US20210321960A1

US20210321960A1 - Machine learning using simulated cardiograms

Info

Publication number: US20210321960A1
Application number: US17/110,101
Authority: US
Inventors: Christopher Villongco
Original assignee: Vektor Medical Inc
Current assignee: Vektor Medical Inc
Priority date: 2018-04-26
Filing date: 2020-12-02
Publication date: 2021-10-21
Also published as: US20190333639A1; US11344263B2; US11504073B2; US10959680B2; US20190328456A1; US20190328336A1; US20190328255A1; US20230190114A1; US20190328254A1; US20190328335A1; US11253206B2; US11547369B2; US20210068763A1; US11259756B2; US10856816B2; US11576624B2; US11622732B2; US20190328257A1; US11564641B2; US20190333643A1

Abstract

A system is provided for generating a classifier for classifying electromagnetic data (e.g., ECG) derived from an electromagnetic source (e.g., heart). The system accesses a computational model of the electromagnetic source. The computational model models the electromagnetic output of the electromagnetic source over time based on a source configuration (e.g., rotor location) of the electromagnetic source. The system generates, for each different source configuration (e.g., different rotor locations), a modeled electromagnetic output (e.g., ECG) of the electromagnetic source for that source configuration. For each modeled electromagnetic output, the system derives the electromagnetic data for the modeled electromagnetic output and generates a label (e.g., rotor location) for the derived electromagnetic data from the source configuration for the modeled electromagnetic data. The system trains a classifier with the derived electromagnetic data and the labels as training data. The classifier can then be used to classify the electromagnetic output collected from patients.

Description

CROSS-REFERENCE TO RELATED APPLICATION

This application is a continuation of U.S. patent application Ser. No. 16/162,695 filed Oct. 17, 2018, which claims the benefit of U.S. Provisional Application No. 62/663,049, filed on Apr. 26, 2018, entitled “MACHINE LEARNING USING SIMULATED CARDIOGRAMS,” which is hereby incorporated by reference it its entirety.

BACKGROUND

Many heart disorders can cause symptoms, morbidity (e.g., syncope or stroke), and mortality. Common heart disorders caused by arrhythmias include atrial fibrillation (“AF”), ventricular fibrillation (“VF”), atrial tachycardia (“AT”), ventricular tachycardia (“VT”), atrial flutter (“AFL”), premature ventricular complexes (“PVCs”), atrioventricular nodal reentrant tachycardia (“AVNRT”), atrioventricular reentrant tachycardia (“AVRT”), and junctional tachycardia (“JT”). The sources of arrhythmias may include stable electrical rotors, recurring electrical focal sources, anatomically-based reentry, and so on. These sources are important drivers of sustained or clinically significant episodes. Arrhythmias can be treated with ablation using different technologies including radiofrequency energy ablation, cryoablation, ultrasound ablation, laser ablation, external radiation sources, and so on by targeting the source of the heart disorder. Since the sources of heart disorders and the locations of the source vary from patient to patient, even for common heart disorders, targeted therapies require the source of the arrhythmia to be identified.
Unfortunately, current methods for reliably identifying the source locations of the source of a heart disorder can be complex, cumbersome, and expensive. For example, one method uses an electrophysiology catheter having a multi-electrode basket catheter that is inserted into the heart (e.g., left ventricle) intravascularly to collect from within the heart measurements of the electrical activity of the heart, such as during an induced episode of VF. The measurements can then be analyzed to help identify a possible source location. Presently, electrophysiology catheters are expensive (and generally limited to a single use) and may lead to serious complications, including cardiac perforation and tamponade. Another method uses an exterior body surface vest with electrodes to collect measurements from the patient's body surface, which can be analyzed to help identify an arrhythmia source location. Such body surface vests are expensive, complex and difficult to manufacture, and may interfere with the placement of defibrillator pads needed after inducing ventricular fibrillation to collect measurements during the arrhythmia. In addition, the vest analysis requires a computed tomography (“CT”) scan and is unable to sense the interventricular and interatrial septa where approximately 20% of arrhythmia sources may occur.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram that illustrates the overall processing of the MLMO system in some embodiments.

FIG. 2 is a flow diagram that illustrates the overall processing of generating a classifier by the MLMO system in some embodiments.

FIG. 3 is a block diagram that illustrates training and classifying using a convolutional neural network in some embodiments.

FIG. 4 is a flow diagram that illustrates processing of a generate classifier component of the MLMO system in some embodiments.

FIG. 5 is a flow diagram that illustrates the processing of a generate simulated VCGs component of the MLMO system in some embodiments.

FIG. 6 is a flow diagram that illustrates the processing of a generate training data component for cycles in some embodiments.

FIG. 7 is a flow diagram that illustrates the processing of an identify cycles component of the MLMO system in some embodiments.

FIG. 8 is a block diagram that illustrates the processing of a normalize cycle component of the MLMO system in some embodiments.

FIG. 9 is a flow diagram that illustrates processing of a generate training data for a sequence of similar cycles component of the MLMO system in some embodiments.

FIG. 10 is a flow diagram that illustrates the processing of a classify component of the MLMO system in some embodiments.

