DESC:FORM 2

THE PATENTS ACT, 1970
(39 of 1970)
&
THE PATENT RULES, 2003

COMPLETE SPECIFICATION
(See Section 10 and Rule 13)

Title of invention:
RECONSTRUCTION OF BODY SURFACE POTENTIAL FROM 12-LEAD ELECTROCARDIOGRAM USING A CONDITIONAL GENERATIVE ADVERSARIAL NETWORK ARCHITECTURE

Applicant:
Tata Consultancy Services Limited
A company Incorporated in India under the Companies Act, 1956
Having address:
Nirmal Building, 9th Floor,
Nariman Point, Mumbai 400021,
Maharashtra, India

The following specification describes the invention and the manner in which it is to be performed.
CROSS-REFERENCE TO RELATED APPLICATIONS AND PRIORITY
[001] The present application claims priority from Indian provisional patent application no. 202321008155, filed on February 08, 2023. The entire contents of the aforementioned application are incorporated herein by reference.

TECHNICAL FIELD
The disclosure herein generally relates to the field of biomedical signal reconstruction, and, more particularly, to reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network architecture.

BACKGROUND
Cardiovascular Diseases (CVD) have a huge prevalence and are regarded as the leading cause of death globally. Among the vast plethora of CVD diagnosis tools, Electrocardiogram (ECG) remains the most commonly conducted procedure. A 12-lead ECG is a standard assessment for cardiac disorders. The 12-lead ECG is used for screening as well as monitoring, and aiding healthcare professionals to provide both prevention and treatment. However, despite being a first level diagnostic tool, the 12-lead ECG lacks spatial resolution and is insufficient in arrhythmia localization, cardiac activation pattern mapping through myocardium, and atrial and ventricular activation abnormalities.
Electrical activity at myocardium level and successive activation mapping can be reconstructed non-invasively from dense Body Surface Potential Maps (BSPM). BSP, like conventional 12-lead ECG, measures electrical potential of heart at body surface but employs relatively large number of electrodes (often 50 to 300) distributed throughout a thorax surface. Dense distribution of electrodes results in higher degree of accuracy in detecting cardiac conditions and source level abnormalities. However, in spite of the rich information derived in terms of diagnostic yield, BSP is still not used widely in clinical practices. Main reason for its non-inclusion in standard medical practice is management of huge number of leads spread across torso surface. Standardization of these leads in terms of placement, signal to noise ratio, type of electrodes, and/or the like are generic hindrance that has prevented use of BSP over the standard 12-lead ECG. BSP electrodes are often made available in the form of electrode vests, but custom electrode manufacturing across multiple institutions leads to incompatible electrode interfaces apart from high cost associated. Selection of lead numbers and location also varies among manufactures and determining optimal number of electrodes required on the torso surface to regenerate cardiac activation maps is also a challenge.
Conventionally, the limitations associated with managing large number of physical electrodes and yet generate dense cardiac activation information are overcome with use of a partial and optimal set of body surface electrodes and reconstruct complete BSPM data matrix synthetically, often using Machine Learning (ML) techniques. In the conventional approaches, ML is coupled with information from 12-lead ECG and have been mostly utilized to localize ventricular activation origin, and improve efficacy of localizing Atrial Fibrillation (AF) ablation sites. There exist a few approaches for reconstruction of BSPM from a reduced set of electrodes by using deep generative models such as Variational Autoencoders (VAE). In VAE, the encoding distribution is regularized to generate new BSP data from the latent space, but the newly generated BSP data is often noisy.
BRIEF DESCRIPTION OF THE DRAWINGS
The accompanying drawings, which are incorporated in and constitute a part of this disclosure, illustrate exemplary embodiments and, together with the description, serve to explain the disclosed principles:
FIG. 1 illustrates an exemplary system for reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network architecture according to some embodiments of the present disclosure.
FIG. 2, with reference to FIG. 1, illustrates an exemplary flow diagram illustrating a method for reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network architecture, using the system of 100 of FIG. 1, in accordance with some embodiments of the present disclosure.
FIGS. 3A and 3B show electrode distribution of a standard 12-lead ECG configuration and a Nijmegen BSP configuration according to some embodiments of the present disclosure.
FIG. 4, with reference to FIGS. 1-3, depicts a functional block diagram of the generator model of the trained time-series conditional generative adversarial network (TSGAN) architecture as implemented by the system of 100 of FIG. 1, in accordance with an embodiment of the present disclosure.
FIGS. 5A an 5B depict graphical illustrations of cardiac electrical activity recorded at torso from lead II and lead V3 for reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network, in accordance with an embodiment of the present disclosure.
FIGS. 6A through 6K depict graphical representations illustrating waveform of reconstructed 65-lead BSP data from a 12-lead ECG in a test set as generated by time-series conditional generative adversarial network architecture (TSGAN) along with original waveform according to some embodiment of the present disclosure.

