method also includes training, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons, wherein each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair, and wherein reconstruction losses and alignment losses are jointly optimized during the training of each autoencoder of the N autoencoders.

[0010] Optionally, in any preceding aspect, the obtaining, for each person in the batch of persons, the ECG data for each of the N-channels sensed using the one or more ECG recorder devices, comprises obtaining, for each person in the batch of persons: ECG data for each of M-channels sensed using an M-channel first type of ECG recorder device (e.g., an M-channel medical grade ECG recorder device); and ECG data for each of X-channels sensed using an X-channel second type of ECG recorder device (e.g., an X-channel consumer grade ECG recorder devices, such as a smartwatch or chest patch). For each person in the batch of persons, the ECG datasensed using the M-channel first type of ECG recorder device and the ECG data sensed using the X-channel second type of ECG recorder device are sensed concurrently and correspond to the same window of time. M is an integer that is at least 2. X is an integer that is at least 1 and is less than M. M + X = N, and thus, the N-channels of ECG data are obtained for each person in the batch of persons.

[0011] Optionally, in any preceding aspect, the N autoencoders include: M autoencoders trained using the ECG data sensed using the M-channel first type of ECG recorder device; and X autoencoder(s) trained using the ECG data sensed using the X-channel second type of ECG recorder device. After the training of the N autoencoders, the method further includes obtaining further ECG data for each of at least one of the X-channels using one of the X-channel second type of ECG recorder device, using at least one of the encoders of the X autoencoder(s) to generate at least one embedding based on the further ECG data, and providing the at least one embedding to a further artificial neural network model to at least one of detect a cardiac abnormality, perform biometric identification, or perform biometric authentication.

[0012] Optionally, in any preceding aspect, during the training of each encoder and decoder pair, of the N autoencoders, the alignment losses are optimized by reducing alignment losses for embeddings produced from the N-channels of ECG data obtained for a same person in the batch of persons, and increasing alignment losses for embeddings produced from the N-channels of ECG data obtained for different persons in the batch of persons.

[0013] Optionally, in any preceding aspect, during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, each encoder of the N autoencoders produces respective embeddings that are shared with other encoders of the N autoencoders to enable the alignment losses to be calculated and optimized.

[0014] Optionally, in any preceding aspect, during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, the reconstruction losses are weighted greater than the alignment losses during earlier epochs compared to during later epochs, andthe alignment losses are weighted greater than the reconstruction losses during the later epochs compared to during the earlier epochs.

[0015] Optionally, in any preceding aspect, the N autoencoders are trained using masked autoencoding.

[0016] Optionally, in any preceding aspect, the X-channel second type of ECG device comprises a wearable device.

[0017] Optionally, in any preceding aspect, M = 12, X = 1 , and N = 13.

[0018] Optionally, in any preceding aspect, M = 12 or 14, X = 1 or 2, and N = 13,14, 15, or 16.

[0019] Another aspect of the present disclosure relates to a system comprising a storage medium comprising computer instructions, and one or more processors coupled to communicate with the storage medium. The one or more processors execute the instructions to cause the system to obtain, for each person in a batch of persons, electrocardiogram (ECG) data for each of N-channels sensed using one or more ECG recorder devices. For each person, in the batch of persons, the ECG data sensed for each of N-channels are sensed concurrently and correspond to a same window of time, wherein N is an integer that is at least 3. The one or more processors execute the instructions to also cause the system to train, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons. Each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair. The one or more processors execute the instructions to cause reconstruction losses and alignment losses to be jointly optimized during the training of each autoencoder of the N autoencoders.

[0020] Optionally, in any preceding aspect, to obtain, for each person in the batch of persons, the ECG data for each of the N-channels sensed using the one or more ECG recorder devices, the one or more processors execute the instructions to cause the system to obtain, for each person in the batch of persons: ECG data for each of M-channels sensed using an M-channel first type of ECG recorder device; and ECGdata for each of X-channels sensed using an X-channel second type of ECG recorder device. For each person in the batch of persons, the ECG data sensed using the M- channel first type of ECG recorder device and the ECG data sensed using the X- channel second type of ECG recorder device are sensed concurrently and correspond to the same window of time. M is an integer that is at least 2, X is an integer that is at least 1 and is less than M, and M + X = N, and thus, the N-channels of ECG data are obtained for each person in the batch of persons.

[0021] Optionally, in any preceding aspect, the N autoencoders include: M autoencoders trained using the ECG data sensed using the M-channel first type of ECG recorder device; and X autoencoder(s) trained using the ECG data sensed using the X-channel second type of ECG recorder device. After the training of the N autoencoders, the one or more processors execute the instructions to cause the system to: obtain further ECG data for each of at least one of the X-channels using the X-channel second type of ECG recorder device; use at least one of the encoders of the X autoencoder(s) to generate at least one embedding based on the further ECG data; and provide the at least one embedding to a further artificial neural network model to at least one of detect a cardiac abnormality, perform biometric identification, or perform biometric authentication.

[0022] Optionally, in any preceding aspect, during the training of each encoder and decoder pair, of the N autoencoders, the one or more processors execute the instructions to cause the system to optimize the alignment losses by: reducing alignment losses for embeddings produced from the N-channels of ECG data obtained for a same person in the batch of persons; and increasing alignment losses for embeddings produced from the N-channels of ECG data obtained for different persons in the batch of persons.

[0023] Optionally, in any preceding aspect, during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, each encoder of the N autoencoders produces respective embeddings that are shared with other encoders of the N autoencoders to enable the alignment losses to be calculated and optimized.

[0024] Optionally, in any preceding aspect, during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, the one or more processors execute the instructions to cause the reconstruction losses to be weighted greater than the alignment losses during earlier epochs compared to during later epochs, and the alignment losses to be weighted greater than the reconstruction losses during the later epochs compared to during the earlier epochs.

[0025] Optionally, in any preceding aspect, the N autoencoders are trained using masked autoencoding.

[0026] Optionally, in any preceding aspect, the X-channel second type of ECG device comprises a wearable device.

[0027] Optionally, in any preceding aspect, M = 12, X = 1 , and N = 13.

[0028] Optionally, in any preceding aspect, M = 12 or 14, X = 1 or 2, and N = 13,14, 15, or 16.

[0029] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter. The claimed subject matter is not limited to implementations that solve any or all disadvantages noted in the Background.BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Aspects of the present disclosure are illustrated by way of example and are not limited by the accompanying Figures (FIGS.) for which like references indicate elements.

[0031] FIG. 1 is a high level block diagram used to describe embodiments of the present disclosure that provide for modally reduced (aka channel-reduced)representation learning of multi-lead (aka multi-channel) ECG signals through simultaneous alignment and reconstruction optimization.

[0032] FIG. 2 is a high level block diagram used to explain how a coordinator can be used during the training of the autoencoders introduced in the high level block diagram of FIG. 1 .

[0033] FIG. 3A is a high level block diagram used to explain how a single encoder trained in accordance with an embodiment of the present disclosure, can be used together with a downstream further artificial neural network (ANN) to perform one or more specific tasks.

[0034] FIG. 3B is a high level block diagram used to explain how two encoders trained in accordance with an embodiment of the present disclosure, can be used together with a downstream further ANN to perform one or more specific tasks.

[0035] FIG. 4A is high level flow diagram used to summarize computer implemented methods according to certain embodiments of the present disclosure.

[0036] FIG. 4B is high level flow diagram used to summarize specific embodiments of the methods previously summarized with reference to FIG. 4A.

[0037] FIGS. 5A-5C illustrates example types of ECG recorder devices with which embodiments of the present disclosure can be used.

[0038] FIG. 6 illustrates an example of a computing system with which embodiments disclosed herein may be implemented.

