Neural network trained preprocessing using intracardiac electrograms to detect activation in electrograms

Abstract

The invention is entitled "Neural Network Trained Preprocessing Detection of Activation in Electrograms Using Intracardiac Electrograms". The invention discloses a method comprising acquiring a plurality of bipolar electrograms and corresponding unipolar electrograms of a patient, the electrograms including annotations in which one or more human reviewers have identified and marked windows of interest and one or more activation times within the windows of interest. Generating a ground truth dataset from the electrograms for training at least one electrogram preprocessing step of a machine learning (ML) algorithm. Applying the ML algorithm to the electrograms to train at least the at least one electrogram preprocessing step in order to detect an occurrence of an activation in a given bipolar electrogram within the window of interest.

Description

Technical Field

[0001] The present invention relates generally to the analysis of intracardiac electrophysiological signals, and more specifically to the use of machine learning (ML) to assess electrical propagation in the heart. Background Technology

[0002] In patients with arrhythmias, impulses of electrical activity can travel along pathological pathways within the heart tissue. In electrophysiological studies, one goal for physicians is to understand the pathways of these impulses, for example, by inserting a multi-electrode catheter into the patient's heart and measuring intracardiac ECG signals via the catheter. Typically, the catheter is inserted through a blood vessel to reach the heart to measure electrical activity in the endocardium. Alternatively, the catheter may be inserted through an incision in the chest to measure electrical activity in the epicardium.

[0003] To measure electrical activity, some electrodes of the catheter are brought into contact with cardiac tissue (endocardium or epicardium) to obtain the potential difference between the electrodes and ground potential (e.g., grounded at the Wilson center). The potential difference between the catheter electrodes and ground is called the unipolar signal of the electrode. The potential difference between two adjacent electrodes of the catheter is called the bipolar signal between the electrodes.

[0004] In electrophysiological studies, to understand the path taken by the pulses of electrical activity, physicians need to know the time it takes for the pulses to travel under each electrode. However, by looking at the potential difference between bipolar pairs, it is not possible to know when the pulses of electrical activity travel under each electrode.

[0005] As described in U.S. Patent Application Publication 2018 / 0042504, this problem can be overcome by simultaneously analyzing bipolar and unipolar signals, as further described below. However, interpreting unipolar signals is challenging, particularly due to noise in the unipolar signal itself. For example, unlike bipolar signals, unipolar signals can contain excessive far-field noise. This makes the analysis of unipolar signals challenging.

[0006] As described above, U.S. Patent Application Publication 2018 / 0042504 previously proposed using both bipolar and unipolar signals for computer-aided analysis of intracardiac signals to determine the timing of cardiac activation. This U.S. patent application describes a rule-based algorithm and method that includes receiving bipolar signals from a pair of electrodes near the myocardium of a human subject, and receiving unipolar signals from a selected electrode of the pair. The method also includes calculating the local unipolar minimum derivative of the unipolar signal and the time of occurrence of the unipolar minimum derivative. Furthermore, the method includes calculating the bipolar derivative of the bipolar signal, evaluating the ratio of the bipolar derivative to the local unipolar minimum derivative, and identifying the time of occurrence as the myocardial activation time when the ratio is greater than a preset threshold ratio. Summary of the Invention

[0007] The embodiments of the present invention described below provide a method comprising acquiring multiple bipolar electrograms and corresponding unipolar electrograms of a patient, the electrograms including annotations, wherein one or more human reviewers have identified and labeled windows of interest and one or more activation times within said windows of interest. A ground truth labeled dataset is generated from the electrograms for at least one electrogram preprocessing step to train a machine learning (ML) algorithm. The ML algorithm is applied to the electrograms to train at least one electrogram preprocessing step to detect the occurrence of activations within a given bipolar electrogram within a window of interest.

[0008] In some implementations, each bipolar electrogram is obtained from a pair of electrodes placed at a location in the patient's heart, and a corresponding unipolar electrogram is obtained from one of the electrodes in the pair.

[0009] In some implementations, acquiring bipolar electrograms and corresponding unipolar electrograms includes acquiring multiple bipolar electrograms and corresponding unipolar electrograms from multiple electrode pairs of a multi-electrode conduit.

[0010] In one implementation, at least one electrogram preprocessing step includes performing one or more convolutions of the electrogram using a set of convolution kernels, and wherein training at least one electrogram preprocessing step includes specifying the coefficients of the convolution kernels.

[0011] In another embodiment, at least one preprocessing step of the electrogram includes pointwise multiplication between a bipolar electrogram filtered by one of the convolution kernels and a corresponding unipolar electrogram filtered by another of the convolution kernels, and wherein applying the ML algorithm includes inputting the multiplied signal obtained by the pointwise multiplication into the ML algorithm.

[0012] In some implementations, the method further includes presenting a given bipolar electrogram to a user, on which annotations are provided to mark the times of detected activation.

[0013] In some implementations, applying ML algorithms includes applying artificial neural networks (ANNs).

[0014] In some implementations, the method further includes receiving a bipolar electrogram and a corresponding unipolar electrogram of the patient, along with a window of interest for inference. Using a trained machine learning algorithm, the occurrence of activations in the bipolar electrogram is detected, and the activations are correlated with cardiac tissue locations that are in contact with the electrodes acquiring the corresponding unipolar electrograms.