DETAILED DESCRIPTION OF EMBODIMENTS
Exemplary embodiments are described with reference to the accompanying drawings. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. Wherever convenient, the same reference numbers are used throughout the drawings to refer to the same or like parts. While examples and features of disclosed principles are described herein, modifications, adaptations, and other implementations are possible without departing from the scope of the disclosed embodiments. It is intended that the following detailed description be considered as exemplary only, with the true scope being indicated by the following embodiments described herein.
Cardiovascular Diseases (CVD) have grown more common over the years and are one of the a leading causes of death globally. Among the vast plethora of CVD diagnosis tools, Electrocardiogram (ECG) remains the most commonly conducted procedure for detection of CVDs. BSP is an augmented representation of standard ECG measurement with a large number of leads (65 to 512 leads) which are particularly useful in source localization in detection of arrhythmias. However, they are difficult to physically implement.
Embodiments of the present disclosure provide systems and methods for reconstruction of body surface potential from a 12-lead electrocardiogram using a conditional generative adversarial network architecture. In the present disclosure, a conditional generative adversarial network (GAN) architecture is used to generate BSP data such as 65-lead BSP data from the standard 12-lead ECG. More Specifically, the present disclosure describes the following:
Implementation of a time-Series conditional generative adversarial network (TSGAN) architecture, a modified pix2pix GAN architecture for reconstruction of BSP time-series from reduced set of ECG leads.
Morphology-preserving regularization terms in the loss function for an accurate reconstruction.
Referring now to the drawings, and more particularly to FIGS. 1 through 6K, where similar reference characters denote corresponding features consistently throughout the figures, there are shown preferred embodiments and these embodiments are described in the context of the following exemplary system and/or method.
FIG. 1 illustrates an exemplary system 100 for reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network architecture according to some embodiments of the present disclosure. In an embodiment, the system 100 includes or is otherwise in communication with one or more hardware processors 104, communication interface device(s) or input/output (I/O) interface(s) 106, and one or more data storage devices or memory 102 operatively coupled to the one or more hardware processors 104. The one or more hardware processors 104, the memory 102, and the I/O interface(s) 106 may be coupled to a system bus 108 or a similar mechanism.
The I/O interface(s) 106 may include a variety of software and hardware interfaces, for example, a web interface, a graphical user interface, and the like. The I/O interface(s) 106 may include a variety of software and hardware interfaces, for example, interfaces for peripheral device(s), such as a keyboard, a mouse, an external memory, a plurality of sensor devices, a printer and the like. Further, the I/O interface(s) 106 may enable the system 100 to communicate with other devices, such as web servers and external databases.
The I/O interface(s) 106 can facilitate multiple communications within a wide variety of networks and protocol types, including wired networks, for example, local area network (LAN), cable, etc., and wireless networks, such as Wireless LAN (WLAN), cellular, or satellite. For the purpose, the I/O interface(s) 106 may include one or more ports for connecting a number of computing systems with one another or to another server computer. Further, the I/O interface(s) 106 may include one or more ports for connecting a number of devices to one another or to another server.
The one or more hardware processors 104 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuitries, and/or any devices that manipulate signals based on operational instructions. Among other capabilities, the one or more hardware processors 104 are configured to fetch and execute computer-readable instructions stored in the memory 102. In the context of the present disclosure, the expressions ‘processors’ and ‘hardware processors’ may be used interchangeably. In an embodiment, the system 100 can be implemented in a variety of computing systems, such as laptop computers, portable computer, notebooks, hand-held devices, workstations, mainframe computers, servers, a network cloud and the like.
The memory 102 may include any computer-readable medium known in the art including, for example, volatile memory, such as static random access memory (SRAM) and dynamic random access memory (DRAM), and/or non-volatile memory, such as read only memory (ROM), erasable programmable ROM, flash memories, hard disks, optical disks, and magnetic tapes. In an embodiment, the memory 102 includes a plurality of modules 102a and a repository 102b for storing data processed, received, and generated by one or more of the plurality of modules 102a. The plurality of modules 102a may include routines, programs, objects, components, data structures, and so on, which perform particular tasks or implement particular abstract data types.