[0039] FIG. 7 illustrates a schematic diagram of a general-purpose network component or computer system with which embodiments of the present disclosure may be implemented.DETAILED DESCRIPTION

[0040] Certain embodiments of the present disclosure described herein relate to modally reduced representation learning methods for ECG signals that are capable ofgenerating channel agnostic and unified representations for ECG signals. Through joint optimization of reconstruction and alignment losses, embeddings of different channels include an amalgamation of the overall heart (aka cardiac) information across channels while also retaining their specific information. Related systems are also described herein. More specifically, certain embodiments described herein involve training, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons, wherein each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair, and wherein reconstruction losses and alignment losses are jointly optimized during the training of each autoencoder of the N autoencoders. The training of the N autoencoders using the N-channels of ECG data obtained for each person in the batch of persons, such that reconstruction losses and alignment losses are jointly optimized, highly enhances the capability of the encoder of each individual channel (of the N-channels) to generate meaningful embeddings as extracted features for down-stream tasks, such as detecting a cardiac abnormality, performing biometric identification, and / or performing biometric authentication, but not limited thereto. This training provides for high-quality single-channel embeddings, or more generally high-quality X-channel embeddings, where X < N. As will be described in additional detail below, the training can occur using an auto-differentiation platform that backpropagate differentiable losses to optimize the encoders and decoders of the N-channels in order to reduce statistical losses. Further, as will be described in additional detail below, this optimization process enhances the capability of each single-channel encoder of each channel to capture heart semantics in the latent embedding the single-channel encoder produces from its corresponding input signal, which enables these embeddings as extracted features to be used for downstream tasks. This beneficially enables consumer grade ECG recorder devices (e.g., that can only obtain one or two channels of ECG data) to be used for performing useful downstream tasks, such as detecting a cardiac abnormality, performing biometric identification, and / or performing biometric authentication, with accuracies that previously could only be achieved using medical grade ECG recorder devices.

[0041] In accordance with certain embodiments, N-channels of ECG signals, which can be represented by the equation x(t) = {x[1](t), x[2](t), x[3](t), . . . , x[N](t)} e R12, aretreated as snapshots of different projections of the states and activities of a heart H at an instant in time t, where the different channels xl'l(t) corresponds to different projections Pi, i.e. , x[i](t) = Pi(H@t). In order to extract a complete state of the heart H, suitable inverse functions of the projections can be used. Using a plurality of autoencoders (AEs), each of which includes a respective encoder and decoder pair, encoders are used to extract latent representations (aka latent embeddings) of the heart h[i](t) from input signals x[i](t) and decoders are used to reconstruct the input signals x[i](t) from the latent representations (aka latent embeddings). Thus, in certain embodiments, a set of encoder networks Ei, i = 1 , 2, 3, . . . , N are assigned to extract the latent heart states from the different ECG channel signals. Further, the latent representations (aka latent embeddings) are correlated as it is hypothetically assumed the embeddings are learning the representation of the same heart H, i.e., Ei(xl'l(t)) = h[i](t) « H@t. This correlated representation learning is performed as the individual ECG channel signals are retrieved from the latent embeddings by a set of decoders, i.e., Di, i = 1 , 2, 3, . . . , N, i.e., Di(h[i](t)) = x[i](t) « x[i](t). Therefore, under this formulation, the decoders try to learn the projections, whereas the encoders aim at learning their inverse.

[0042] In accordance with certain embodiments, artificial neural network (ANN) models are implemented using a masked autoencoder (MAE) architecture that takes N single-channel ECG signals as inputs to each of N models, and more specifically, N autoencoders (AEs). This enables 1 -Dimensional (1 D) positional and patch embeddings to be implemented using sine-cosine position embedding and 1 D convolution, respectively. Each model can, for example, take T = 5 sec, fs= 500 Hz signals as inputs and break them into 100 samples or 0.2 sec long non-overlapping patches. In accordance with certain embodiments, the autoencoder (AE) of each channel includes a strong encoder in conjunction with a weak decoder, i.e., the encoder in each channel has more adjustable parameters than the decoder in each channel. Explained another way, the encoder in each channel has more embedding dimensions than the decoder in each channel, e.g., the embedding dimensions of the encoder and decoder of each channel are 768 and 512, respectively, but are not limited thereto.

[0043] In accordance with certain embodiments, the encoder of each AE (of each channel) is implemented as a transformer, which is an artificial neural network (ANN) that learns context and understanding through sequential data analysis. More specifically, in accordance with certain embodiments, the encoder of each AE (of each channel) is implemented as a Vision Transformer (ViT). Where each AE is implemented as a Masked Autoencoder (MAE), there are two ViTs put end-to-end. The first ViT takes in image patches of an ECG with positional encoding, and outputs vectors representing each image patch. The second ViT takes in the vectors with positional encoding and outputs image patches again. During training of each MAE, both ViTs are used. More specifically, during training an ECG image is cut into patches, and only a percentage (e.g., 25%) of the patches are put into the first ViT. Further, during training the second ViT takes the encoded vectors and outputs a reconstruction of the full image, i.e. , the ECG image. After training, only the first ViT is used. In alternative embodiments, other types of sequence processing models can be used instead of a transformer, such as, but not limited to, a long short-term memory (LSTM) type of recurrent neural network (RNN).

[0044] In accordance with certain embodiments, each separate channel (of N channels) is assigned the task of training one of a plurality N of autoencoders, while collectively the models aim to achieve two objectives. Each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair. In specific embodiments, each of the N autoencoders is a one-dimensional (1 -D) Masked AutoEncoder (MAE) model. The aforementioned two objectives that the models aim to achieve include ensuring that an individual model (and more specifically, an individual encoder and decoder pair) can reconstruct their designated ECG channel while the collective models (and more specifically, the plurality of encoder and decoder pairs) correlate their embeddings as an approximation of the hypothetical heart states. Therefore, both reconstruction and alignment losses are jointly optimized.

[0045] Reconstruction loss measures how close the decoder output is to the original input. In other words, reconstruction loss measures how well the output of a machine learning model can reconstruct the original input data after being encoded and decoded. Accordingly, it can be appreciated that reconstruction loss quantifies thediscrepancy between the original input and its reconstructed version, often using measures such as mean squared error (MSE) for continuous data or cross-entropy loss for categorical data. Typically, in a machine learning model, the aim is to minimize this discrepancy, leading to a model that can capture and replicate essential features of the input data.

[0046] Alignment in multimodal machine learning is used for synchronizing and correlating data from different modalities so that these diverse forms of data can be effectively integrated and interpreted. With respect to the embodiments described herein, the different modalities can be ECG data sensed using different channels (aka leads). The process of alignment can involve adjusting the data from these different modalities so that they correspond to each other in a meaningful way, allowing for a more holistic analysis and decision-making by an artificial neural network. Such alignment can, for example, provide for temporal synchronization and / or semantic coherence, but is not limited thereto. Alignment loss measures distances between similar items across different modalities, such as different channels (aka leads). Techniques such as triplet loss or contrastive loss can be used, for example, to determine alignment loss. The effectiveness of alignment techniques impacts performance of downstream tasks, such as classification, prediction, and decisionmaking processes. Proper alignment may result in better data synchronization, which enhances a model’s ability to perform these tasks accurately.

[0047] In accordance with certain embodiments, a plurality N (e.g., N = 12, 13, 14, or 15, but is not limited thereto) of autoencoders are trained, using machine learning, such that a different one of a plurality N of channels of ECG data obtained for each person in a batch of persons is used to train a different one of the N autoencoders, with each autoencoder of N autoencoders including a respective encoder and decoder pair. The ECG data obtained for each person in a batch of persons are examples of input signals x[i](t). For input signals x[i](t) and predictions x[i](t) the reconstruction loss can be calculated using the mean-squared error (MSE) defined below in Equation 1 (EQ1 ), averaged over a minibatch. reconstruction loss

[0048] In accordance with certain embodiments of the present disclosure, the alignment loss aims at correlating (i.e. , making similar) the embeddings of the different ECG channels of the same heart H, while making them dissimilar for different hearts. Explained another way, during the training of each encoder and decoder pair, of the N autoencoders, the alignment losses are optimized by reducing alignment losses for embeddings produced from the N-channels of ECG data obtained for a same person in the batch of persons, and increasing alignment losses for embeddings produced from the N-channels of ECG data obtained for different persons in the batch of persons. In accordance with certain embodiments, Triplet loss defined below in Equation 2 (EQ2) is used for this purpose. alignment loss = Triplet(A, P, N) (EQ2)