DETAILED DESCRIPTION

A method and a system are provided for generating a classifier for classifying electromagnetic data derived from an electromagnetic source within a body. A body may be, for example, a human body, and the electromagnetic source may be a heart, a brain, a liver, a lung, a kidney, or another part of the body that generates an electromagnetic field that can be measured, preferably, from outside the body and represented via a cardiogram such as an electrocardiogram (“ECG”), a vectorcardiogram (“VCG”), and an electroencephalogram (“EEG”). In some embodiments, a machine learning based on modeled output (“MLMO”) system is provided to generate a classifier by modeling electromagnetic output of the electromagnetic source for a variety of source configurations and using machine learning to train a classifier using derived electromagnetic data that is derived from the modeled electromagnetic output as training data. The MLMO system is described below primarily to generate a classifier for electromagnetic data of the heart.
In some embodiments, the MLMO system employs a computational model of the electromagnetic source to generate training data for training the classifier. A computational model models electromagnetic output of the electromagnetic source over time based on a source configuration of the electromagnetic source. The electromagnetic output may represent electrical potential, a current, a magnetic field, and so on. When the electromagnetic (“EM”) source is a heart, the source configuration may include information on geometry and muscle fibers of the heart, torso anatomy, scar locations, rotor locations, electrical properties, and so on, and the EM output is a collection of the electric potentials at various heart locations over time. To generate the EM output, a simulation may be performed for simulation steps of a step size (e.g., 1 ms) to generate an EM mesh for that step. The EM mesh may be a finite-element mesh that stores the value of the electric potential at each heart location for that step. For example, the left ventricle may be defined as having approximately 70,000 heart locations with the EM mesh storing an electromagnetic value for each heart location. If so, a three-second simulation with a step size of 1 ms would generate 3,000 EM meshes that each include 70,000 values. The collection of the EM meshes is the EM output for the simulation. A computational model is described in C. T. Villongco, D. E. Krummen, P. Stark, J. H. Omens, & A. D. McCulloch, “Patient-specific modeling of ventricular activation pattern using surface ECG-derived vectorcardiogram in bundle branch block,” Progress in Biophysics and Molecular Biology, Volume 115, Issues 2-3, August 2014, Pages 305-313, which is hereby incorporated by reference.
In some embodiments, the MLMO system generates the training data by running many simulations, each based on a different source configuration, which is a set of different values for the configuration parameters of the computational model. For example, the configuration parameters for the heart may be cardiac geometry, rotor location, focal source location, ventricular orientation in the chest, ventricular myofiber orientation, cardiomyocyte intracellular potential electrogenesis and propagation, and so on. Each configuration parameter may have a set or range of possible values. For example, the rotor location may be 78 possible parameter sets corresponding to different locations within a ventricle. Since the MLMO system may run a simulation for each combination of possible values, the number of simulations may be in the millions.
In some embodiments, the MLMO system uses EM outputs of the simulations to train the classifier for the generation of a classification based on EM data collected from a patient. The MLMO system may generate derived EM data, such as an ECG or VCG, for each EM output of a simulation. The ECG and VCG are equivalent source representations of the EM output. The MLMO then generates a label (or labels) for each derived EM data to specify its corresponding classification. For example, the MLMO system may generate a label that is the value of a configuration parameter (e.g., rotor location) used when generating the EM output from which the EM data was derived. The collection of the derived EM data, which correspond to feature vectors, and their labels compose the training data for training the classifier. The MLMO system then trains the classifier. The classifier may be any of a variety or combination of classifiers including neural networks such as fully-connected, convolutional, recurrent, autoencoder, or restricted Boltzmann machine, a support vector machine, a Bayesian classifier, and so on. When the classifier is a deep neural network, the training results in a set of weights for the activation functions of the deep neural network.
In some embodiments, the MLMO system may augment the training data with additional features from the configuration parameters of the source configuration used to generate the training data. For example, the MLMO system may generate additional features to represent the geometry of the heart, the orientation of the heart, scar location, ablation location, ablation shape, and so on. The MLMO system may input these additional features into the fully connected layer along with the output generated by the layer before the fully connected layer of a convolutional neural network (“CNN”), which is described below. The output of the layer before the fully connected layer (e.g., pooling layer) may be “flattened” into a one-dimensional array, and the MLMO system may add the additional features as further elements of the one-dimensional array. The output of the fully connected layer may provide a probability for each label used in the training data. The probabilities will thus be based on the combination of the derived EM data and the additional features. The classifier will be able to output different probabilities even when the derived EM data is the same or similar to reflect, for example, that the same or similar EM data may be generated for patients with different heart geometries and different scar locations. The MLMO system may alternatively employ an additional classifier that (1) inputs the probabilities generated by the CNN based only on the derived EM data and (2) inputs the additional features and then outputs a final probability for each classification that factors in the additional features. The additional classifier may be, for example, a support vector machine. The CNN and the additional classifier may be trained in parallel.