[0015] According to another embodiment of the invention, a system comprising one or more displays and one or more processors is also provided. The one or more displays are configured to present annotated electrograms. The one or more processors are configured to (a) acquire multiple bipolar electrograms and corresponding unipolar electrograms of a patient, the electrograms including annotations, wherein one or more human reviewers have used the one or more processors and corresponding one or more displays to identify and label windows of interest and one or more activation times within those windows of interest; (b) generate a real-label dataset based on the electrograms in at least one of the one or more processors for at least one electrogram preprocessing step to train a machine learning (ML) algorithm; and (c) apply the ML algorithm to the electrograms to train at least one electrogram preprocessing step to detect the occurrence of activations within a given bipolar electrogram within a window of interest.

[0016] According to another embodiment of the invention, a method is also provided for identifying activations in cardiac electrograms using a machine learning (ML) model having at least one electrogram preprocessing step, the method comprising at least one electrogram preprocessing step for training the ML model. Using the ML model, activations in bipolar electrograms are identified within a window of interest using corresponding unipolar electrograms.

[0017] In some implementations, training at least one electrogram preprocessing step includes optimizing the coefficients of the convolution kernel that convolves with the electrogram during preprocessing.

[0018] In some implementations, identifying activations involves generating a probability for each of a plurality of possible activation values ​​in a window of interest using an ML model.

[0019] In one implementation, identifying activations includes selecting one of the following: (i) the activation with the highest probability, (ii) one or more activations with a probability higher than a given threshold, and (iii) one or more activations with a probability higher than a variable threshold.

[0020] According to another embodiment of the present invention, a computer software product is also provided, the product comprising a tangible non-transitory computer-readable medium therein storing program instructions, which, when read by a processor, cause the processor to:

[0021] A machine learning (ML) algorithm, including at least one trainable electrogram preprocessing step, is applied to a ground-labeled dataset of annotated electrograms to train at least one electrogram preprocessing step of the ML algorithm to detect the occurrence of activations within a given bipolar electrogram within a window of interest. The ground-labeled dataset of annotated electrograms is generated by acquiring multiple bipolar electrograms and corresponding unipolar electrograms from patients, the electrograms including annotations, wherein one or more human reviewers have identified and labeled the window of interest and one or more activation times within the window of interest. Attached Figure Description

[0022] This disclosure will be more fully understood through the following detailed description of embodiments thereof, taken in conjunction with the accompanying drawings, wherein:

[0023] Figure 1 This is a schematic diagram of a catheter-based electrophysiological (EP) mapping system according to an exemplary embodiment of the present invention, which is configured to detect activation time in an electrogram (EGM);

[0024] Figure 2 This is a schematic diagram illustrating the workflow of training and deploying an algorithm for detecting activation times in an electrogram (EGM) according to an exemplary embodiment of the present invention.

[0025] Figure 3 A block diagram illustrating an algorithm for detecting activation time in an electrogram (EGM) according to an exemplary embodiment of the present invention;

[0026] Figure 4 The flowchart illustrates a method and algorithm for detecting activation time, annotating activation, and generating activation maps in an electrogram (EGM) according to an exemplary embodiment of the present invention.

[0027] Figure 5 To illustrate an exemplary embodiment of the present invention Figure 4 A flowchart of the algorithm's training method; and

[0028] Figures 6A to 6C To illustrate the following exemplary embodiments according to the present invention Figure 4 The algorithm determines the probability distribution of LAT values ​​by selecting one or more Local Activation Time (LAT) values ​​from a graph. Detailed Implementation

[0029] Overview

[0030] Intracardiac electrophysiological (EP) mapping is a catheter-based method that can sometimes be used to characterize abnormalities in cardiac EP wave propagation, such as those leading to arrhythmias. In a typical catheter-based procedure, the distal end of a catheter, including multiple sensing electrodes, is inserted into the heart to acquire a set of data points. This set of data points includes (i) the measurement location on the wall tissue of the heart chambers and (ii) the corresponding EP signal, from which the EP mapping system generates an EP mapping map. An example of an EP mapping map that can be used to diagnose arrhythmias is an EP timing mapping map of a region of the heart chamber wall tissue (called a local activation time (LAT) mapping map).

[0031] To generate a LAT mapping, the processor may have to analyze intracardiac ECG signals (hereinafter referred to as electrograms (EGMs)) acquired at various points in the heart to identify activations in each signal (i.e., in the waveform of the EGM), and annotate the activations and calculate LAT values.

[0032] Annotation time refers to the time it takes for a cardiac impulse to travel to a specific location in cardiac tissue, such as when measured by catheter. Since different EGM signals are acquired at different times, gating measurement techniques are needed to align signals acquired at different times. In cardiac electrophysiology, the "gate" used for gating measurements is called the reference annotation. For example, the R peak of the QRS signal on a surface ECG, or activation detected in the coronary sinus, activation detected in the high right atrium, or more complex methods fed by multiple signals can be used as the reference annotation. The LAT value is defined as the difference between the mapped annotation time and the reference annotation time.