The plurality of modules 102a may include programs or computer-readable instructions or coded instructions that supplement applications or functions performed by the system 100. The plurality of modules 102a may also be used as, signal processor(s), state machine(s), logic circuitries, and/or any other device or component that manipulates signals based on operational instructions. Further, the plurality of modules 102a can be used by hardware, by computer-readable instructions executed by the one or more hardware processors 104, or by a combination thereof. Further, the memory 102 may include information pertaining to input(s)/output(s) of each step performed by the processor(s) 104 of the system 100 and methods of the present disclosure.
The repository 102b may include a database or a data engine. Further, the repository 102b amongst other things, may serve as a database or includes a plurality of databases for storing the data that is processed, received, or generated as a result of the execution of the plurality of modules 102a. Although the repository 102b is shown internal to the system 100, it will be noted that, in alternate embodiments, the repository 102b can also be implemented external to the system 100, where the repository 102b may be stored within an external database (not shown in FIG. 1) communicatively coupled to the system 100. The data contained within such external database may be periodically updated. For example, new data may be added into the external database and/or existing data may be modified and/or non-useful data may be deleted from the external database. In one example, the data may be stored in an external system, such as a Lightweight Directory Access Protocol (LDAP) directory and a Relational Database Management System (RDBMS). In another embodiment, the data stored in the repository 102b may be distributed between the system 100 and the external database.
FIG. 2, with reference to FIG. 1, depicts an exemplary flow chart illustrating a method 200 for reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network architecture, using the system 100 of FIG. 1, in accordance with an embodiment of the present disclosure.
Referring to FIG. 2, in an embodiment, the system(s) 100 comprises one or more data storage devices or the memory 102 operatively coupled to the one or more hardware processors 104 and is configured to store instructions for execution of steps of the method by the one or more processors 104. The steps of the method 200 of the present disclosure will now be explained with reference to components of the system 100 of FIG. 1, the flow diagram as depicted in FIG. 2, and one or more examples. Although steps of the method 200 including process steps, method steps, techniques or the like may be described in a sequential order, such processes, methods and techniques may be configured to work in alternate orders. In other words, any sequence or order of steps that may be described does not necessarily indicate a requirement that the steps be performed in that order. The steps of processes described herein may be performed in any practical order. Further, some steps may be performed simultaneously, or some steps may be performed alone or independently.
In an embodiment, at step 202 of the present disclosure, the one or more hardware processors 104 are configured to receive a plurality of electrocardiogram (ECG) time series data from a 12-lead system. In the present disclosure, the plurality of electrocardiogram (ECG) time series data and a boy surface potential time series data is used from a known in the art data which is EDGAR database. The EDGAR dataset is also referred as ‘Nijmegen’ data an contains 65-lead BSP signals acquired using 53 electrodes on front and 12 on back of the torso. FIGS. 3A and 3B show electrode distribution of standard 12-lead ECG configuration and the Nijmegen BSP configuration according to some embodiments of the present disclosure. As shown in FIG. 3A, in 12-lead ECG configuration, there are three limb leads, namely Right Arm (RA), Left Arm (LA), Left Leg (LL) and 6 precordial leads (V1 to V6) placed at specific anatomically standardized locations. Remaining 3 augmented leads (aVR, aVL and aVF) are calculated from the limb leads. In the Nijmegen BSP dataset, along with the potential profiles of 65 leads, co-ordinate information to extract the 9 leads (limb and precordial) of 12-lead ECG are also provided. The 12-lead ECG information is extracted and used as input.