In some embodiments, the MLMO system normalizes the VCGs of each cycle of the training data in both the voltage and time axes. A cycle may be defined as a time interval (e.g., start time to end time) defining a single unit or beat of periodic electrical activity during normal or abnormal rhythms. Cycles facilitate beat-by-beat analysis of source configuration evolution over time and enable subsequent voltage and time normalization over each cycle. Normalization preserves salient features of voltage-time dynamics and improves generalizability of the training data to variations in source configuration parameters (e.g. torso conductivities, lead placement and resistance, myocardial conduction velocity, action potential dynamics, overall heart size, etc.) anticipated in real patients. The MLMO system may normalize the voltages to a range between −1 and 1 and the time to a fixed range of 0 to 1 in increments of milliseconds or percentages. To normalize the voltages for a cycle, the MLMO system may identify the maximum magnitude of the vectors across the axes. The MLMO system divides each voltage by the maximum magnitude. To normalize the time axis, the MLMO system performs an interpolation from the number of points in the VCG, which may be more or less than 1000, to the 1000 points of the normalized cycle.
In some embodiments, after the classifier is trained, the MLMO system is ready to generate classifications based on EM and other routinely available clinical data collected from patients. For example, an ECG may be collected from a patient, and a VCG may be generated from the ECG. The VCG is input to the classifier to generate a classification indicating, for example, a rotor location for the patient. As a result, even though the geometry of the patient's heart is not known or no simulation was based on the same geometry as the patient's heart, the MLMO system can be used to generate a classification. If other patient measurements such as cardiac dimensions and orientation, scar configuration, etc. are available, they may be included as input with the EM data to improve accuracy. This allows the classifier to effectively learn complex hidden features in various clinical data that are not directly represented by the training data.
In some embodiments, the MLMO system may classify source stability (i.e. the beat-to-beat consistency of a dominant arrhythmia source localized to a particular region in the heart) by generating training data that is based on sequences of consecutive cycles that have similar EM features. A technique for determining the stability of arrhythmia sources is described in Krummen, D., et. al., Rotor Stability Separates Sustained Ventricular Fibrillation From Self-Terminating Episodes in Humans, Journal of American College of Cardiology, Vol. 63, No. 23, 2014, which is hereby incorporated by reference. This reference demonstrates the efficacy of targeted ablation at stable source sites for preventing recurring arrhythmic episodes. For example, given a VCG, the MLMO system may identify the cycles and then identify sequences of two consecutive cycles, three consecutive cycles, four consecutive cycles, and so on in which all the VCG cycles in the sequence are of similar morphology to each other. Each identified sequence may be labeled based on the value of a parameter of the source configuration used to generate the VCG. The MLMO system may then train a separate classifier for each sequence length (e.g., 2, 3, 4, and so on) using the training data for the sequences of that sequence length. For example, the MLMO system may train a classifier for sequences of two cycles and a separate classifier for sequences of three cycles. To generate a classification for a patient, the MLMO system may identify sequences of similar cycles of varying sequence lengths in the VCG of the patient and input those sequences into the classifier for the appropriate sequence length. The MLMO system may then combine the classifications from all the classifiers to arrive at a final classification or may simply output all the classifications.
Although a classifier could be trained using actual patient ECGs or VCGs and corresponding intracardiac basket catheter measurements of source location, the cost of collecting, preparing, and labeling a sufficient number of data would be prohibitive. Moreover, training data based on actual patients would likely be too sparse and noisy to be effective at training a classifier for a large population. In some embodiments, the MLMO system could be trained using a combination of actual patient VCGs and VCGs derived from simulations.
FIG. 1 is a block diagram that illustrates the overall processing of the MLMO system in some embodiments. The MLMO system includes classifier generation components 110 and classification components 120. The computational model for a heart is a heart model that may include data and code stored in a heart model data store 111. A generate simulations component 112 inputs the heart model and the parameter sets for the simulations. The parameter sets, also referred to as source configurations, may include a parameter set for each combination of possible values of the parameters or may specify how to generate (e.g., via a computer code) the parameter sets. For example, the computer code for the rotor location parameter may include a list of possible rotor locations and for the ventricle orientation parameter may dynamically generate the values from a base orientation axis along with code for generating possible tilt angles from that base orientation such as an x-axis and a y-axis increment. The output of the generate simulations component is stored in a voltage solutions data store 113 where a voltage solution is an EM output. A voltage solution is an example of an EM mesh. A generate VCGs component 114 generates a VCG from the voltage solutions and stores the VCG in a VCG data store 115. The generate VCGs component may generate an ECG from the voltage solutions and then generate a VCG from the ECG. The generation of a VCG from an ECG is described in J. A. Kors, G. Van Herpen, A. C. Sittig, & J. H. Van Bemmel, “Reconstruction of the Frank vectorcardiogram from standard electrocardiographic leads: diagnostic comparison of different methods,” European Heart Journal, Volume 11, Issue 12, 1 Dec. 1990, Pages 1083-1092, which is hereby incorporated by reference. A generate training data component 116 inputs the VCGs and labels each VCG with a label or labels that may be derived from the parameter sets and stores the training data in a training data store 117. A label may be, for example, the value of a parameter of the parameter set used to generate the EM output from which the VCG is derived. A train classifier component 118 inputs the training data, trains a classifier, and stores the weights (e.g., of activation functions of a convolutional neural network) in a classifier weights data store 119. To generate a classification, a collect ECG component 121 inputs an ECG collected from a patient. A generate VCG component 122 generates a VCG from the ECG. A classify component 123 inputs the VCG and generates a classification using the classifier weights of the trained classifier.