[0033] The embodiments of the present invention described below provide a machine learning (ML) technique and algorithm configured to train both a preprocessing module and an artificial neural network (ANN) model. After preprocessing pairs of bipolar signals and corresponding unipolar EGM signals by the trained preprocessing module, the processor analyzes the preprocessed signals to determine the precise timing of activation under each catheter electrode in the catheter electrode assembly.

[0034] In preparation for training, multiple electrocardiograms are collected from multiple sites (e.g., hospitals), including annotations, where one or more human reviewers have identified and labeled activation times using one or more processors. Typically, the activation annotations are validated and / or corrected by multiple experts skilled in analyzing diagnostic cardiac electrocardiograms.

[0035] One or more technicians (e.g., clinical application engineers or algorithm engineers) compile a set of annotated electrograms into a real-label dataset, which will be used to train the provided ML algorithm.

[0036] During training and inference, two components of the algorithm (i.e., preprocessing and ANN) work together to identify activations in the electrograph input. To do this, the processor applies an ML algorithm to the electrograph to train at least one electrograph preprocessing step to detect the occurrence of activations in a given bipolar electrograph.

[0037] In some implementations, during training or inference, the disclosed model searches for activations present in both the bipolar EGM signal and the corresponding unipolar EGM signal. Using the unipolar EGM, the model determines the timing and location of activation on the bipolar EGM (e.g., below the electrode acquiring the unipolar EGM). When applied to multiple electrograms, the disclosed algorithm generates a reliable database for generating electrophysiologically accurate LAT mappings of at least a portion of the cardiac chambers.

[0038] To explicitly define the algorithm, an ML model (e.g., an ANN model) is applied to both unipolar and bipolar signals via a sliding window. The output of the ANN model is a probability associated with the LAT value, or a set of probabilities associated with each possible LAT value in the window of interest.

[0039] Training is typically performed using loss functions such as cross-entropy, where the loss is defined as zero (0), for example, when the calculated LAT value equals the true label LAT value. Loss function parameters can include a set of differences between the algorithm-estimated LAT value and the corresponding validated LAT value (true label). The LAT differences are explicitly limited to a window of interest (WOI), such as an effective width having the length of the cardiac cycle. Parameters outside the WOI are ignored.

[0040] The loss function is defined such that zero LAT difference contributes zero loss to the total loss, while the contributed loss asymptotically approaches a maximum value (e.g., normalized value 1) when the difference between the calculated LAT value and the expected LAT value is higher than a predefined value (such as 5 milliseconds). For the EP signal analyzed in this patent application, a 5-millisecond width of the loss function (e.g., at loss = 1 / 2) was found to be sufficient for the ML algorithm to converge fast enough with sufficient estimation accuracy (e.g., ±1 millisecond) in the presence of noise.

[0041] Typically, during training, the preprocessing module of the disclosed technical optimization algorithm includes appropriate high-pass, low-pass, band-pass, and / or band-stop convolutional filters (i.e., kernels) to flatten and smooth the EGM signal and enhance the desired features of the EGM signal, thereby enabling accurate determination of the presence of activation indicator features that exist simultaneously in both bipolar and unipolar signals.

[0042] Searching for such activation indicator features that exist simultaneously in both bipolar and unipolar EGMs is roughly equivalent to searching for such patterns in the multiplication of signals. Therefore, in one implementation, the preprocessing module multiplies the bipolar and unipolar EGMs acquired at the same locations point-by-point. The resulting multiplied signal is concatenated with the signal itself and input into the ANN. Besides using the signal itself, the advantage of using the multiplied signal in the ANN's input layer is that it improves the ANN model's ability to detect activation indicator amplitudes, because multiplication enhances the temporal coincidence of such amplitudes in the corresponding unipolar signal within the bipolar signal.

[0043] Convolutional filters are optimized through training, for example, to overcome low-frequency drift baseline artifacts and high-frequency noise artifacts. First, the algorithm learns a sufficiently optimized convolutional filter (by training the neural network using multiple signals). In the very first iteration of training, the convolutional kernel is typically composed of random numbers or other digits such as "all ones" or "all zeros". Regardless, the training iterations optimize the number of kernels and the ANN module as a whole.

[0044] Once the algorithm applies a fully optimized convolutional filter to both the bipolar signal and its corresponding unipolar signal, the multiplication of the convolutions has an enhancing effect on training, further optimizing the preprocessing and ANN parameters.

[0045] In some implementations, multiple convolutional filters can be applied in parallel to the same signal. Each convolutional filter can be trained to extract different features of either a bipolar or unipolar signal.

[0046] This training generates optimized parameters for the preprocessing module, such as the kernel parameter vector. And the optimization parameters of ANN, such as weights and bias vectors. The optimized parameters can be transmitted to the user who is using the technology, for example, via a secure memory stick or via the Internet, and the user has the algorithm pre-installed on the user's site.

[0047] Many types of ML models are available, and those skilled in the art can choose from various ML models besides the ANN model used as an example herein, including decision tree learning, support vector machines (SVM), and Bayesian networks. ANN models include, for example, convolutional neural networks (CNN), recurrent neural networks (RNN), long short-term memory networks (LSTM), reinforcement learning, autoencoders, and probabilistic neural networks (PNN). Typically, one or more processors used (collectively referred to as "processors") are programmed in software containing a specific algorithm that enables the processor to perform each of the processor-related steps and functions listed above. Training is typically performed using a computing system that includes multiple processors, such as graphics processing units (GPUs) or tensor processing units (TPUs). However, any one of these processors can also be a central processing unit (CPU).