In an embodiment, at step 204 of the present disclosure, the one or more hardware processors 104 are configured to input the plurality of electro-cardiogram (ECG) time series data to a trained time-series conditional generative adversarial network (TSGAN) architecture. The TSGAN architecture comprises a generator model and a discriminator model. FIG. 4, with reference to FIGS. 1-3, depicts a functional block diagram of the generator model of the trained TSGAN architecture as implemented by the system of 100 of FIG. 1, in accordance with an embodiment of the present disclosure. The generator model of the trained TSGAN architecture comprises a one-dimensional (1D) convolutional encoder-decoder structure, wherein an encoder in the 1D convolutional encoder-decoder structure comprises three one-dimensional (1D) convolutional layers and a decoder in the 1D convolutional encoder-decoder structure comprises three transposed one-dimensional (1D) convolutional layers with associated batch normalization and leaky Rectified linear unit (ReLU) activation layer. The discriminator model of the trained TSGAN architecture comprises eight 1D convolution layers with one or more associated batch normalization and Leaky ReLU activation layers, and wherein the Leaky ReLU activation layers are used in first seven layers of the eight 1D convolution layers and an eighth layer of the eight 1D convolution layers uses a sigmoid function for classification.
In an embodiment, at step 206 of the present disclosure, the one or more hardware processors 104 are configured to generate a body surface potential (BSP) time-series signal using the generator model of the trained TSGAN architecture. The body surface potential (BSP) time-series signal may include but not limited to a 65-lead BSP, 90-lead BSP, 120-lead BSP, and/or the like. In the context of the present disclosure, the body surface potential (BSP) time-series signal is a 65-lead MSP. The generator model takes x as input and generates G(x,z). Generally, the generator model takes random noise as input and creates new plausible synthetic examples. Traditional GANs have no control over the types of generated examples which can be improved by conditional GANs. The generator model takes 12-lead ECG time-series data having 3000 samples in each lead (channel) as input (dimension = (3000, 1, 12)) and maps it into 65-lead BSP (dimension = (3000, 1, 65)) using a 1D convolutional encoder-decoder structure shown in FIG. 4. The output tensor dimension of each block is provided at the bottom of the block. The encoder has three 1D convolutional layers, with associated batch normalization and leaky ReLU activation layer. In each layer, convolution is performed with a kernel dimension of 5 and stride length of 2 by applying zero-padding to the inputs. 65 filters to each convolutional layer are applied in the generator model. Output is flattened after the third layer and is applied to a dense layer with leaky ReLU activation to get an encoded vector. The decoder comprises a series of transposed-convolution layers having similar kernel dimension with stride length of 2 to increase the output dimension to eventually map into desired output. In order to maintain stochasticity in the output, 40% dropout to different layers of the encoder are added. Additionally, the input data (x) is added with low amplitude random Gaussian noise (z).
In an embodiment, at step 208 of the present disclosure, the one or more hardware processors 104 are configured to determine a difference between a target body surface potential (BSP) time-series signal and the generated body surface potential (BSP) time-series signal using the discriminator model of the trained TSGAN architecture. The discriminator model takes x and G(x,z) (fake) as well as x and target body surface potential (BSP) time-series signal (i.e., actual target data y (real)) as inputs. The discriminator model classifies it’s input as one of fake (generated) and real. In an embodiment, the discriminator model is a one-dimensional (1D) convolutional patchGAN discriminator model which takes two sets of paired inputs. The input 12-lead ECG (x) paired with the real (y -the actual) and fake (G(x,z) which is the generated body surface potential (BSP) time-series signal conditioned on the input are applied to the discriminator model in separate batches. The discriminator model aims to classify whether the pair of data is real or fake by minimizing the likelihood of a negative log identifying real and fake data. In patchGAN discriminator model, the model outputs a tensor where each element is corresponding to a patch of the input and the value indicates whether the patch is real or fake. The discriminator model is applied convolutionally across the input data, averaging all responses to provide the final output prediction. The patch size is selected as 70x1. The discriminator model comprises eight one-dimensional (1D) convolution layers with associated batch normalization and activation layers. Leaky ReLU is used in the first seven layers. The final layer uses sigmoid function for classification. The output at the end of the final block is a (30,1) tensor, where each point represents a 70x1 patch in the input.
Referring to steps of FIG. 2 of the present disclosure, at step 210 of the present disclosure, the one or more hardware processors 104 are configured to reconstruct the target body surface potential (BSP) time-series signal for a plurality of incoming electro-cardiogram (ECG) time series data by optimizing an objective function of the TSGAN architecture. The difference between the target body surface potential (BSP) time-series signal and the generated body surface potential (BSP) time-series signal is minimized and one or more morphological properties of the reconstructed target body surface potential (BSP) time-series signal is preserved while restructuring the BSP time-series signal. The objective function is optimized by minimizing one or more loss functions using one or more morphology-preserving regularization terms. In an embodiment, plurality of incoming electro-cardiogram (ECG) time series could be interchangeably referred as test data or test set throughout the description.