The computing systems (e.g., network nodes or collections of network nodes) on which the MLMO system may be implemented may include a central processing unit, input devices, output devices (e.g., display devices and speakers), storage devices (e.g., memory and disk drives), network interfaces, graphics processing units, cellular radio link interfaces, global positioning system devices, and so on. The input devices may include keyboards, pointing devices, touch screens, gesture recognition devices (e.g., for air gestures), head and eye tracking devices, microphones for voice recognition, and so on. The computing systems may include high-performance computing systems, cloud-based servers, desktop computers, laptops, tablets, e-readers, personal digital assistants, smartphones, gaming devices, servers, and so on. For example, the simulations and training may be performed using a high-performance computing system, and the classifications may be performed by a tablet. The computing systems may access computer-readable media that include computer-readable storage media and data transmission media. The computer-readable storage media are tangible storage means that do not include a transitory, propagating signal. Examples of computer-readable storage media include memory such as primary memory, cache memory, and secondary memory (e.g., DVD) and other storage. The computer-readable storage media may have recorded on them or may be encoded with computer-executable instructions or logic that implements the MLMO system. The data transmission media are used for transmitting data via transitory, propagating signals or carrier waves (e.g., electromagnetism) via a wired or wireless connection. The computing systems may include a secure cryptoprocessor as part of a central processing unit for generating and securely storing keys and for encrypting and decrypting data using the keys.
The MLMO system may be described in the general context of computer-executable instructions, such as program modules and components, executed by one or more computers, processors, or other devices. Generally, program modules or components include routines, programs, objects, data structures, and so on that perform tasks or implement data types of the MLMO system. Typically, the functionality of the program modules may be combined or distributed as desired in various examples. Aspects of the MLMO system may be implemented in hardware using, for example, an application-specific integrated circuit (“ASIC”) or field programmable gate array (“FPGA”).
FIG. 2 is a flow diagram that illustrates the overall processing of generating a classifier by the MLMO system in some embodiments. A generate classifier component 200 is executed to generate a classifier. In block 201, the component accesses the computational model to be used to run the simulations. In block 202, the component selects the next source configuration (i.e., parameter set) to be used in a simulation. In decision block 203, if all the source configurations have already been selected, then the component continues at block 205, else the component continues at block 204. In block 204, the component runs the simulation using the selected source configuration to generate an EM output for the simulation and then loops to block 202 to select the next source configuration. In block 205, the component selects the next EM output that was generated by a simulation. In decision block 206, if all the EM outputs have already been selected, then the component continues at block 210, else the component continues at block 207. In block 207, the component derives the EM data from the EM output. For example, the EM output may be a collection of EM meshes, and the EM data may be an ECG or a VCG derived from the electromagnetic values of the EM mesh. In some embodiments, the component may in addition identify cycles (periodic intervals of arrhythmic activity) within the ECG or VCG. A cycle may be delimited by successive crossings from a negative voltage to a positive voltage (“positive crossings”) or successive crossings from a positive voltage to a negative voltage (“negative crossings”) with respect to a spatial direction or set of directions comprising a reference frame or set of reference frames. A reference frame may coincide with anatomical axes (e.g. left-to-right with x, superior-to-inferior with y, anterior-to-posterior with z), imaging axes (e.g. CT, MR, or x-ray coordinate frames), body-surface lead vectors, principal axes computed by principal component analysis of measured or simulated EM source configurations and outputs, or user-defined directions of interest. For example, a three-second VCG may have three cycles, and each cycle may be delimited by the times of the positive crossings along the x-axis. Alternatively, the cycles may be delimited by crossings along the y-axis or z-axis. In addition, cycles may be defined by negative crossings. Thus, in some embodiments, the component may generate training data from a single VCG based on various cycle definitions that are various combinations of positive crossings and negative crossings with the cycles for all the axes being defined by crossings on one of the x-axis, y-axis, and z-axis or the cycles for each defined by crossings on that axis. Moreover, the training data may include cycles identified based on all possible cycle definitions or a subset of the cycle definition. For example, the training data may include, for each axis, a cycle defined by positive crossings of the x-axis, negative crossings of the y-axis, and positive crossings