[0048] In some implementations, the disclosed technique detects a single activation (i.e., finding a LAT value for each input electrogram waveform). In other implementations, the technique is used to detect fragmented ECG signals, regional anomalous ventricular activation (LAVA) indication signals, delayed potentials, biphasic potentials, and splitting potentials. Implementations designed to detect regular activations may select the LAT value with the highest probability. Implementations designed to detect biphasic potentials, splitting potentials, fragmented potentials, fragmented potentials, or any kind of multiple potentials may select all LAT values ​​with a probability higher than a given constant threshold. Implementations designed to detect delayed potentials may use a variable threshold (e.g., decreasing) such that the threshold is higher the later the LAT. Implementations for detecting early potentials may do the opposite.

[0049] After training with 800,000 atrial ECG signals, the inventors achieved a 95% success rate in analyzing 3.1 million ECGs that were not involved in the training (i.e., by performing inference using a trained algorithm).

[0050] The disclosed ANN-based EP signal analysis technology can improve the value of diagnostic catheter insertion procedures by providing physicians with reliable diagnoses.

[0051] System Description

[0052] Figure 1 This is a schematic illustration of a catheter-based electrophysiological (EP) mapping system 21 according to an exemplary embodiment of the present invention, which is configured to detect activation time in an electrogram (EGM). Figure 1A physician 27 is shown performing EP mapping of the heart 23 of a patient 25 using an EP mapping catheter 29. The mapping catheter 29 includes one or more arms 20 at its distal end, each of which is coupled to a bipolar electrode 22 including adjacent electrodes 22a and 22b. The bipolar signal acquired by the catheter 29 is defined herein as the potential difference between the adjacent electrodes 22a and 22b.

[0053] To acquire unipolar signals from electrodes 22a and / or 22b, a Wilson central terminal (WCT) grounded by three surface electrodes 24 (attached to the patient's skin) is formed. For example, the three surface electrodes 24 may be coupled to the patient's chest. (For ease of illustration, Figure 1 The image shows only one external electrode 24.

[0054] Bipolar signals are easier to interpret, but less precise. Correspondingly, unipolar signals are more precise, but more prone to far-field noise. If a user only views a bipolar signal, he / she cannot distinguish activation on the first or second electrode. If a user only views a unipolar signal, the user may obtain multiple false detections. To find the most accurate activation time, the disclosed technique applies an ML algorithm to simultaneously search for features predicting electrical activation in pairs of unipolar and bipolar signals, as described below.

[0055] During the mapping procedure, the position of electrode 22 is tracked while it is located within the patient's heart 23. For this purpose, electrical signals are transmitted between electrode 22 and surface electrodes 24 (typically, three additional surface electrodes may be coupled to the patient's back). Based on the signals and given the known location of electrode 24 on the patient's body, processor 28 calculates the estimated location of each electrode 22 within the patient's heart. This tracking can be performed using an active current location (ACL) system manufactured by Biosense-Webster (Irvine, California) as described in U.S. Patent No. 8,456,182, the disclosure of which is incorporated herein by reference.

[0056] The processor correlates the signals acquired by electrode 22 (such as unipolar EGM or bipolar EGM) with the location of the acquired signals. The processor 28 receives the signals via electrical interface 35 and uses the information contained in these signals to construct EP mapping map 31 and EGM 40, and presents these signals on display 26.

[0057] In the illustrated implementation, algorithms including ML algorithms (such as...) are used. Figure 3 (Disclosed) The processor 28 calculates local activation time (LAT) values ​​to generate LAT mapping maps. The ML algorithm is trained to find LAT values ​​at specific tissue locations.

[0058] like Figure 3 The training is primarily based on the following approach that a human expert reviewer might perform: examining the bipolar EGM signal (V22a-V22b) and, from that location, understanding the approximate timing of cardiac wavefront activation, then searching for activation pattern V22a in the corresponding unipolar EGM to find the precise timing of cardiac wavefront activation. As can be seen in the magnifying glass image, the identified irregularity 101 (albeit a very small feature) in the unipolar signal is located at the precise location of activation time (102), and the LAT value is calculated using the time difference between time 102 and baseline time 103. As the training steps iterate, the aforementioned ML algorithm is expected to mimic the decision-making of a human expert reviewer in the following way:

[0059] • The preprocessing parameters of the bipolar signal are shaped so that these parameters will enhance the activation features seen in the bipolar EGM.

[0060] • The preprocessing parameters of the unipolar signal are shaped to remove the drift baseline and enhance the activation features seen on the unipolar EGM.

[0061] • The machine learning parameters are shaped to interpret these features, thereby calculating the probability of activation for each LAT value within the window of interest.

[0062] The signals illustrated are simplified examples used only to illustrate the concept. In real life, signals may be much noisier than those illustrated.