The step 210 is further better understood by way of the following exemplary explanation.
A pix2pix model is a conditional GAN popularly used in image to image translation. Unlike traditional GANs, the generator model of pix2pix GAN takes a source image as input and transforms into a translated image. The discriminator model determines whether the translated image is a plausible transformation of the source image. In an embodiment, TSGAN is a modified pix2pix model to generate 65-lead BSP time-series from 12-lead ECG. If x is the input (i.e., 12-lead ECG time series data), z is the random noise, y is a target output (original 65-lead BSP), G(x,z) is the generated 65-lead BSP, then the objective function of a conditional GAN is represented by:
L_cGAN (G,D)=E_(x,y) [log?(1-D(x,G(x,z))) ] ---- (1)
FIGS. 5A an 5B depict graphical illustrations of cardiac electrical activity recorded at torso from lead II and lead V3 for reconstruction of body surface potential from 12-lead electrocardiogram using a conditional generative adversarial network, in accordance with an embodiment of the present disclosure. In other words, the normal ECG template of lead II and lead V3 along with major morphological points of interest like P,Q,R,S,T points and their intervals are shown in FIGS. 5A and 5B. These specific points in ECG correspond to major electrophysiological events leading into generation of the ECG patterns, like P wave signifies atrial depolarization, QRS complex represents ventricular depolarization and T wave corresponds to ventricular repolarization. The morphological patterns of these specific regions along with their intervals vary for normal and pathological conditions. A L2 loss of the position differences of P,Q,R,S,T peaks in the target and generated waveform on time axis is also minimized along with a generator loss function. The final objective of the proposed GAN is:
G* =arg min-G?max-D??L_cGAN (G,D)+?_1.L_(?L2?_sig ) ? (G)+?_2.L_(?L2?_P ) (G)+?_3.L_(?L2?_Q ) (G)+?_4.L_(?L2?_R ) (G)+?_5.L_(?L2?_S ) (G)+?_6.L_(?L2?_T ) (G) (2)
Here,
L_(?L2?_sig ) (G)=E_(x,y,z) [?y-G(x,z)?_2 ]
The other terms, L_(?L2?_P ), L_(?L2?_Q ), L_(?L2?_R ), L_(?L2?_S ), L_(?L2?_T ) denote the L2 loss of time difference between the P,Q,R,S,T locations in the target and the generated waveform. The constants, denoted by lambdas in equation (2) are regularization constants and adjusted during training to assign higher weightage to selected portions of the BSP waveform requiring more accurate reconstruction. In a traditional pix2pix model, a L1 loss of pixel difference between the target and the generated image is added with the generator loss function for minimizing during training. However, in TSGAN, a L2 loss of the amplitude difference between the target and generated 65-lead BSP is minimized which causes a more accurate and noise-free reconstruction of time-series. Few more regularization terms are also added to preserve the one or more morphological properties in the generated waveform.
The TSGAN is implemented in Python 3.8.10 using TensorFlow 2.6.0 library. The model is trained on a computer system having Intel® Xeon(R) 16-core processor, 64 GB of RAM and and NVidia GeForce® GTX 1080 Ti graphics processing unit. NeuroKit2®, a Python package for neurophysiological data analysis is used to extract the P,Q,R,S,T regions. The generator model takes x as input and generates G(x,z). The discriminator model takes x and G(x,z) (fake) as well as x and the actual target data y (real) as inputs. The real and fake data are labeled by arrays of ones and zeros. Binary cross entropy loss is defined to model the objectives of the generator model and the discriminator model. For the generator model, the cross entropy loss of the generated images and an array of ones is measured. A discriminator loss is calculated by averaging the sum of the cross entropy loss for the real and the fake data. In each iteration, first the discriminator loss is measured followed by the generator loss. Next, the gradient of the loss is measured with respect to model weights using backpropagation to update the weighs. The regularization constants are selected by trial and error. The value of ?_1 is selected as 50. The other constants are set as ?_2 = 1, ?_3 = 10, ?_4 = 20, ?_5= 15, ?_6 = 20 for an optimum reconstruction. An Adam optimizer with a learning rate of 0.0002, and momentum parameters ß_1= 0.5, ß_2= 0.999 is used for training. The batch size is taken as 4. The model weights are initialized from a normal distribution with zero mean and standard deviation of 0.02. The end-to-end GAN is trained for 200 epochs. Once the training is done, the discriminator model is discarded and the generator model is used for BSP reconstruction on the test data.