of that axis itself. Cycles definitions may also be defined by the timing of electrical events derived from the values stored in the EM mesh. For example, a point or set of points in the mesh may periodically cross voltage thresholds signifying electrical activation and deactivation. Thus, a cycle may be defined by activation-deactivation, or successive activation-activation or deactivation-deactivation intervals corresponding to a point or set of points within the mesh. The resulting timings of these intervals can be co-localized to the ECG or VCG for cycles identification. In block 208, the component labels the EM data based on the source configuration (e.g., a source location). When cycles are identified, the component may label each cycle with the same label. For example, the component may label the identified cycles with the same rotor location. In block 209, the component adds the EM data along with the label to the training data and then loops to block 205 to select the next EM output. In block 210, the component trains the classifier using the training data and then completes.
FIG. 3 is a block diagram that illustrates training and classifying using a convolutional neural network in some embodiments. The convolutional neural network may be one-dimensional in the sense that it inputs an image that is a single row of pixels with each pixel having a red, green, and blue (“RGB”) value. The MLMO system sets the values of the pixels based on the voltages of a VCG of the training data. The image has the same number of pixels as vectors of a VCG of the training data. The MLMO system sets the red, green, and blue values of a pixel of the image to the x, y, and z values of the corresponding vector of the VCG. For example, if a cycle of a VCG is 1 second long, and the VCG has a vector for each millisecond, then the image is 1 by 1000 pixels. The one-dimensional convolutional neural network (“1D CNN”) trainer 310 learns the weights of activation functions for the convolutional neural network using the training data 301. To generate a classification for a patient, the MLMO system provides the VCG 302 for the patient as a one-dimensional image. The 1D CNN 320 then classifies the VCG based on the weights and outputs the classification, such as rotor location.
CNNs are a type of neural network that has been developed specifically to process images. A CNN may be used to input an entire image and output a classification of the image. For example, a CNN can be used to automatically determine whether a scan of a patient indicates the presence of an anomaly (e.g., tumor). The MLMO system considers the derived EM data to be a one-dimensional image. A CNN has multiple layers such as a convolution layer, a rectified linear unit (“ReLU”) layer, a pooling layer, a fully connected (“FC”) layer, and so on. Some more complex CNNs may have multiple convolution layers, ReLU layers, pooling layers, and FC layers.
A convolution layer may include multiple filters (also referred to as kernels or activation functions). A filter inputs a convolution window of an image, applies weights to each pixel of the convolution window, and outputs an activation value for that convolution window. For example, if the image is 256 by 256 pixels, the convolution window may be 8 by 8 pixels. The filter may apply a different weight to each of the 64 pixels in a convolution window to generate the activation value also referred to as a feature value. The convolution layer may include, for each filter, a node (also referred to as a neuron) for each pixel of the image assuming a stride of one with appropriate padding. Each node outputs a feature value based on a set of weights for the filter that are learned during a training phase for that node. Continuing with the example, the convolution layer may have 65,536 nodes (256*256) for each filter. The feature values generated by the nodes for a filter may be considered to form a convolution feature map with a height and width of 256. If an assumption is made that the feature value calculated for a convolution window at one location to identify a feature or characteristic (e.g., edge) would be useful to identify that feature at a different location, then all the nodes for a filter can share the same set of weights. With the sharing of weights, both the training time and the storage requirements can be significantly reduced. If each pixel of an image is represented by multiple colors, then the convolution layer may include another dimension to represent each separate color. Also, if the image is a 3D image, the convolution layer may include yet another dimension for each image within the 3D image. In such a case, a filter may input a 3D convolution window.
The ReLU layer may have a node for each node of the convolution layer that generates a feature value. The generated feature values form a ReLU feature map. The ReLU layer applies a filter to each feature value of a convolution feature map to generate feature values for a ReLU feature map. For example, a filter such as max(0, activation value) may be used to ensure that the feature values of the ReLU feature map are not negative.
The pooling layer may be used to reduce the size of the ReLU feature map by downsampling the ReLU feature map to form a pooling feature map. The pooling layer includes a pooling function that inputs a group of feature values of the ReLU feature map and outputs a feature value. For example, the pooling function may generate a feature value that is an average of groups of 2 by 2 feature values of the ReLU feature map. Continuing with the example above, the pooling layer would have 128 by 128 pooling feature map for each filter.
The FC layer includes some number of nodes that are each connected to every feature value of the pooling feature maps. For example, if an image is to be classified as being a cat, dog, bird, mouse, or ferret, then the FC layer may include five nodes whose feature values provide scores indicating the likelihood that an image contains one of the animals. Each node has a filter with its own set of weights that are adapted to the type of the animal that the filter is to detect.
In the following, the MLMO system is described in reference to the following data structures. The brackets indicate an array. For example, VCG[2].V[5].x represents the voltage for the x-axis for the fifth time interval in the second VCG. The data structures are further described below when first referenced.