[0049] In Equation 2, anchor A corresponds to all the ECG signals (aka ECG data) in the minibatch, positive P and negative N correspond to signals (aka ECG data) of different channels from the same vs. different records, respectively. Jointly optimizing these losses could potentially result in a side effect that the embeddings generated by the encoders become trivial, in that they are highly correlated with one another but unable to be used by the decoders to reconstruct the ECG signals. In order to avoid this, in accordance with certain embodiments, during the training of the autoencoders, and more generally the models, the focus is initially on reducing the reconstruction loss, so that the embeddings contain meaningful information, and thereafter the embedding are correlated through alignment. Explained another, during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, the reconstruction losses are weighted greater than the alignment losses during earlier epochs compared to during later epochs, and the alignment losses are weighted greater than the reconstruction losses during the later epochs compared to during the earlier epochs. In accordance with certain embodiments, this is performed by setting loss weights using sine and cosine functions of the epochs as defined below in Equation 3 (EQ3).Loss@epOch-i = sin ( — - - - ) x alignment loss + cos ( — - - - ) x reconstruction loss (EQ3) epochs 2 / \N epochs 2 /

[0050] In accordance with certain embodiments, the alignment objectives can be achieved by performing a split of the embeddings produced by the encoders (e.g., 102 in FIG. 1 ) of the autoencoders (e.g., 105 in FIG. 1 ). How the split is performed can depend, at least in part, on the specific encoder architecture used. For example, if the encoders 102 map sequences to sequences (as could be the case if the encoders 102 are implemented as transformers), then the embeddings (aka encodings) could be split such that vectors are chosen from all members of each of the embeddings. For example, assume that embeddings output by the encoders 102 are triply indexed values v_ij_l, which are actual real numbers, where v_ij is the jth vector in the ith embedding. More specifically, assume that embeddingl output by the encoder 102i includes a sequence of vectors (v1_1 , v1_2, v1_3, ... v_1_k), and more generally, each embeddingj output by the encoder 102i includes a sequence of vectors (vi_1 , vi_2, vi_3, ... , vi_k). In accordance with certain embodiments, the vector sequence of each embedding is split individually at some point in the vector (not across 1 to k). Each vector v_ij includes of elements v_ij_l where I runs from 1 to, e.g., 1024. If so, each vector could be split (e.g., using 50-50 split, but not limited thereto) into two parts (v_ij_2, v_ij_1 , > v_ij_512) to be used for alignment and the whole vector (v_ij_1 , v_ij_2, .... v_ij_1024) is used for reconstruction.

[0051] For a more specific example, assume that each embeddingj produced by each of the plurality of encoders 102i, during each epoch, includes a sequence of 8 vectors, with each of the 8 vectors including 1024 elements, which can be real numbers. Also assume that each embedding is split using a 50-50 split, although other splits are also possible and within the scope of the embodiments described herein. Using the 50-50 split, the alignment objectives focus on half of the 1024 elements of each of the 8 vectors, i.e., on 512 elements (out of the 1024 elements) of each of the 8 vectors. By contrast, the reconstruction objective focuses on the full extent of all of the 8 vectors, i.e., on all 1024 elements of each of the 8 vectors.

[0052] In accordance with certain embodiments, a distributed training protocol is used where MaskedAutoEncoder (MAE) models are trained separately on their individual channels, with a coordinator exchanging embeddings generated by encoders of the different channels to enable computing of alignment losses. In certain such embodiments, the MAE models are trained with a base leaning rate of 1e-3accompanied by a cosine learning rate scheduler, for 200 epochs with a batch size of 256. These are just examples, and one of skill in the art should appreciate that other quantities of epochs and batch sizes are also possible and within the scope of the embodiments described herein. Using such embodiments of the present disclosure, a unified representation for ECG signals can be generated irrespective of which channel is used as an input signal to a model.

[0053] As described below, initially with reference to the high level block diagram of FIG. 1 , certain embodiments of the present disclosure can be used to provide for modally reduced (aka channel-reduced) representation learning of multi-lead (aka multi-channel) ECG signals through simultaneous alignment and reconstruction optimization. As also described below, certain such embodiments of the present disclosure can be used to improve usefulness of ECG data obtained using a consumer grade ECG device. The term ECG data, as used herein, refers to data that represents an ECG signal that is either sensed or reconstructed. Accordingly, the term ECG data and the term ECG (or ECG signal) are often used interchangeably, since ECG data refers to data that represents an ECG (aka ECG signal).

[0054] Referring to FIG. 1 , shown therein are 13 autoencoders 105, labeled 105i, 1052, 1053, ... 105i2, 105i3, which can be collectively referred to as the autoencoders 105, and can be individually referred to as an autoencoder 105. Accordingly, in FIG. 1 , because there are 13-channels shown, N = 13 in FIG. 1. Each of the autoencoders 105 includes a respective encoder 102 and a respective decoder 104, which can be referred to collectively as an encoder and decoder pair 102, 104. For example, the autoencoder 105i includes the encoder 1021 and the decoder 104i , which can be referred to as the encoder and decoder pair 102i , 104i . For another example, the autoencoder 1052 includes the encoder 1022 and the decoder 1042, which can be referred to as the encoder and decoder pair 1022, 1042, and so on. More generally, the autoencoder 105i includes the encoder 102i and the decoder 104i, which can be referred to as the encoder and decoder pair 102i, 104i.

[0055] In the embodiments described with reference to FIG. 1 , twelve of the autoencoders 105 (i.e. , the autoencoders 105i through 105i2) are models (aka artificial neural networks) that are trained using ECG data for 12-channels sensed using a 12- channel first type of ECG recorder device, and one of the autoencoders (i.e., theautoencoder 105is), which includes the encoder 102i3 and the decoder 104i3 (aka the encoder and decoder pair 102i3, 104i3), is a model (aka artificial neural network) that is trained using ECG data for 1 -channel sensed using a 1 -channel second type of ECG recorder device (e.g., a smartwatch or other single channel wearable). More specifically, the 12 autoencoders 105i through 105i2 are each simultaneously trained using a different one of the 12-channels of ECG data obtained using a 12-channel first type of ECG recorder device (e.g., a 12-channel medical grade ECG recorder device), while the autoencoder 105i3 is simultaneously trained using 1 -channel of ECG data sensed using a 1 -channel second type of ECG recorder device (e.g., a 1 -channel consumer grade ECG recorder device, such as a smartwatch), wherein the 13- channels (and more generally, N-channels) of ECG data are sensed concurrently and correspond to a same window of time for each person in a batch of persons. The 12- channel first type of ECG recorder device discussed above is an example of a first type of ECG recorder device that senses M-channels of ECG data, wherein M is an integer that is at least 3. The 1 -channel second type of ECG recorder device discussed above is an example a second type of ECG recorder device that senses X-channels of ECG data, where X is an integer that is at least 1 and is less than M. Most likely, X will equal 1 or 2, but could potentially be as large as 3, or greater. Since the first type of ECG recorder device senses M-channels of ECG data, and the second type of ECG recorder device senses X-channels of ECG data, collectively the first and second types of ECG recorder devices sense N-channels of ECG data for each person in the batch of persons, where N = M + X. In accordance with certain embodiments M is an integer that is at least 3, and X is an integer that is at least 1 and is less than M. In block diagram shown in FIG. 1 , M = 12, X = 1 , and N = 13. In certain embodiments, the M-channel (e.g., 12-channel) first type of ECG recorder device is an M-channel medical grade ECG recorder device, and the X-channel (e.g., 1 -channel) is an X- channel consumer grade ECG recorder device.

[0056] Still referring to FIG. 1 , during the training of the 13 autoencoders 105i through 105i3, the encoder and decoder of each autoencoder 105, and thus of each channel, are trained through inter-channel alignment and intra-channel reconstruction losses. The training can occur on auto-differentiation platforms (e.g., PyTorch, TensorFlow, Keras, etc.) which can backpropagate differentiable losses through an optimization algorithm to optimize the encoders 102 and decoders 104, in order toreduce these statistical losses. This optimization process enhances the capability of each single-channel encoder 102i of each channel to capture heart semantics in the latent embedding the single-channel encoder 102i produces from its corresponding input signal 101 i, which enables these embeddings as extracted features to be used for downstream tasks. Examples of such downstream tasks include, but are not limited to identifying cardiac abnormalities (e.g., arrhythmias, myocardial infarction, heart failure, etc.), performing biometric identification, and / or performing biometric authorization, just to name a few. In order for an embodiment of the present disclosure to be used for biometric identification and / or biometric authentication of a person, the person will need to register their heart, which can be achieved by acquiring N-channels of ECG data for the person, and registering the embedding(s) from X channels that can be sensed using an X-channel second type of (e.g., consumer grade) ECG recorder device. The registered embedding(s) can later be used for identification and / or authentication, similar to how a person can register their finger print or iris for later identification and / or authentication.