EXPERIMENTAL RESULTS:
FIGS. 6A through 6K depict graphical representations illustrating waveform of reconstructed 65-lead BSP data from a 12-lead ECG in a test set as generated by the time-series conditional generative adversarial network architecture (TSGAN) along with original waveform according to some embodiment of the present disclosure. In FIGS. 6A through 6K, original waveform is shown by a solid line and the reconstructed waveform is shown by the dotted line. As observed from FIGS. 6A through 6K, the reconstructed waveform closely matches the morphology of the original waveform in almost all leads except lead: 10, 21, 24, 34, 48, 61, 62, where the reconstructed data is noisy. In general, the output leads, located far from any of the base 9 leads of the input 12-lead ECG time series data yield higher reconstruction loss. A comparison of the performance of the system of present disclosure with variational autoencoder (VAE) and baseline conditional generative adversarial network (cGAN) in terms of reconstruction loss is provided in Table 1.
Model architecture Minimum reconstruction loss Maximum reconstruction loss Average reconstruction loss
VAE 0.08 (lead 13) 0.61 (lead 48) 0.31
Baseline cGAN 0.06 (lead 24) 0.32 (lead 62) 0.18
TSGAN (system of present disclosure) 0.02 (lead 30) 0.12 (lead 62) 0.07
Table 1
It is observed from Table 1 that overall reconstruction loss on the test set is measure in terms of Normalized Root Mean Square Error (NRMSE) between the original and the reconstructed BSP data across various leads. Also, it is observed form Table 1 that the baseline cGAN has an identical structure to the TSGAN, but it does not consider morphology-specific regularization terms in equation (2). In Table 1, lead position and NRMSE value corresponding to the minimum and the maximum reconstruction loss are reported along with the average reconstruction loss for all the leads. It can be clearly observed from Table 1 that the TSGAN yields the least reconstruction loss. The GAN-based models inevitably outperform the VAE model. However, utilizing the morphology-preserving regularization, a much improved reconstruction can be achieved compared to the baseline cGAN.
Table 2 provides morphological analysis of extracted features from the reconstructed data with respect to the original data in terms of NRMSE.
Model RR QRS QT PR
ME AE ME AE ME AE ME AE
VAE 0.26
(L24) 0.08 0.55
(L10) 0.36 0.25
(L62) 0.19 0.91
(L62) 0.77
Baseline cGAN 0.08
(L24) 0.03 0.82
(L62) 0.35 0.26
(L24) 0.12 0.68
(L61) 0.47
TSGAN (system of present disclosure) 0.08
(L61) 0.02 0.38
(L62) 0.13 0.12
(L30) 0.06 0.42
(L62) 0.29
Table 2
In Table 2, maximum and average NRMSE for RR, QRS, QT and PR interval time distances between the actual and target waveform in the test set are reported. As observed from Table 2, errors are higher for PR and QRS segments as an accurate detection of P and S points are often unreliable by a known in the rat tool box (NeuroKit2) in certain leads. The TSGAN still reports the least error for all the features, justifying its usability in morphology-preserving reconstruction.
The written description describes the subject matter herein to enable any person skilled in the art to make and use the embodiments. The scope of the subject matter embodiments is defined herein and may include other modifications that occur to those skilled in the art. Such other modifications are intended to be within the scope of the present disclosure if they have similar elements that do not differ from the literal language of the embodiments or if they include equivalent elements with insubstantial differences from the literal language of the embodiments described herein.
The embodiments of present disclosure provide a Generative Adversarial Network (GAN) architecture to reconstruct 65-lead BSP from standard 12-lead ECG. The Time-Series GAN (TSGAN), a specially designed modified pix2pix GAN is implemented for an accurate reconstruction of time-series BSP data. Further, some regularization terms are used in the generator loss function to preserve morphological properties of the generated waveform. The present disclosure outperforms a Variational Autoencoder (VAE) and a baseline GAN on publicly available dataset in reconstructing 65-lead BSP with morphological preservation.