VCG data structures
VCG[ ]

	Size
	V[ ]

	x
	y
	z

nVCG

V[ ]

	x
	y
	z

Cycles data structure

#C

C[ ]

	start
	end

Training data structure

#TD

TD[ ]

	nVCG[ ]
	label(s)

FIG. 4 is a flow diagram that illustrates detailed processing of the generate classifier component of the MLMO system in some embodiments. The generate classifier component 400 is invoked to generate a classifier. In block 401, the component invokes a generate simulated VCGs component to simulate VCGs (VCG[ ]) for a variety of parameter sets. In blocks 402-405, the component loops, generating the training data for each simulation. In block 402, the component sets an index i to 1 for indexing the parameter sets. In decision block 403, if index i is equal to the number of parameter sets, then all the training data has been generated and the component continues at block 406, else the component continues at block 404. In block 404, the component invokes a generate training data component, passing an indication of the indexed parameter set. In block 405, the component increments index i and then loops to block 403. In block 406, the component invokes a train classifier component to train the classifier based on the generated training data and then completes.
FIG. 5 is a flow diagram that illustrates the processing of a generate simulated VCGs component of the MLMO system in some embodiments. The generate simulated VCGs component 500 is invoked to generate a simulated VCG for each parameter set. In block 501, the component sets an index i to 1 for indexing through the parameter sets. In decision block 502, if index i is greater than the number of parameter sets, then the component completes, else the component continues at block 503. In block 503, the component sets an index j to 1 for indexing through the simulation steps. In decision block 504, if index j is greater than the number of simulation steps, then the simulation for the indexed parameter set is complete and the component continues at block 507, else the component continues at block 505. In block 505, the component applies the computational model based on the indexed parameter set and the indexed simulation step to generate a voltage solution (VS[j]) for the indexed simulation step. In block 506, the component increments index j and then loops to block 504 to process the next simulation step. In block 507, the component generates a VCG (VCG[i]) for the indexed parameter set from the voltage solution (VS[ ]) that was calculated for the parameter set. In block 508, the component increments index i and then loops to block 502 to process the next parameter set.
FIG. 6 is a flow diagram that illustrates the processing of a generate training data component for cycles of the MLMO system in some embodiments. The generate training data component 600 is invoked, passing an index i, that indexes a VCG generated for a parameter set and generates the training data from the VCG. In block 601, the component invokes an identify cycles component, passing an indication of the indexed VCG (VCG[i]) and receiving a normalized VCG (nVCG[ ]) for each cycle along with a count (#C) of the cycles that were identified. In block 602, the component sets an index k to 1 for indexing through the cycles. In decision block 603, if index k is greater than the count of the cycles, then the training data for all the cycles of the indexed VCG has been generated and the component completes, else the component continues at block 604. In block 604, the component increments a running count (#TD) of the training data (TD) that is used as an index into the training data. In block 605, the component sets the normalized nVCG of the indexed training data (TD[#TD].nVCG) to the portion of the VCG specified by the indexed cycle. The component extracts the portion from the x-axis, y-axis, and z-axis as defined by the start and end points of the cycle. In block 606, the component sets the label(s) of the indexed training data based on the function of the indexed parameter set (e.g., rotor location). In block 607, the component increments index k to index to the next cycle and then loops to block 603 to process the next cycle.
FIG. 7 is a flow diagram that illustrates the processing of an identify cycles component of the MLMO system in some embodiments. The identify cycles component 700 is invoked to identify the cycles within a VCG and provides the normalized VCGs (nVCG[ ]) for the cycles. In block 701, the component initializes an index j to 2 for indexing through the VCG and sets an index k to 0 for indexing through the identified cycles. In decision block 702, if index j is greater than the size of the VCG, then the component has identified all the cycles and the component completes, providing the normalized nVCG, else the component continues at block 703. In block 703, if the prior voltage of the x-axis of the VCG (VCG.V[j−1].x) is greater than or equal to zero and the indexed voltage of the x-axis of the VCG (VCG.V[j].x) is less than zero (i.e., a negative crossing of the x-axis), then the start of a possible cycle has been identified and the component continues at block 704 to identify the cycle, else the component continues at block 709. In block 704, the component sets the start of the indexed cycle (C[k].start) equal to index j. In decision block 705, if at least one cycle has already been identified, then the end of the prior cycle is known and the component increments index k and continues at block 706, else the component increments index k and continues at block 709. In block 706, the component sets the end of the prior cycle to index j−1. In block 707, the component extracts the VCG (eVCG) for the prior indexed cycle delimited by the start and the end of the prior cycle. In block 708, the component invokes a normalize cycle component, passing an indication of the extracted VCG (eVCG), and receives the normalized cycle (nVCG). In block 709, the component increments the index j for indexing through the VCG and loops to block 702.
FIG. 8 is a block diagram that illustrates the processing of a normalize cycle component of the MLMO system in some embodiments. The normalize cycle component 800 is invoked, passing an indication of the VCG of a cycle, and normalizes the cycle. In block 801, the component identifies the maximum vector magnitude V′ of the vectors in the cycle. For example, a vector magnitude of a vector may be calculated by taking the square root of the sum of the squares of the x, y, and z values of the vector. In block 802, the component sets index i to index a next axis of the VCG. In decision block 803, if all the axes have already been selected, then the component completes, providing the normalized VCG, else the component continues at block 804. In block 804, the component initializes an index j to 1 for indexing through the vectors of a normalized cycle. In decision block 805, if index j is greater than the number of vectors of a normalized cycle, then the component loops to block 802 to select the next axis, else the component continues at block 806. In block 806, the component sets the normalized VCG for the indexed vector for the indexed axis to an interpolation of the passed VCG, the indexed vector, and the maximum vector magnitude V′. The interpolation effectively compresses or expands the VCG to the number of vectors in the normalized VCG and divides the x, y, and z values of the vector by the maximum vector magnitude V′. In block 807, the component increments the index j and then loops to block 805.
FIG. 9 is a flow diagram that illustrates processing of a generate training data for a sequence of similar cycles component of the MLMO system in some embodiments. The generate training data for a sequence of similar cycles component 900 is invoked to identify sequences of two consecutive cycles of the VCG indexed by the passed index i that are similar and generate training data based on the identified sequences of similar cycles. The cycles in a sequence are similar according to a similarity score that reflects the stability of the cycles. In block 901, the component invokes the identify cycles component to identify the cycles (nVCG[ ]) for the VCG. In block 902, the component sets an index j to 2 for indexing through the identified cycles. In decision block 903, if index j is greater than the number of identified cycles, then all the cycles have been indexed and then component completes, else the component continues at block 904. In block 904, the component generates a similarity score for the cycles indexed by j−1 and j. The similarity score may be based on, for example, a cosine similarity, a Pearson correlation, and so on. In decision block 905, if similarity score is above a similarity score threshold (T) indicating similar cycles, then a sequence of similar cycles has been identified and the component continues at block 906, else the component continues at block 909. In block 906, the component increments a running count (#TD) of the training data. In block 907, the component sets the training data to the sequence of similar cycles. In block 908, the component sets the label for the training data to a label derived from the parameter set (PS[i]) used to generate the VCG and then continues at block 909. In block 909, the component increments index i to select the next sequence of cycles and loops to block 903.