[0057] In accordance with certain embodiments of the present disclosure, during the training of the 13 autoencoders 105i through 105i3 (or more generally, each of the N autoencoders 105), reconstruction losses and alignment losses are jointly optimized to achieve reconstructive objectives for each of the 13-channels (or more generally, each of the N-channels) and to collectively achieve two alignment objectives. The reconstructive objective, for each channel of the 13-channels (or more generally, each of the N-channels) is to train the encoder and decoder pair 102i, 104i, of the ithchannel such that in response to being provided with respective 1 -channel of ECG data 101 i (aka a respective input signal 101 i), the encoder 102i outputs a respective embedding 103i (aka, a respective latent representation 103i) that when input to the decoder 104 results in the decoder 104 outputting an accurate reconstruction 106i of the input signal 101 i. More specifically, each of the autoencoders 105i can be trained to minimize reconstruction errors, using a loss function such as mean squared error (MSE) between the input ECG signal 101 i and its reconstructed ECG signal 106i. An example equation for calculating the MSE was described above with reference to Equation 1 . During training, the weights of the autoencoders are adjusted to minimize the loss function, and the process is performed using backpropagation. The two alignment objectives, as alluded to above, include reducing alignment losses forembeddings produced from the 13-channels (and more generally, N-channels) of ECG data obtained for a same person in the batch of persons, and increasing alignment losses for embeddings produced from the 13-channels (and more generally, N- channels) of ECG data obtained for different persons in the batch of persons.

[0058] In accordance with certain embodiments, the aforementioned alignment losses are measured using cosine similarity, which is the cosine of an angle between two vectors, and more specifically, is the dot product of the vectors divided by the product of their lengths. Such cosine similarity does not depend on the magnitudes of the vectors, but only on their angle. The cosine similarity always belongs to the interval [-1 , 1 ], For example, two proportional vectors have a cosine similarity of 1 , two orthogonal vectors have a similarity of 0, and two opposite vectors have a similarity of -1. In certain such embodiments, the aforementioned alignment objectives are performed using stochastic gradient decent, which is an iterative method for optimizing an objective function with suitable smoothness properties.

[0059] In accordance with certain embodiments, the autoencoders of each the N channels are trained until certain specified criteria is / are satisfied, depending upon the specific implementation. Examples of such criteria relate to convergence of loss function, validation performance, reconstruction quality, and early stopping. Using the criteria related to convergence of loss function, training can be performed until reconstruction loss (e.g., calculated using mean squared error) stops decreasing significantly over several epochs, indicating that the model has learned to reconstruct the input data as accurately as possible. Using the criteria related to validation performance, training can be performed until model performance on a validation set (ECG data not seen during training) is stable and does not show signs of overfitting. Using the criteria related to reconstruction quality, training can be performed until reconstructed outputs are within a specified tolerance of the original inputs. Using the criteria related to early stopping, the training can be performed until reconstruction loss does not improve for a predefined number of epochs. Combinations of the above criteria can be used to determine when to stop the training of the models. The use of additional and / or alternative criteria are also possible and within the scope of the embodiments described herein.

[0060] Referring to FIG. 2, in accordance with certain embodiments, a coordinator 202 is used during the training of the autoencoders 105i through 105i3 (and more generally, during the training of the N autoencoders 105 of the N-channels). More specifically, during the training, each of the N autoencoders 105 provides the latent embedding that its encoder 102 outputs during each epoch to the coordinator 202. This results in the coordinator 202 receiving latent embeddings from each of the encoders 102 during each epoch. The coordinator 202 then provides each of the N autoencoders 105 with the embeddings received from the other N-1 autoencoders 105 during each epoch, so that each of the N autoencoders 105 can adjusts parameters of their respective encoder 102 to achieve the two alignment objective described above. In accordance with certain embodiments, this training is performed as data- parallel training, where N sequential calls are made and N backpropagation calls are made. Additionally, in accordance with certain embodiments, the training is performed using a distributed multi-modal (aka multi-channel) model-parallel training protocol. During each epoch of this distributed multi-modal (aka multi-channel) model-parallel training protocol, the coordinator 202 reads the input signal 101 i for each of the N- channels and proposes a mask, and sends the input signal 101 i to its appropriate channel, and more specifically, to its corresponding autoencoder 105i. Additionally, the coordinator 202 collectors or receives the computed embedding from each of the N autoencoders 105, sends to each of the N autoencoders 105 the embeddings from the other N-1 autoencoders 105, receives computed reconstruction losses (e.g., mean square errors) from each of the N autoencoders 105, and performs a snapshot request.

[0061] The high level block diagram of FIG. 3A will now be used to explain how an encoder, e.g., 102i3, trained in accordance with an embodiment of the present disclosure, can be used together with a downstream further artificial neural network (ANN) 310, to perform one or more specified tasks, such as detect cardiac abnormalities (e.g., arrythmias), perform biometric identification, and / or perform biometric authorization, but not limited thereto. Referring to FIG. 3A, shown therein is an encoder 102i (e.g., 102is) that has been pretrained along with N-1 other encoders 102 (e.g., twelve other encoders 102i through 102i2) to jointly optimize both reconstruction and alignment losses, in a manner described above with reference to FIG. 1. Also shown in FIG. 3A is task specific ANN 310, which has been trained usingsupervised learning to be perform a downstream task based on channel i embedding data output by the encoder 102i (e.g., 102is). The task specific artificial neural network (ANN) 310 can be implemented, e.g., using a multi-layer perceptron (MLP) or a deep neural network (DNN), but is not limited thereto. The downstream task that the ANN 310 is trained to perform using supervised learning can be, e.g., to detect cardiac abnormalities (e.g., arrythmias), perform biometric identification, and / or perform biometric authorization, but not limited thereto. For the following discussion, it is assumed that the downstream task is to detect one or more types of cardiac abnormalities. Accordingly, the ANN 310 can be trained based on physician adjudicated ECGs, wherein one or more physicians reviewed numerous instances of ECG data from numerous different persons and specified for each of the instances of ECG data whether or not one or more specific types of cardiac abnormalities (e.g., arrhythmias, myocardial infarction, heart failure, etc.) is present in the ECG data. More generally, the ANN 310 is trained using labeled datasets to predict outcomes and recognize patterns.

[0062] Still referring to FIG. 3A, further channel i ECG data (e.g., further channel 13 ECG data), which is data not seen during the training, is shown as being provided to the pretrained encoder 102i (e.g., 102is). Based on the further channel i ECG data (e.g., further channel 13 ECG data), the pretrained encoder 102i (e.g., 102is) outputs channel i (e.g., channel 13) embedding data, which is provided to the task specific ANN 310. Assuming the task specific ANN 310 has been trained using supervised learning to detect one or more types of cardiac abnormalities, the task specific ANN 310 outputs a prediction (aka detection) in response to receiving the channel i (e.g., channel 13) embedding data from the pretrained encoder 102i (e.g., 102is).