[0063] Processor 28 typically includes a general-purpose computer having software programmed to perform the functions described herein. This software can be downloaded to the computer electronically via a network, or alternatively or additionally, it can be located and / or stored on a non-transitory tangible medium (such as magnetic storage, optical storage, or electronic storage). Specifically, processor 28 operates as disclosed herein, including... Figure 3 , Figure 4 and Figure 5 The processor 28 incorporates a dedicated algorithm that enables it to perform the disclosed steps, as further described below. In some cases, the processor 28 includes enhanced computing power, for example, through the use of the aforementioned GPU or TPU.

[0064] Figure 1 The exemplary illustrations shown are chosen solely for clarity of concept. Other types of electrophysiological sensing catheter geometries may also be employed, such as... The catheter (manufactured by Biosense-Webster Inc., Irvine, California). Additionally, a contact sensor may be fitted at the distal end of the mapping catheter 29 and transmit data indicating the physical quality of electrode contact with tissue. In an embodiment, if the physical contact quality of one or more electrodes 22 is indicated as poor, their measurements may be discarded, and if the contact quality of other electrodes is indicated as sufficient, their measurements may be considered valid.

[0065] Training and deploying trained ML models

[0066] Figure 2 This is a schematic diagram illustrating the training and deployment workflow of an algorithm (e.g., an algorithm including an ANN and a trainable preprocessing model) for detecting activation times in an electrophysiological record (EGM) according to one embodiment of the present invention. As shown, the acquisition of the EGM training set is completed at multiple sites 202, wherein at each site, a physician acquires data verified by the physician to utilize activation times (e.g., such as...). Figure 1 (Note 102) Multiple associated bipolar and unipolar EGMs are correctly annotated. Physicians can manually correct erroneous annotations. This validated annotation set is used during the ANN training session at training site 204 to fine-tune the annotations performed by the ANN during training by minimizing the ANN's loss function, thereby generating a set of ANN parameters. and preprocessing parameters For inferences used in the new EGM, such as Figure 3 As stated above.

[0067] The resulting trained ANN LAT activation detection algorithm can be stored in a memory stick (206) and provided to... Figure 1 Users of system 21 or other related diagnostic systems.

[0068] Figure 2 The exemplary cases shown are chosen solely for clarity of concept. For example, a trained model can be sent via a network instead of using a memory stick. Training can be performed using distributed computing instead of training stations.

[0069] Preprocessing of electrophoresis records trained by neural networks

[0070] Figure 3 A block diagram illustrating an algorithm for detecting activation time in an electrogram (EGM) according to one embodiment of the present invention.

[0071] As shown in the figure, the disclosed deep learning model is divided into two parts: a preprocessing module 300 and an ANN module 301. The preprocessing module 300 includes parameters to be optimized. A set of convolutional kernels (310). The ANN module 301 has a set of weights and biases to be optimized.

[0072] The raw EGM input to the model consists of associated portions of the raw unipolar and bipolar signals (93, 95), which are selected one by one by applying sliding windows 303 and 305 respectively.

[0073] During training, the intermediate optimized parameters and The parameters (93, 94) are backpropagated to different modules to optimize modules 300 and 301 respectively. The training process is iterative and terminates when a given termination criterion is reached, thus delivering the final set of optimized parameters. The termination criteria can be the convergence of the loss sum, the convergence of the parameters, the loss value reaching a value below a certain threshold, timeout, or a combination thereof.

[0074] In the illustrated implementation scheme, in order to optimize the parameters The loss function 311 of the ANN module 301 minimizes the difference loss between the known annotations of the EGM and the annotations calculated by the ANN.

[0075] As described above, the central feature of the disclosed technique (e.g., algorithm) is the training of the preprocessing step, for example, by optimizing the EGM signal (93, 95) applied to the time window (303, 305) (i.e., applied to the unipolar signal 302 (e.g., Figure 1 V22a) and bipolar signal 304 (e.g., Figure 1 The convolution kernels of V22a-V22b).

[0076] The preprocessing module 300 may include any number of convolutional kernels, with a typical number between four and eight, to increase training flexibility in finding the optimal kernel. Each convolutional kernel can "learn" to enhance different features of the signal. The preprocessing stage and the ML algorithm are optimized through training iterations to identify the features that need enhancement. For example, while it is reasonable to assume that the convolutional kernels associated with unipolar signals will learn to remove drift baselines, this is not defined by the algorithm's designer and should be a natural consequence of training.

[0077] Generally speaking, convolutional filters function as finite impulse response (FIR) filters. Convolution can be used as a low-pass filter, a high-pass filter, a band-stop filter, or a band-pass filter.

[0078] As mentioned above, many kernels can be included in the optimization parameters. (In searches for things like optimizing kernel type and sharpness).

[0079] A key feature of the preprocessing module 300 is the input multiplication signal. As shown in the figure, the unipolar and bipolar signals of the convolution are multiplied point by point (313), and the multiplication signal 306 is input together with the other convolution signals. As mentioned above, any number of convolution kernels "A" can be used, and the same number of different resulting multiplication signals are input into the ANN module 301 in the form of a concatenated input vector.

[0080] The motivation for using multiplied signals is to train modules 300 and 301 using the temporal correlation between unipolar and bipolar signals, such as... Figure 1 The example is illustrated using signals V22a-V22b and V22a.