FIG. 10 is a flow diagram that illustrates the processing of a classify component of the MLMO system in some embodiments. The classify component 1000 is invoked, passing a VCG derived from a patient, and outputs a classification. In block 1001, the component invokes the identify cycles component, passing an indication of the VCG, and receives the normalized VCGs for the cycles and a count of cycles. In block 1002, the component sets an index k to 1 for indexing through the cycles. In decision block 1003, if index k is greater than the number of cycles, then the component completes with the classifications, else the component continues at block 1004. In block 1004, the component applies the classifier to the indexed cycle to generate the classification. In block 1005, the component increments the index and then loops to block 1003 to process the next cycle. A different classification (e.g., different rotor location) may be generated for each cycle. In such a case, the overall classification may be derived from the combination of the different classifications (e.g., average of the rotor locations).
The following paragraphs describe various embodiments of aspects of the message interface system. An implementation of the message interface system may employ any combination of the embodiments. The processing described below may be performed by a computing system with a processor that executes computer-executable instructions stored on a computer-readable storage medium that implements the message interface system.
In some embodiments, a method performed by one or more computing systems is provided for generating a classifier for classifying electromagnetic data derived from an electromagnetic source within a body. The method accesses a computational model of the electromagnetic source, the computational model for modeling electromagnetic output of the electromagnetic source over time based on a source configuration of the electromagnetic source. For each of a plurality of source configurations, the method generates using the computational model a modeled electromagnetic output of the electromagnetic source for that source configuration. The method, for each modeled electromagnetic output, derives the electromagnetic data for the modeled electromagnetic output and generates a label for the derived electromagnetic data based on the source configuration for the modeled electromagnetic data. The method trains a classifier with the derived electromagnetic data and the labels as training data. In some embodiments, the modeled electromagnetic output for a source configuration includes, for each of a plurality of time intervals, an electromagnetic mesh with a modeled electromagnetic value for each of a plurality of locations of the electromagnetic source. In some embodiments, the derived electromagnetic data, for a time interval, is an equivalent source representation of the electromagnetic output. In some embodiments, the equivalent source representation is generated using principal component analysis. In some embodiments, the method further identifies cycles within the derived electromagnetic data for a modeled electromagnetic output. In some embodiments, the same label is generated for each cycle. In some embodiments, the method further identifies a sequence of cycles that are similar and wherein the same label is generated for each sequence. In some embodiments, the deriving of the electromagnetic data for a modeled electromagnetic output includes normalizing the modeled electromagnetic output on a per-cycle basis. In some embodiments, the classifier is a convolutional neural network. In some embodiments, the convolutional neural network inputs a one-dimensional image. In some embodiments, the classifier is a recurrent neural network, an autoencoder, or a restricted Boltzmann machine. In some embodiments, the classifier is a support vector machine. In some embodiments, the classifier is Bayesian. In some embodiments, the electromagnetic source is a heart, a source configuration represents source location and other properties of a heart disorder, the modeled electromagnetic output represents activation of the heart, and the electromagnetic data is based on body-surface measurements such as an electrocardiogram. In some embodiments, the heart disorder is selected from a set consisting of atrial fibrillation, ventricular fibrillation, atrial tachycardia, ventricular tachycardia, atrial flutter, and premature ventricular contractions.
In some embodiments, a method performed by a computing system is provided for classifying electromagnetic output collected from a target that is an electromagnetic source within a body. The method accesses a classifier to generate a classification for electromagnetic output of an electromagnetic source. The classifier is trained using training data generated from modeled electromagnetic output for a plurality of source configurations of an electromagnetic source. The modeled electromagnetic output is generated using a computational model of the electromagnetic source that models the electromagnetic output of the electromagnetic source over time based on a source configuration. The method collects target electromagnetic output from the target. The method applies the classifier to the target electromagnetic output to generate a classification for the target. In some embodiments, the training data is generated by running, for each of the source configurations, a simulation that generates an electromagnetic mesh for each of a plurality of simulation intervals, each electromagnetic mesh having an electromagnetic value for a plurality of locations of the electromagnetic source. In some embodiments, the electromagnetic source is a heart, a source configuration represents a source location of a heart disorder, and the modeled electromagnetic output represents activation of the heart, and the classifier is trained using electromagnetic data derived from an electrocardiogram representation of the electromagnetic output.
In some embodiments, one or more computing systems are provided for generating a classifier for classifying electromagnetic output of an electromagnetic source. The one or more computing systems include one or more computer-readable storage mediums and one or more processors for executing the computer-executable instructions stored in the one or more computer-readable storage mediums. The one or more computer-readable storage mediums store a computational model of the electromagnetic source. The computational model is modeling electromagnetic output of the electromagnetic source over time based on a source configuration of the electromagnetic source. The one or more computer-readable storage mediums store computer-executable instructions for controlling the one or more computing systems to for each of a plurality of source configurations, generate training data from the electromagnetic output of the computational model that is based on the source configuration and train the classifier using the training data. In some embodiments, the computer-executable instructions to generate the training data for a source configuration further control the one or more computing systems to generate derived electromagnetic data from the electromagnetic output for the source configuration and generate a label for the electromagnetic data based on the source configuration.
Although the subject matter has been described in language specific to structural features and/or acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims. In some embodiments, the MLMO system can be employed to classify electromagnetic output of an electromagnetic source based on different types of classifications. For example, the classifications may include location of a heart disorder (e.g., rotor location), scar location, heart geometry (e.g., ventricle orientation), and so on. To generate the training data, the MLMO system labels the training data with the classification type that the classifier is to generate. Accordingly, the invention is not limited except as by the appended claims.