It is to be understood that the scope of the protection is extended to such a program and in addition to a computer-readable means having a message therein; such computer-readable storage means contain program-code means for implementation of one or more steps of the method, when the program runs on a server or mobile device or any suitable programmable device. The hardware device can be any kind of device which can be programmed including e.g. any kind of computer like a server or a personal computer, or the like, or any combination thereof. The device may also include means which could be e.g. hardware means like e.g. an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or a combination of hardware and software means, e.g. an ASIC and an FPGA, or at least one microprocessor and at least one memory with software processing components located therein. Thus, the means can include both hardware means and software means. The method embodiments described herein could be implemented in hardware and software. The device may also include software means. Alternatively, the embodiments may be implemented on different hardware devices, e.g. using a plurality of CPUs.
The embodiments herein can comprise hardware and software elements. The embodiments that are implemented in software include but are not limited to, firmware, resident software, microcode, etc. The functions performed by various components described herein may be implemented in other components or combinations of other components. For the purposes of this description, a computer-usable or computer readable medium can be any apparatus that can comprise, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.
The illustrated steps are set out to explain the exemplary embodiments shown, and it should be anticipated that ongoing technological development will change the manner in which particular functions are performed. These examples are presented herein for purposes of illustration, and not limitation. Further, the boundaries of the functional building blocks have been arbitrarily defined herein for the convenience of the description. Alternative boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed. Alternatives (including equivalents, extensions, variations, deviations, etc., of those described herein) will be apparent to persons skilled in the relevant art(s) based on the teachings contained herein. Such alternatives fall within the scope of the disclosed embodiments. Also, the words “comprising,” “having,” “containing,” and “including,” and other similar forms are intended to be equivalent in meaning and be open ended in that an item or items following any one of these words is not meant to be an exhaustive listing of such item or items, or meant to be limited to only the listed item or items. It must also be noted that as used herein, the singular forms “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.
Furthermore, one or more computer-readable storage media may be utilized in implementing embodiments consistent with the present disclosure. A computer-readable storage medium refers to any type of physical memory on which information or data readable by a processor may be stored. Thus, a computer-readable storage medium may store instructions for execution by one or more processors, including instructions for causing the processor(s) to perform steps or stages consistent with the embodiments described herein. The term “computer-readable medium” should be understood to include tangible items and exclude carrier waves and transient signals, i.e., be non-transitory. Examples include random access memory (RAM), read-only memory (ROM), volatile memory, nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and any other known physical storage media.
It is intended that the disclosure and examples be considered as exemplary only, with a true scope of disclosed embodiments being indicated herein by the following claims.
,CLAIMS:
1. A processor implemented method, comprising:
receiving (202), via one or more hardware processors, a plurality of electrocardiogram (ECG) time series data from a 12-lead system;
inputting (204), via the one or more hardware processors, the plurality of electro-cardiogram (ECG) time series data to a trained time-series conditional generative adversarial network (TSGAN) architecture, wherein the TSGAN architecture comprises a generator model and a discriminator model;
generating (206), via the one or more hardware processors, a body surface potential (BSP) time-series signal using the generator model of the trained TSGAN architecture;
determining (208), via the one or more hardware processors, a difference between a target body surface potential (BSP) time-series signal and the generated body surface potential (BSP) time-series signal using the discriminator model of the trained TSGAN architecture; and
reconstructing (210), via the one or more hardware processors, the target body surface potential (BSP) time-series signal for a plurality of incoming electro-cardiogram (ECG) time series data by optimizing an objective function of the TSGAN architecture, wherein the difference between the target body surface potential (BSP) time-series signal and the generated body surface potential (BSP) time-series signal is minimized and one or more morphological properties of the reconstructed target body surface potential (BSP) time-series signal is preserved while restructuring the BSP time-series signal.