[0063] In accordance with certain embodiments, the further channel i ECG data (e.g., further channel 13 ECG data) is obtained (and more specifically, sensed) by a consumer grade ECG recorder device, which can be a wearable (e.g., a smartwatch or a body worn patch). In such an embodiment, the pretrained encoder 102i (e.g., 102i3) and the task specific ANN 310 can both be deployed on the consumer grade ECG recorder device, e.g., by one or more processors thereof. Alternatively, the pretrained encoder 102i (e.g., 102is) and the task specific ANN 310 can both be deployed on another consumer grade device (e.g., a smartphone or tablet computer)that is communicatively coupled to (e.g., using Bluetooth, or the like) the consumer grade ECG recorder device, e.g., by one or more processors of the other consumer grade device (e.g., the smartphone or tablet computer). Electrodes coupled to sensing circuitry (e.g., a differential amplifier, and / or the like) of the consumer grade ECG recorder device can be used to sense an ECG signal. An analog-to-digital (ADC) converter of the consumer grade ECG recorder device can be used to digitize the ECG signal, such that the ECG data that is provided to the pretrained encoder 102i (e.g., 10213) of the consumer grade ECG recorder device is digital data. In response to receiving the ECG data, the pretrained encoder 102i (e.g., 102is) outputs channel i (e.g., channel 13) embedding data, which is provided to the ANN 310 that had been pretrained using supervised learning. In response to receiving the channel i (e.g., channel 13) embedding data, the ANN 310 outputs detections (aka predictions) based on the embedding data. The detections (aka predictions) output by the ANN 310 can be saved in memory of the consumer grade ECG recorder device. Additionally, or alternatively, where the consumer grade ECG recorder device is a smartwatch or other type of device that includes a display and / or a speaker, or is communicative coupled (e.g., using Bluetooth, or the like) to another device (e.g., a smartphone) that includes a display and / or a speaker, the detections (aka predictions) output by the ANN 310 can be used to generate alerts (aka alarms), which may instruct a person to visit or otherwise contact their physician. Such alerts (aka alarms) may also instruct a person to sit down, or stop operating a vehicle or other types of machinery. If the ANN 310 is alternatively, or additionally, used to perform biometric identification and / or biometric authentication, then the ANN 310 can be used to provide or deny a person access to certain content and / or capabilities.

[0064] In the embodiments described above with reference to FIG. 3A, it was assumed that the consumer grade ECG recorder device with which the embodiments are used is only capable of obtaining 1 -channel (aka a single-channel) of ECG data, which is the case with many wearables, such as smartwatches. In the embodiments described below with reference to FIG. 3B, it is assumed that consumer grade ECG recorder device(s) with which the embodiments are used are capable of obtaining 2- channels (or more generally, multiple channels) of ECG data.

[0065] Referring to FIG. 3B, shown therein is are two encoder 102i and 102j that have been pretrained along with N-2 other encoders 102 (e.g., 12 other encoders 102) to jointly optimize both reconstruction and alignment losses, in a manner described above with reference to FIG. 1 . Also shown in FIG. 3B is the ANN 310, which has been trained using supervised learning to be perform a downstream task based on both channel i embedding data output by the encoder 102i and channel j embedding data output by the encoder 102j using supervised learning. Examples details of the supervised learning and the types of tasks that may be performed by the ANN 310 were described above in the discussion of FIG. 3A, and thus need not be described again. In certain embodiments, the channel i embedding data output by the encoder 102i and channel j embedding data output by the encoder 102j are concatenated or added prior to be provided to the ANN 310, depending upon how the ANN 310 was trained.

[0066] The high level flow diagram of FIG. 4A will now be used to summarize computer implemented methods, according to certain embodiments of the present disclosure. Additional details of the methods summarized below with reference to FIG. 4A (as well as with reference to the high level flow diagram of FIG. 4B) can be appreciated from the above discussion of FIGS. 1 , 2, 3A, and 3B.

[0067] Referring to FIG. 4A, step 402 involves obtaining, for each person in a batch of persons, ECG data for each of N-channels sensed using one or more ECG recorder devices, wherein for each person, in the batch of persons, the ECG data sensed for each of N-channels are sensed concurrently and correspond to a same window of time. In certain embodiments, N is an integer that is at least 3. In certain embodiments, the ECG data sensed for each of N-channels, which is obtaining at step 402, is sensed using N sensing vectors, each of which includes a different pair of electrodes that are communicatively coupled to a respective sense amplifier. As will be explained in more detail below with reference to FIG. 4B, and as can be appreciated from the above discussion of FIG. 1 , some of the N-channels of ECG data can be sensed using a first type of ECG recorder device (e.g., an M-channel medical grade ECG recorder device), and one or more of the N-channel of ECG data can be sensed using a second type of ECG reorder device (e.g., an X-channel consumer grade ECG recorder device, such as a smartwatch).

[0068] Still referring to FIG. 4A, step 404 involves training, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons, wherein each autoencoder of the N autoencoders comprises an artificial neural network (ANN) model that includes a respective encoder and decoder pair, wherein reconstruction losses and alignment losses are jointly optimized during the training of each autoencoder of the N autoencoders. Referring briefly back to FIG. 1 , the autoencoders 105, each of which includes a respective encoder 102 and a respective decoder 104, are examples of the N autoencoders trained at step 404. Further, as was described above in the discussion of FIG. 1 , in accordance with certain embodiments, the N autoencoder trained at step 404 are trained using masked autoencoding.

[0069] Referring again to FIG. 4A, in accordance with certain embodiments, during the training of each encoder and decoder pair, of the N autoencoders, at step 404, the reconstruction losses are optimized by reducing alignment losses for embeddings produced from the N-channels of ECG data obtained for a same person in the batch of persons, and increasing alignment losses for embeddings produced from the N- channels of ECG data obtained for different persons in the batch of persons. In certain such embodiments, the alignment losses are calculated using mean-squared error (MSE), e.g., using Equation 1 discussed above.

[0070] Further, in accordance with certain embodiments, during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, each encoder of the N autoencoders produces respective embeddings that are shared with other encoders of the N autoencoders to enable the alignment losses to be calculated and optimized. Such sharing can be performed, for example, using the coordinator 202 discussed above with reference to FIG. 2. Additionally, in accordance with certain embodiments, during the training of each encoder and decoder pair, of the N autoencoders, the reconstruction losses are weighted greater than the alignment losses during earlier epochs compared to during later epochs, and the alignment losses are weighted greater than the reconstruction losses during the later epochs compared to during the earlier epochs.

[0071] Referring again to FIG. 4A, after the N autoencoders have been trained such that reconstruction losses and alignment losses are jointly optimized, step 406 involves obtaining further ECG data for each of a subset of the N-channels, wherein the subset can include as few as one of the N-channels. Thereafter, step 408 involves using a subset of the trained encoders, of the N autoencoders that have been trained, to generate at least one embedding based on the further ECG data obtained at step 406. The subset of the trained encoders, used at step 408, can include as few as one of the trained encoders.

[0072] Still referring to FIG. 4A, step 410 involves providing the at least one embedding (generated at step 408) to a further artificial neural network (ANN) model that had been trained using supervised learning to perform one or more specific tasks, such as detect one or more cardiac abnormalities (e.g., arrythmias), perform biometric identification, and / or perform biometric authorization, but not limited thereto. Step 412 involves using the further ANN model, to which the embedding(s) was / were provided at step 410, to perform one or more specific tasks, such as detecting cardiac abnormalities (e.g., arrythmias), performing biometric identification, and / or perform biometric authorization, but not limited thereto. Referring briefly back to FIGS. 3A and 3B, the task specific ANNs 310 shown therein are examples of the further ANN referred to in steps 410 and 412.

[0073] The high level flow diagram of FIG. 4B will now be used to describe how the computer implemented methods initially summarized with reference to FIG. 4A can be used to improve the usefulness of ECG data obtained using a consumer grade ECG device, such as a smartwatch or a patch. Further, as will be described in additional detail below, with reference to the high level flow diagram of FIG. 4B, the ECG data obtained for each of N-channels at step 402 can include ECG data obtained for each of M-channels sensed using an M-channel first type of (e.g., medical grade) ECG recorder device, and ECG data obtained for each of X-channels sensed using an X- channel second type of (e.g., consumer grade) ECG recorder device, where M + X = N. In certain such embodiments, M is an integer that is at least 3, X is an integer that is at least 1 and is less than M, and M + X = N. For example, referring back to the high level block diagram of FIG. 1 , in FIG. 1 M = 12, and X = 1 , and thus N = 13. More specifically, in FIG. 1 , ECG data is obtained for each of 12-channels sensed using a12-channel first type of ECG recorder device (e.g., a 12-channel medical grade ECG recorder device), and an additional channel of data is sensed using a 1 -channel second type of ECG recorder device (e.g., a 1 -channel consumer grade ECG recorder device), and thus, 13-channels of ECG data are obtained for each person in the batch of persons. The batch of persons preferably includes many thousands of person, e.g., 100,000 persons, or more, but is not limited thereto.