[0081] During inference, the same deep learning model is used to analyze the new EGM signal, where the deep learning model parameters include optimized parameters. During inference, ML models (e.g., CNNs including the RESNET network) provide annotations of the activation times of the bipolar signals so that LAT values ​​can be extracted subsequently.

[0082] Some embodiments of the disclosed invention receive a fixed number of signal samples as input (e.g., receive a digitized signal for a given duration); however, physicians are typically interested in a specific window of interest, which depends on the cycle length of the patient's current arrhythmia. To overcome this problem, during both the training and inference phases, just before the distribution is normalized, for example, using a SoftMax function, the probability of LAT values ​​outside the window of interest is zeroed.

[0083] Setting all probability values ​​outside the window of interest to zero ensures that the sum of probabilities remains 1 after the normalization step, and also ensures that the window of interest is correctly considered during both the training and inference phases.

[0084] The following pseudocode example illustrates the convolution preprocessing and signal multiplication steps, where k convolutions are performed on the unipolar signal and k convolutions are performed on the bipolar signal. For each bipolar signal and its corresponding unipolar signal, p ≤ k, they are multiplied point-by-point to generate a p-multiplied signal. In this notation, The value is a parameter for different cores.

[0085]

[0086] Figure 3The exemplary block diagrams shown are chosen solely for conceptual clarity. ANN module 301 is a schematic module shown conceptually only, where ANN module 301 represents a selection from many possible ANN implementations (including library-coded ANN functions).

[0087] Methods for detecting activations in the EGM using a preprocessing module trained with an ANN. Figure 4 A flowchart illustrating a method and algorithm for detecting activations, annotating activations, and generating activation mapping maps according to an embodiment of the present invention is provided. According to the presented embodiment, the algorithm is divided into two parts: algorithm and training dataset preparation 1101 and algorithm usage 1102.

[0088] The algorithm is prepared to execute the following process, which begins with ANN modeling step 1201, where an ANN algorithm is generated for detecting activations in the EGM. This algorithm includes preprocessing training of associated unipolar and bipolar signals, such as... Figure 3 The algorithm includes an ANN-trainable preprocessing module 300.

[0089] Independently, the training database set preparation phase begins with real label acquisition step 1202, which includes acquiring electrograms from multiple sites (e.g., hospitals) including activated annotations validated and / or corrected by multiple experts skilled in analyzing diagnostic cardiac electrograms.

[0090] At step 1203, when generating the real-label dataset, a person skilled in the art (e.g., a clinical application engineer or algorithm engineer) compiles the real-label dataset that will be used to train the algorithm prepared in step 1201, based on the set of annotated electrographs.

[0091] Next, at the ML algorithm training step 1204, the processor uses a dataset of real-labeled annotated electrographs to train the algorithm (e.g., ANN and preprocessing parts), where the annotations were verified to be accurate in step 1202.

[0092] At step 1206, where the trained model is stored, the algorithm preparation concludes by storing the trained model on a non-transitory computer-readable medium such as a cryptographic disk (Memory Stick). In an alternative implementation, the model is sent in advance, and its optimization parameters are sent separately after training.

[0093] The algorithm uses step 1102 to execute the process that begins at step 1208 of the algorithm upload process, during which the user uploads the entire ML model or its optimization parameters (e.g., The signal is uploaded to the processor. Next, at EGM receiving step 1210, the processor, such as processor 28, receives bipolar ECG signals and unipolar ECG signals from conduit 29, for example.

[0094] Next, at EGM inference step 1212, the processor uses the trained ANN model to generate a set of probability distributions for activation times (typically, one distribution is generated for each bipolar EGM).

[0095] In some implementations, the peaks of the distribution, i.e., those exceeding those included in the ANN model, can be selected in subsequent steps to determine the annotations and corresponding LAT values. Therefore, in LAT extraction step 1214, the processor extracts (e.g., calculates) the corresponding LAT values, for example, based on the standard annotation bipolar EGM provided in Figure 6. Note that any annotations falling outside the WOI are ignored.

[0096] Finally, at LAT mapping generation step 1216, processor 28 uses a set of LAT values ​​to generate LAT mappings of at least a portion of the heart chambers.

[0097] Figure 4 The exemplary flowchart shown is chosen solely for clarity of concept. This embodiment may also include additional steps of the algorithm, such as receiving multiple bipolar and unipolar ECM signals, and receiving an indication of the degree of physical contact between the electrodes and the tissue being diagnosed from a contact force sensor. This step, and other possible steps, have been intentionally omitted from this disclosure to provide a more simplified flowchart.

[0098] Figure 5 To illustrate one embodiment of the invention Figure 4 The flowchart illustrates the training method of the algorithm. According to the presented implementation, the algorithm begins at step 502 with random parameters.

[0099] The following procedure begins by feeding random or pseudo-random parameters into the algorithm. Additionally, in the initialization step 504, the sumOfTheLoss variable is set to zero.

[0100] Next, at training step 506, the parameters are... Used in conjunction with a database of real labels for the annotated EGM signals (e.g., verified LAT values) (5061). In step 5062, the processor calculates the loss for each signal using one of the loss functions described above, and calculates the loss and sumOfTheLoss.