Claims

1. A method performed by one or more computing systems for generating a classifier for classifying electromagnetic data derived from an electromagnetic source within a body, the method comprising:

accessing a computational model of the electromagnetic source, the computational model for modeling electromagnetic output of the electromagnetic source over time based on a source configuration of the electromagnetic source;

for each of a plurality of source configurations, generating using the computational model a modeled electromagnetic output of the electromagnetic source for that source configuration;

for each modeled electromagnetic output,

deriving the electromagnetic data for the modeled electromagnetic output; and

generating a label for the derived electromagnetic data based on the source configuration for the modeled electromagnetic data; and

training a classifier with the derived electromagnetic data and the labels as training data.

2. The method of claim 1 wherein the modeled electromagnetic output for a source configuration includes, for each of a plurality of time intervals, an electromagnetic mesh with a modeled electromagnetic value for each of a plurality of locations of the electromagnetic source.

3. The method of claim 2 wherein the derived electromagnetic data, for a time interval, is an equivalent source representation of the electromagnetic output.

4. The method of claim 3 wherein the equivalent source representation is generated using principal component analysis.

5. The method of claim 3 further comprising identifying cycles within the derived electromagnetic data for a modeled electromagnetic output.

6. The method of claim 5 wherein the same label is generated for each cycle.

7. The method of claim 5 further comprising identifying a sequence of cycles that are similar and wherein the same label is generated for each sequence.

8. The method of claim 5 wherein the deriving of the electromagnetic data for a modeled electromagnetic output includes normalizing the modeled electromagnetic output on a per-cycle basis.

9. The method of claim 1 wherein the classifier is a convolutional neural network.

10. The method of claim 9 wherein the convolutional neural network inputs a one-dimensional image.

11. The method of claim 1 wherein the classifier is a recurrent neural network, an autoencoder, or a restricted Boltzmann machine.

12. The method of claim 1 wherein the classifier is a support vector machine.

13. The method of claim 1 wherein the classifier is Bayesian.

14. The method of claim 1 wherein the electromagnetic source is a heart, a source configuration represents source location and other properties of a heart disorder, the modeled electromagnetic output represents activation of the heart, and the electromagnetic data is based on body-surface measurements such as an electrocardiogram.

15. The method of claim 14 wherein the heart disorder is selected from a set consisting of atrial fibrillation, ventricular fibrillation, atrial tachycardia, ventricular tachycardia, atrial flutter, premature ventricular complexes, atrioventricular nodal reentrant tachycardia, atrioventricular reentrant tachycardia, and junctional tachycardia.

16. A method performed by a computing system for classifying electromagnetic output collected from a target that is an electromagnetic source within a body, the method comprising:

accessing a classifier to generate a classification for electromagnetic output of an electromagnetic source, the classifier trained using training data generated from modeled electromagnetic output for a plurality of source configurations of an electromagnetic source, the modeled electromagnetic output being generated using a computational model of the electromagnetic source that models the electromagnetic output of the electromagnetic source over time based on a source configuration;

collecting target electromagnetic output from the target; and

applying the classifier to the target electromagnetic output to generate a classification for the target.

17. The method of claim 16 wherein the training data is generated by running, for each of the source configurations, a simulation that generates an electromagnetic mesh for each of a plurality of simulation intervals, each electromagnetic mesh having an electromagnetic value for a plurality of locations of the electromagnetic source.

18. The method of claim 16 wherein the electromagnetic source is a heart, a source configuration represents a source location of a heart disorder, and the modeled electromagnetic output represents activation of the heart, and the classifier is trained using electromagnetic data derived from an electrocardiogram representation of the electromagnetic output.

19. One or more computing systems for generating a classifier for classifying electromagnetic output of an electromagnetic source, the systems comprising:

one or more computer-readable storage mediums storing:

a computational model of the electromagnetic source, the computational model for modeling electromagnetic output of the electromagnetic source over time based on a source configuration of the electromagnetic source; and

computer-executable instructions for controlling the one or more computing systems to:

for each of a plurality of source configurations, generate training data from the electromagnetic output of the computational model that is based on the source configuration; and

train the classifier using the training data; and

one or more processors for executing the computer-executable instructions stored in the one or more computer-readable storage mediums.

20. The one or more computing systems of claim 19 wherein the computer-executable instructions to generate the training data for a source configuration further control the one or more computing systems to generate derived electromagnetic data from the electromagnetic output for the source configuration and generate a label for the electromagnetic data based on the source configuration.