[0074] Referring now to FIG. 4B, steps 402’, 404’, 406’, and 408’ in FIG. 4B, are specific embodiments of steps 402, 404, 406, and 408, respectively, which were discussed above with reference to FIG. 4A. Referring to FIG. 4B, step 402’ involves obtaining, for each person in the batch of persons, ECG data for each of M-channels sensed using an M-channel first type of ECG recorder device (e.g., an M-channel medical grade ECG recorder device), and ECG data for each of X-channels sensed using an X-channel second type of ECG recorder device (e.g., an X-channel consumer grade ECG recorder device), wherein M is an integer that is at least 3, and X is an integer that is at least 1 and is less than M. In such embodiments, for each person in the batch of persons, the ECG data sensed using the M-channel first type of ECG recorder device and the ECG data sensed using the X-channel second type of ECG recorder device are sensed concurrently and correspond to the same window of time. Further, as noted above, M + X = N, and thus, at step 402’, N-channels of ECG data are obtained for each person in the batch of persons. However, it is noted that for different persons (or subsets of persons), in the batch of persons, the ECG data obtained for each of M-channels and each of the X-channels can be sensed using different instances of the M-channel first type of ECG recorder device and different instances of the X-channel second type of ECG recorder device. For example, at step 402’, a first instance of an M-channel medical grade ECG recorder device and a first instance of ab X-channel medical grade ECG recorder device (e.g., a specific model of a smartwatch) that are co-located in a first facility (e.g., in a first city or town) can be used to sense a first subset of the N-channels of ECG data obtained at step 402’, a second instance of the M-channel medical grade ECG recorder device and a second instance of the X-channel consumer grade ECG recorder device that are co-located in a second facility (e.g., in a second city or town) can be used to sense a second subset of the N-channels of ECG data obtained at step 402’, etc. In certain embodiments, M = 12 or 14. In certain embodiments, X = 1 or 2. Further, it is notedthat for different persons in the batch of persons, the ECG data sensed using the M- channel first type of ECG recorder device and the ECG data sensed using the X- channel second type of ECG recorder device can correspond to different windows of time and need not be sensed concurrently.

[0075] Still referring to FIG. 4B, step 404’ involves training, using machine learning, M autoencoders using the ECG data sensed using the M-channel first type of ECG recorder device, and X autoencoder(s) using the ECG data sensed using the X- channel second type of ECG recorder device. For example, referring briefly back to FIG. 1 , and assuming M = 12, and X = 1 , step 404’ can involve training the autoencoders 105i through 105i2 using the ECG data sensed using a 12-channel first type of ECG recorder device, and training the autoencoder 105i3 using the ECG data sensed using the 1 -channel second type of ECG recorder device. Such a 12-channel first type of ECG recorder device can include or be communicatively coupled to 10 skin electrodes, which are known in the art as the RA, LA, RL, LL, V1 , V2, V3, V4, V5, and V6 electrodes, as was discussed above in the Background. The 1 -channel second type of ECG recorder device can be, e.g., a smartwatch, or a patch, but is not limited thereto.

[0076] Referring again to FIG. 4B, after the training of the N autoencoders, step 406’ involves obtaining at least one channel of further ECG data using the X-channel (e.g., 1 -channel) second type of ECG recorder device. Step 408’ involves using at least one of the encoders of the X autoencoder(s) to generate at least one embedding based on the at least one channel of further ECG data obtained at step 406’. Step 410’ involves providing the at least one embedding, generated at step 408’, to a further ANN model (e.g., 310 in FIG. 3A or 3B) to perform one or more specific task, e.g., to detect a cardiac abnormality, perform biometric identification, and / or perform biometric authentication, but not limited thereto.

[0077] Referring back to FIG. 1 again, in the example embodiment shown therein, M = 12, X = 1 , and N = 12 (since 12 + 1 = 13). In another embodiment, M = 12, and X = 2, and N = 14. In still another embodiment, M = 12, X = 2, and N = 14. In a further embodiment, M = 14, X = 1 , and N = 15. In still another embodiment, M = 14, X = 2, and N = 16. Other values for M, X, and N are also possible and within the scope of the embodiments described herein. For example, if it is known that a specific subset ofECG channels (e.g., two specific ECG channels) of a 12-channel medical grade ECG recorder device are the most relevant for detecting a specific type of cardiac abnormality, or some other type of detection, then it is possible that M can be as small as 2, in which case if X = 1 , then N can be as small as 3 (since N = M + X = 2 + 1 = 3).

[0078] As can be appreciated from the above description, embodiments described herein can be used to improve the usefulness of ECG data obtained from a consumer grade ECG recorder device, making such a second type of ECG recorder device just as good, or almost as good, as a medical grade ECG recorder device for certain types of detections, predictions and / or other tasks. Certain such embodiments can enable, e.g., reconstructed 12-channel ECG data to be generated and sent to a physician to analyze, based on X-channel(s) of ECG data obtained using an X-channel consumer grade ECG recorder device. Alternatively, or additionally, certain such embodiments can be used by models and / or other types of algorithms to perform disease prediction (e.g., arrythmia detection), as can be appreciated from the above description. Such embodiments achieve mode reduction (aka channel reduction) and provide for embeddings that represent a heart, such that if ECG data for one channel (or just a few channels) is / are obtained, corresponding embeddings that represent the heart provide enough representational power that disease prediction (and other types of predictions) is almost as good as can be achieve using 12-channels of ECG data.

[0079] FIG. 5A illustrates an example 12-channel medical grade ECG recorder device 502. FIG. 5B illustrates an example 1 -channel consumer grade ECG recorder device 512, which in this example is a smartwatch that is configured to be strapped on a person’s wrist. FIG. 5C illustrates an example 1 - or 2-channel consumer grade ECG recorder device 522, which in this example is a patch the is configured to adhered to a person’s chest. Details of such types of ECG recorder devices are well known in the art, and thus, need not be described in detail herein. Nevertheless, for completeness, it noted that such devices each include or are communicatively coupled to electrodes, which together with one or more sense amplifiers, can be used to sense respective ECG signals, and more generally, channels of ECG data. Each such device can also include a controller, e.g., implemented using a processor, to control the sensing of ECG data. Additionally, each such device can include memory for storing ECG data.Further, each such device can include telemetry circuitry for transferring ECG data, and / or detections or predictions based thereon, to another device. Optionally, each such device can include a display that can display ECG segments, detections or predictions, alerts, and / or the like. The controller of each such device can be used to perform steps and / or other functions of the embodiments described herein.

[0080] Certain embodiments of the present disclosure can be used to synthesize multi-lead ECG signals (or more generally, M-channels of ECG data) from a single channel ECG signal, by performing multi-modal (aka multi-channel) alignment and reconstruction of the multi-channel ECG data so that a single channel encoding (or more generally, an X-channel encoding) can be used as extracted feature for signal reconstruction and / or down-stream classification or other similar tasks. Certain embodiments of the present disclosure use these capabilities, along with a novel embedding alignment strategy, to correlate different ECG channel information and obtain a universal, channel agnostic representation of a heart. The aligned embedding of single (or in some cases, a smaller subset of channels than medical-grade 12 or 14 channel ECGs) can then be used as extracted latent features for down-stream tasks to detect arrythmias, provide a myocardial infarction diagnosis, perform activity and / or emotion monitoring, detect one or more biometrics, coronary heart disease, heart attacks, cardiomyopathy, and / or the like.

[0081] On an independent test dataset, embodiments of the present disclosure were used to generate highly correlated channel embeddings from very different ECG channels, leading to a moderate approximation of the 12-channel ECG signals from a single-channel embedding. Using embodiments described herein, embeddings can be generated and used as strong features for ECG signals for performing one or more downstream tasks, such as, but not limited to, detecting cardiac abnormalities (e.g., arrythmias), performing biometric identification, and / or performing biometric authorization.