[0101] In gradient descent calculation step 508, the processor calculates the gradient of the sum of losses relative to each parameter. By examining the gradients, each parameter is adjusted in the direction that will reduce the loss.

[0102] In one implementation, stochastic gradient descent can be used to perform training. During training, calculations can be performed frequently. and The values ​​of x and w are adjusted in the direction that will reduce the loss. In this notation, the value of x is related to the convolution kernel, and the value of w is related to the ML algorithm. The training iterations optimize all values ​​of x and w globally to produce the desired set of optimized parameters when the termination criterion (510) is met. At this point, the process ends (512). If the termination condition has not yet been met, the process returns to step 504. The process can also be terminated by another criterion, such as reaching a timeout.

[0103] Various methods, such as stochastic gradient descent, batch gradient descent, mini-batch gradient descent, gradient descent, and Newton's method (also known as the Newton-Raphson method), can be used to converge the parameters by using the gradient values.

[0104] The termination condition can be defined as the convergence of the sumOfTheLoss variable, the convergence of the parameters, reaching a sumOfTheLoss value below a certain threshold, timeout, or a combination thereof.

[0105] Analysis of activation values

[0106] As described above, in some embodiments, the disclosed technique detects a single activation (i.e., finding a LAT value for each input electrogram waveform). In other embodiments, the technique is used to detect fragmented ECG signals, regional abnormal ventricular activation (LAVA) indication signals, delayed potentials, dual potentials, and split potentials. Implementations designed to detect regular activations may select the LAT value with the highest probability. Implementations designed to detect dual potentials, split potentials, fragmented potentials, fragmented potentials, or any kind of multiple potentials may select all LAT values ​​with a probability higher than a certain threshold. Implementations designed to detect delayed potentials may use a variable threshold, such that the threshold is higher the later the LAT. Implementations for detecting early potentials may do the opposite.

[0107] Figures 6A to 6C The following are shown as embodiments of the invention. Figure 4 The algorithm determines the probability distribution of Local Activation Time (LAT) values ​​and selects one or more LAT values.

[0108] Figure 6AThe diagram shows the selection of the LAT value with the highest probability (602). This method finds activations, for example, according to rules.

[0109] Figure 6B The diagram shows all LAT values ​​with probabilities (604, 606) above a given threshold of 608. This method can be used to find, for example, dual potentials, split potentials, fragment potentials, shattered potentials, or any kind of multipotential. Dual potentials play a crucial role in the detection of gaps or block lines in ablation lines. Shattered potentials play a crucial role in various arrhythmias, including atypical flutter and VT, where isthmuses in return pathways or slow zones can exhibit shattered potentials.

[0110] Figure 6C A varying threshold 616 is shown, which is lower for late-arriving LAT values ​​614 and higher for early values ​​(610, 612). This method preferentially identifies potentials arriving late. Identifying late-arriving potentials plays a crucial role in ablation therapy for ischemic VT conditions. Late-arriving potentials can indicate slow conduction pathways in scar-related VT conditions and other cardiac conditions.

[0111] for Figures 6A to 6C For clarity, only LAT values ​​between 10 and 25 milliseconds are shown in all the figures above. In a real system, the ANN should compute the probability of all LAT values ​​within the window of interest. The width of the window of interest is typically analogous to the cycle length of the patient's arrhythmia. For example, if the patient is experiencing a 200 beats / minute arrhythmia, the cycle length is 300 milliseconds. In this case, the physician could limit the window of interest to, for example, -150 to +150 milliseconds around the reference annotation. In this specific example, our ML method will generate a probability for each LAT value between -150 and +150 milliseconds.

[0112] While the implementation described herein primarily relates to cardiac EP mapping systems, the methods and systems described herein can also be used in any medical application that requires detection of cardiac activation based on bipolar intracardiac EGM signals and unipolar intracardiac EGM signals, such as in intracardiac defibrillators (ICDs) and artificial pacemakers.

[0113] It should be understood that the above embodiments are cited by way of example, and the invention is not limited to the specific contents shown and described above. Rather, the scope of the invention includes combinations and sub-combinations of the various features described above, as well as variations and modifications thereof, which will be apparent to those skilled in the art upon reading the above description, and which are not disclosed in the prior art.

Claims

1. A computer-implemented method for at least one electrophoresis preprocessing step in training a machine learning (ML) algorithm, comprising: Multiple bipolar electrograms and corresponding unipolar electrograms of the patient are acquired, the bipolar electrograms and the unipolar electrograms include annotations, wherein one or more human reviewers have identified and marked windows of interest and one or more activation times within the windows of interest; Based on the bipolar electrogram and the unipolar electrogram, a real label dataset is generated, including bipolar signals from the bipolar electrogram and unipolar signals from the unipolar electrogram. as well as Using the bipolar signal and the unipolar signal, at least one electrograph preprocessing step of a machine learning (ML) algorithm is trained to detect the occurrence of activations in a given bipolar electrograph within the window of interest, including: One or more convolutions of the bipolar signal and the unipolar signal are performed using a set of convolution kernels, wherein the coefficients of each convolution kernel in the set of convolution kernels are specified; Perform pointwise multiplication between the bipolar signal filtered by the first convolution kernel in the set of convolution kernels and the corresponding unipolar signal filtered by the second convolution kernel in the set of convolution kernels to obtain the multiplied signal; and The multiplied signal is concatenated with the corresponding bipolar and unipolar signals filtered by the set of convolution kernels to obtain a concatenated signal.