2. The processor implemented method as claimed in claim 1, wherein the generator model of the trained TSGAN architecture comprises a one-dimensional (1D) convolutional encoder-decoder structure, wherein an encoder in the 1D convolutional encoder-decoder structure comprises three one-dimensional (1D) convolutional layers and a decoder in the 1D convolutional encoder-decoder structure comprises three transposed one-dimensional (1D) convolutional layers with associated batch normalization and leaky Rectified linear unit (ReLU) activation layer.

3. The processor implemented method as claimed in claim 1, wherein the discriminator model of the trained TSGAN architecture comprises eight one-dimensional (1D) convolution layers with one or more associated batch normalization and Leaky ReLU activation layers, and wherein the Leaky ReLU activation layers are used in first seven layers of the eight 1D convolution layers and an eighth layer of the eight 1D convolution layers uses a sigmoid function for classification.

4. The processor implemented method as claimed in claim 1, wherein the objective function is optimized by minimizing one or more loss functions using one or more morphology-preserving regularization terms.

5. A system (100), comprising:
a memory (102) storing instructions;
one or more communication interfaces (106); and
one or more hardware processors (104) coupled to the memory (102) via the one or more communication interfaces (106), wherein the one or more hardware processors (104) are configured by the instructions to:
receive a plurality of electrocardiogram (ECG) time series data from a 12-lead system;
input the plurality of electro-cardiogram (ECG) time series data to a trained time-series conditional generative adversarial network (TSGAN) architecture, wherein the TSGAN architecture comprises a generator model and a discriminator model;
generate a body surface potential (BSP) time-series signal using the generator model of the trained TSGAN architecture;
determine a difference between a target body surface potential (BSP) time-series signal and the generated body surface potential (BSP) time-series signal using the discriminator model of the trained TSGAN architecture; and
reconstruct the target body surface potential (BSP) time-series signal for a plurality of incoming electro-cardiogram (ECG) time series data by optimizing an objective function of the TSGAN architecture, wherein the difference between the target body surface potential (BSP) time-series signal and the generated body surface potential (BSP) time-series signal is minimized and one or more morphological properties of the reconstructed target body surface potential (BSP) time-series signal is preserved while restructuring the BSP time-series signal.

6. The system as claimed in claim 5, wherein the generator model of the trained TSGAN architecture comprises a one-dimensional (1D) convolutional encoder-decoder structure, wherein an encoder in the 1D convolutional encoder-decoder structure comprises three one-dimensional (1D) convolutional layers and a decoder in the 1D convolutional encoder-decoder structure comprises three transposed one-dimensional (1D) convolutional layers with associated batch normalization and leaky Rectified linear unit (ReLU) activation layer.

7. The system as claimed in claim 5, wherein the discriminator model of the trained TSGAN architecture comprises eight one-dimensional (1D) convolution layers with one or more associated batch normalization and Leaky ReLU activation layers, and wherein the Leaky ReLU activation layers are used in first seven layers of the eight 1D convolution layers and an eighth layer of the eight 1D convolution layers uses a sigmoid function for classification.

8. The system as claimed in claim 5, wherein the objective function is optimized by minimizing one or more loss functions using one or more morphology-preserving regularization terms.