[0082] A person’s heart is a single identifiable organ producing multi-channel (aka, multi-modal) ECG signals that are related. Through multi-modal (and more precisely, multi-channel or homogenous multi-modal) alignment correlation (of embeddings) and encoder / decoder reconstruction objectives, high-quality single-channel embeddings can be obtained as a representational multi-dimensional feature tensor. Certain suchembodiments involve correlating representations obtained from the different ECG signal channels. Certain such embodiments strengthen mode-alignment with reconstruction targets, where aligned encodings can be used to reconstruct a corresponding mode’s signal or another mode’s signal. With such embodiments, a single-channel signal encoder of a heart can provide for adequate features for downstream classification tasks related to heart mal-condition, identification, etc.

[0083] Certain embodiments of the present disclosure enable use of minimal mode (aka channel) disease detector, which allows consumers to be warned immediately if the more compact single-channel (or other minimal channel) profile’s probes are available on their device.

[0084] In accordance with certain embodiments, specialized data can also be gathered for the consumer device probes to accompany other data used for training stages in order to enhance the exact channel encoder used. Such an embodiment can also be applied to any similar situation involving type-homogenous multi-modal signals from the same object of interest. When used with ECG signals corresponding to a heart, all the ECG signals are essentially of the same signal type - hence, this is an example of homogenous multi-modal (e.g., multi-channel) mode reduction which can be repeated in similar other situations. For example, the techniques described herein can also be used for constructing mode-reduced devices in heterogenous multi-modal situations, e.g., multiple signals from the lungs (x-ray images, capacity, oxygen levels, etc.). In general, certain embodiments described herein can be used to construct enhanced mode-reduced devices in all kinds of medical applications.

[0085] FIGS. 6 and 7 illustrate example hardware devices suitable for implementing embodiments of present disclosure described herein.

[0086] FIG. 6 shows an example of a computing system 600 with which embodiments disclosed herein may be implemented. For example, the computing system 600 can be used to implement the autoencoders 105 and the ANNs 310, but not limited thereto. The computing system 600 includes at least one processor 602, which could be a single central processing unit (CPU) or an arrangement of multiple processor cores of a multi-core architecture. In the depicted example, the processor 602 includes a pipeline 604, an instruction cache 606, and a data cache 608 (andother circuitry, not shown). The processor 602 is connected to a processor bus 610, which enables communication with an external memory system 612 and an input / output (I / O) bridge 614. The I / O bridge 614 enables communication over an I / O bus 616, with various different I / O devices 618A-618D (e.g., disk controller, network interface, display adapter, and / or user input devices such as a keyboard or mouse). Optionally, in embodiments, device 600 may include both a processor 600 and a dedicated graphics processing unit (GPU, not shown.) As understood, a GPU is a type of processing unit that enables very efficient parallel processing of data. Although GPUs may be used in a video card or the like for computer graphics, GPUs have found much broader applications.

[0087] The external memory system 612 is part of a hierarchical memory system that includes multi-level caches, including the first level (L1 ) instruction cache 606 and data cache 608, and any number of higher level (L2, L3, . . . ) caches within the external memory system 612. Other circuitry (not shown) in the processor 602 supporting the caches 606 and 608 includes a translation lookaside buffer (TLB), various other circuitry for handling a miss in the TLB or the caches 606 and 608. For example, the TLB is used to translate an address of an instruction being fetched or data being referenced from a virtual address to a physical address, and to determine whether a copy of that address is in the instruction cache 606 or data cache 608, respectively. If so, that instruction or data can be obtained from the L1 cache. If not, that miss is handled by miss circuitry so that it may be executed from the external memory system 612. It is appreciated that the division between which level caches are within the processor 102 and which are in the external memory system 612 can differ in various examples. For example, an L1 cache and an L2 cache may both be internal and an L3 (and higher) cache could be external. The external memory system 612 also includes a main memory interface 620, which is connected to any number of memory modules (not shown) serving as main memory (e.g., Dynamic Random Access Memory modules).

[0088] FIG. 7 illustrates a schematic diagram of a general-purpose network component or computer system 700 with which embodiments of the present disclosure may be implemented. For example, general-purpose network component or computer system 700 can be used to implement the autoencoders 105 and the further ANNs310, but not limited thereto. FIG. 7 illustrates a schematic diagram of a general- purpose network component or computer system. The general-purpose network component or computer system 700 includes a processor 702 (which may be referred to as a central processor unit or CPU) that is in communication with memory devices including secondary storage 704, and memory, such as ROM 706 and RAM 708, input / output (I / O) devices 710, and a network 712, such as the Internet or any other well-known type of network, that may include network connectively devices, such as a network interface. Although illustrated as a single processor, the processor 702 is not so limited and may comprise multiple processors. The processor 702 may be implemented as one or more CPU chips, cores (e.g., a multi-core processor), FPGAs, ASICs, and / or DSPs, and / or may be part of one or more ASICs. The processor 702 may be configured to implement any of the schemes described herein. The processor 702 may be implemented using hardware, software, or both.

[0089] The secondary storage 704 is typically comprised of one or more disk drives or tape drives and is used for non-volatile storage of data and as an over-flow data storage device if the RAM 708 is not large enough to hold all working data. The secondary storage 704 may be used to store programs that are loaded into the RAM 708 when such programs are selected for execution. The ROM 706 is used to store instructions and perhaps data that are read during program execution. The ROM 706 is a non-volatile memory device that typically has a small memory capacity relative to the larger memory capacity of the secondary storage 704. The RAM 708 is used to store volatile data and perhaps to store instructions. Access to both the ROM 706 and the RAM 708 is typically faster than to the secondary storage 1404. At least one of the secondary storage 704 or RAM 708 may be configured to store routing tables, forwarding tables, or other tables or information disclosed herein.

[0090] It is understood that by programming and / or loading executable instructions onto the node 700, at least one of the processor 720 or the memory 722 are changed, transforming the node 700 in part into a particular machine or apparatus (aka system), e.g., a router, having the novel functionality taught by the present disclosure. Similarly, it is understood that by programming and / or loading executable instructions onto the node device 700, at least one of the processor 702, the ROM 1406, and the RAM 708 are changed, transforming the node 700 in part into a particular machine or apparatus,e.g., a router, having the novel functionality taught by the present disclosure. Optionally, in embodiments, device 700 may include both a processor 702 and a dedicated GPU, not shown.

[0091] It is fundamental to the electrical engineering and software engineering arts that functionality that can be implemented by loading executable software into a computer can be converted to a hardware implementation by well-known design rules. Decisions between implementing a concept in software versus hardware typically hinge on considerations of stability of the design and numbers of units to be produced rather than any issues involved in translating from the software domain to the hardware domain. Generally, a design that is still subject to frequent change may be preferred to be implemented in software, because re-spinning a hardware implementation is more expensive than re-spinning a software design. Generally, a design that is stable that will be produced in large volume may be preferred to be implemented in hardware, for example in an ASIC, because for large production runs the hardware implementation may be less expensive than the software implementation. Often a design may be developed and tested in a software form and later transformed, by well-known design rules, to an equivalent hardware implementation in an application specific integrated circuit that hardwires the instructions of the software. In the same manner as a machine controlled by a new ASIC is a particular machine or apparatus (aka system), likewise a computer that has been programmed and / or loaded with executable instructions may be viewed as a particular machine or apparatus.

[0092] The embodiments described herein can be implemented using hardware, firmware, software, or a combination of these. The software used is stored on one or more of the processor readable storage devices described above to program one or more of the processors to perform the functions described herein. The processor readable storage devices can include computer readable media such as volatile and non-volatile media, removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer readable storage media and communication media. Computer readable storage media may be implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computerreadable storage media include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by a computer. A computer readable medium or media does (do) not include propagated, modulated or transitory signals.

[0093] Communication media typically embodies computer readable instructions, data structures, program modules or other data in a propagated, modulated or transitory data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” means a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as RF and other wireless media. Combinations of any of the above are also included within the scope of computer readable media.

[0094] In alternative embodiments, some or all of the software can be replaced by dedicated hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include Field-programmable Gate Arrays (FPGAs), Application-specific Integrated Circuits (ASICs), Applicationspecific Standard Products (ASSPs), System-on-a-chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), special purpose computers, etc. In one embodiment, software (stored on a storage device) implementing one or more embodiments is used to program one or more processors. The one or more processors can be in communication with one or more computer readable media / storage devices, peripherals and / or communication interfaces.