2. The method according to claim 1, wherein, Each bipolar electrogram is obtained from a pair of electrodes placed at a location in the patient's heart, and a corresponding unipolar electrogram is obtained from one of the electrodes in the pair.

3. The method according to claim 1, wherein, Acquiring the bipolar electrogram and the corresponding unipolar electrogram includes acquiring multiple bipolar electrograms and corresponding unipolar electrograms from multiple electrode pairs of a multi-electrode conduit.

4. The method according to claim 1, further comprising: The given bipolar electrogram is presented to the user, with annotations marking the time of the detected activation.

5. The method according to claim 1, wherein, The ML algorithm further includes an artificial neural network (ANN), and the method further includes training the artificial neural network using the cascaded signals.

6. The method according to claim 5, further comprising: Receive the patient's bipolar electrogram and corresponding unipolar electrogram, as well as the window of interest for inference; as well as Using a trained ML algorithm, the occurrence of activations in the received bipolar electrogram is detected, and the detected activations are associated with a cardiac tissue location that is in contact with the electrode that acquired the corresponding unipolar electrogram.

7. A system for at least one electrophoresis preprocessing step in training a machine learning (ML) algorithm, comprising: One or more displays, the one or more displays being configured to present an annotated electrograph; as well as One or more processors, said one or more processors being configured to: Multiple bipolar electrograms and corresponding unipolar electrograms of the patient are acquired, the bipolar electrograms and the unipolar electrograms including annotations, wherein one or more human reviewers have used the one or more processors and corresponding one or more displays to identify and mark windows of interest and one or more activation times within the windows of interest; Based on the bipolar electrogram and the unipolar electrogram, a real label dataset is generated, including bipolar signals from the bipolar electrogram and unipolar signals from the unipolar electrogram. as well as Using the bipolar signal and the unipolar signal, at least one electrograph preprocessing step of a machine learning (ML) algorithm is trained to detect the occurrence of activations in a given bipolar electrograph within the window of interest, including: One or more convolutions of the bipolar signal and the unipolar signal are performed using a set of convolution kernels, wherein the coefficients of each convolution kernel in the set of convolution kernels are specified; Perform pointwise multiplication between the bipolar signal filtered by the first convolution kernel in the set of convolution kernels and the corresponding unipolar signal filtered by the second convolution kernel in the set of convolution kernels to obtain the multiplied signal; and The multiplied signal is concatenated with the corresponding bipolar and unipolar signals filtered by the set of convolution kernels to obtain a concatenated signal.

8. The system according to claim 7, wherein, Each bipolar electrogram is obtained from a pair of electrodes placed at a location in the patient's heart, and a corresponding unipolar electrogram is obtained from one of the electrodes in the pair.

9. The system according to claim 7, wherein, The one or more processors are further configured to acquire the bipolar electrograms and the corresponding unipolar electrograms by acquiring multiple bipolar electrograms and corresponding unipolar electrograms from multiple electrode pairs of the multi-electrode conduit.

10. The system according to claim 7, wherein, The one or more processors are further configured to present the given bipolar electrogram to a user on the one or more displays, the given bipolar electrogram having annotations marking the time of detected activation.

11. The system according to claim 7, wherein, The ML algorithm further includes an artificial neural network (ANN), and the one or more processors are further configured to train the artificial neural network using the cascaded signals.

12. The system according to claim 11, wherein, The one or more processors are further configured to: Receive the patient's bipolar electrophysiological record and corresponding unipolar electrophysiological record, along with a window of interest for inference; and Using a trained ML algorithm, the occurrence of activations in the received bipolar electrogram is detected, and the detected activations are associated with a cardiac tissue location that is in contact with the electrodes that acquire the corresponding unipolar electrogram.

13. A method for identifying activations in cardiac electrograms using a machine learning (ML) model having at least one electrogram preprocessing step and an artificial neural network (ANN), the method comprising: The at least one electrogram preprocessing step for training the ML model is performed using the method according to any one of claims 1-6; The artificial neural network is trained using the cascaded signals; as well as Using the ML model, activations in a bipolar electrogram are identified within the window of interest using the corresponding unipolar electrogram.

14. The method according to claim 13, wherein, The training step of the at least one electrogram preprocessing step includes: optimizing the coefficients of the convolution kernel that is convolved with the electrogram during preprocessing.

15. The method according to claim 13, wherein, Identifying the activation includes generating a probability for each of the multiple possible activation values ​​in the window of interest using the ML model.

16. The method according to claim 13, wherein, Identifying the activations includes selecting one of the following: (i) the activation with the highest probability, (ii) one or more activations with a probability higher than a given threshold, and (iii) one or more activations with a probability higher than a variable threshold.

17. A computer software product comprising a tangible, non-transitory, computer-readable medium therein storing program instructions that, when read by a processor, cause the processor to perform the method according to any one of claims 1-6 and 13-16.

Images