[0095] It is understood that the present subject matter may be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this subject matter will be thorough and complete and will fully convey the disclosure to those skilled in the art. Indeed, the subject matter is intended to cover alternatives, modifications and equivalents of these embodiments, which are included within the scope and spirit of the subject matter as defined by the appended claims. Furthermore, in the followingdetailed description of the present subject matter, numerous specific details are set forth in order to provide a thorough understanding of the present subject matter. However, it will be clear to those of ordinary skill in the art that the present subject matter may be practiced without such specific details.

[0096] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatuses (systems) and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable instruction execution apparatus, create a mechanism for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0097] The description of the present disclosure has been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to the disclosure in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the disclosure. The aspects of the disclosure herein were chosen and described in order to best explain the principles of the disclosure and the practical application, and to enable others of ordinary skill in the art to understand the disclosure with various modifications as are suited to the particular use contemplated.

[0098] For purposes of this document, each process associated with the disclosed embodiments may be performed continuously and by one or more computing devices. Each step in a process may be performed by the same or different computing devices as those used in other steps, and each step need not necessarily be performed by a single computing device.

[0099] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific featuresor acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

CLAIMSWhat is claimed is:1 . A computer implemented method, comprising: obtaining, for each person in a batch of persons, electrocardiogram (ECG) data for each of N-channels sensed using one or more ECG recorder devices; wherein for each person, in the batch of persons, the ECG data sensed for each of N-channels are sensed concurrently and correspond to a same window of time; and wherein N is an integer that is at least 3; training, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons; wherein each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair; and wherein reconstruction losses and alignment losses are jointly optimized during the training of each autoencoder of the N autoencoders.

2. The computer implemented method of claim 1 , wherein the obtaining, for each person in the batch of persons, the ECG data for each of the N-channels sensed using the one or more ECG recorder devices, comprises: obtaining, for each person in the batch of persons:ECG data for each of M-channels sensed using an M-channel first type of ECG recorder device; andECG data for each of X-channels sensed using an X-channel second type of ECG recorder device; wherein, for each person in the batch of persons, the ECG data sensed using the M-channel first type of ECG recorder device and the ECG data sensed using the X-channel second type of ECG recorder device are sensed concurrently and correspond to the same window of time; wherein M is an integer that is at least 2; wherein X is an integer that is at least 1 and is less than M; andwherein M + X = N, and thus, the N-channels of ECG data are obtained for each person in the batch of persons.

3. The computer implemented method of claim 2, wherein: the N autoencoders include:M autoencoders trained using the ECG data sensed using the M-channel first type of ECG recorder device; andX autoencoder(s) trained using the ECG data sensed using the X-channel second type of ECG recorder device; and after the training of the N autoencoders, the method further comprises: obtaining further ECG data for each of at least one of the X-channels using the X-channel second type of ECG recorder device; using at least one of the encoders of the X autoencoder(s) to generate at least one embedding based on the further ECG data; and providing the at least one embedding to a further artificial neural network model to at least one of detect a cardiac abnormality, perform biometric identification, or perform biometric authentication.

4. The computer implemented method of any one of claims 1 through 3, wherein during the training of each encoder and decoder pair, of the N autoencoders, the alignment losses are optimized by: reducing alignment losses for embeddings produced from the N-channels of ECG data obtained for a same person in the batch of persons; and increasing alignment losses for embeddings produced from the N-channels of ECG data obtained for different persons in the batch of persons.

5. The computer implemented method of any one of claims 1 through 4, wherein during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, each encoder of the N autoencoders produces respective embeddings that are shared with other encoders of the N autoencoders to enable the alignment losses to be calculated and optimized.

6. The computer implemented method of any one of claims 1 through 5, wherein during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, the reconstruction losses are weighted greater than the alignment losses during earlier epochs compared to during later epochs, and the alignment losses are weighted greater than the reconstruction losses during the later epochs compared to during the earlier epochs.

7. The computer implemented method of any one of claims 1 through 6, wherein the N autoencoders are trained using masked autoencoding.

8. The computer implemented method of any one of claims 2 through 7, wherein the X-channel second type of ECG device comprises a wearable device.

9. The computer implemented method of any one of claims 2 through 8, wherein: M = 12;X = 1 ; andN = 13.

10. The computer implemented method of any one of claims 2 through 9, wherein: M = 12 or 14;X = 1 or 2; andN = 13, 14, 15, or 16.

11. A system, comprising: a storage medium comprising computer instructions; and one or more processors coupled to communicate with the storage medium, wherein the one or more processors execute the instructions to cause the system to: obtain, for each person in a batch of persons, electrocardiogram (ECG) data for each of N-channels sensed using one or more ECG recorder devices; wherein for each person, in the batch of persons, the ECG data sensed for each of N-channels are sensed concurrently and correspond to a same window of time; andwherein N is an integer that is at least 3; and train, using machine learning, each autoencoder of N autoencoders such that each autoencoder is trained using a different one of the N-channels of ECG data obtained for each person in the batch of persons; wherein each autoencoder of the N autoencoders comprises an artificial neural network model that includes a respective encoder and decoder pair; and wherein reconstruction losses and alignment losses are jointly optimized during the training of each autoencoder of the N autoencoders.

12. The system of claim 11 , wherein to obtain, for each person in the batch of persons, the ECG data for each of the N-channels sensed using the one or more ECG recorder devices, the one or more processors execute the instructions to cause the system to: obtain, for each person in the batch of persons:ECG data for each of M-channels sensed using an M-channel first type of ECG recorder device; andECG data for each of X-channels sensed using an X-channel second type of ECG recorder device; wherein, for each person in the batch of persons, the ECG data sensed using the M-channel first type of ECG recorder device and the ECG data sensed using the X-channel second type of ECG recorder device are sensed concurrently and correspond to the same window of time; wherein M is an integer that is at least 2; wherein X is an integer that is at least 1 and is less than M; and wherein M + X = N, and thus, the N-channels of ECG data are obtained for each person in the batch of persons.

13. The system of claim 12, wherein: the N autoencoders include:M autoencoders trained using the ECG data sensed using the M-channel first type of ECG recorder device; andX autoencoder(s) trained using the ECG data sensed using the X-channel second type of ECG recorder device; and after the N autoencoders have been trained, the one or more processors execute the instructions to cause the system to: obtain further ECG data for each of at least one of the X-channels using the X- channel second type of ECG recorder device; use at least one of the encoders of the X autoencoder(s) to generate at least one embedding based on the further ECG data; and provide the at least one embedding to a further artificial neural network model to at least one of detect a cardiac abnormality, perform biometric identification, or perform biometric authentication.

14. The system of any one of claims 11 through 13, wherein during the training of each encoder and decoder pair, of the N autoencoders, the one or more processors execute the instructions to cause the system to optimize the alignment losses by: reducing alignment losses for embeddings produced from the N-channels of ECG data obtained for a same person in the batch of persons; and increasing alignment losses for embeddings produced from the N-channels of ECG data obtained for different persons in the batch of persons.

15. The system of any one of claims 11 through 14, wherein during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, each encoder of the N autoencoders produces respective embeddings that are shared with other encoders of the N autoencoders to enable the alignment losses to be calculated and optimized.

16. The system of any one of claims 11 through 15, wherein during the training of each encoder and decoder pair, of the N autoencoders, during which the reconstruction losses and the alignment losses are jointly optimized, the one or more processors execute the instructions to cause the reconstruction losses to be weighted greater than the alignment losses during earlier epochs compared to during laterepochs, and the alignment losses to be weighted greater than the reconstruction losses during the later epochs compared to during the earlier epochs.

17. The system of any one of claims 11 through 16, wherein the N autoencoders are trained using masked autoencoding.

18. The system of any one of claims 12 through 17, wherein the X-channel second type of ECG device comprises a wearable device.

19. The system of any one of claims 12 through 18, wherein:M = 12;X = 1 ; andN = 13.

20. The system of any one of claims 12 through 19, wherein:M = 12 or 14;X = 1 or 2; andN = 13, 14, 15, or 16.

Images