Detection of activation in electrocardiograms using preprocessing trained on a neural network of intracardiac electrocardiograms.

Abstract

To provide evaluation of electrical propagation in a heart using machine leaning (ML).SOLUTION: A method includes collecting a plurality of bipolar electrograms and respective unipolar electrograms of patients, the electrograms including annotations in which one or more human reviewers have identified and marked a window of interest and one or more pieces of activation time inside the window of interest. A ground truth data set is generated from the electrograms, for training at least one electrogram-preprocessing step of a machine learning (ML) algorithm. The ML algorithm is applied to the electrograms to at least train the at least one electrogram-preprocessing step, so as to detect occurrence of activation in a given bipolar electrogram within the window of interest.SELECTED DRAWING: Figure 1

Description

Technical Field

[0005]

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

Background Art

[0002] In patients suffering from arrhythmia, the impulses of electrical activity can follow pathological pathways within the heart tissue. In electrophysiological studies, one of the physician's goals is to understand the pathway followed by the impulses of electrical activity, for example, by inserting a multi-electrode catheter into the patient's heart and measuring intracardiac ECG (electrocardiogram) signals through the catheter. Typically, the catheter is passed through the blood vessels into the heart to measure the electrical activity within the endocardium. Alternatively, the catheter may be inserted through an incision in the chest to measure the electrical activity within the epicardium.

[0003] To measure electrical activity, some electrodes of the catheter are brought into contact with the heart tissue (either endocardium or epicardium) to obtain the potential difference between the electrode and a ground point such as the Wilson central terminal ground point. The potential difference between the catheter electrode and the ground point is called the monopolar signal of the electrode. The potential difference between two adjacent electrodes of the catheter is called the bipolar signal between the electrodes.

[0004] During electrophysiological studies, to understand the pathway followed by the impulses of electrical activity, the physician needs to know the time when the impulse passed under each electrode. However, it is not possible to know when the impulse of electrical activity passed under each electrode by looking at the potential difference between bipolar pairs.

[0005] As further described below, this problem can be overcome by simultaneously analyzing bipolar and unipolar signals, as described in U.S. Patent Application Publication No. 2018 / 0042504. However, interpreting unipolar signals is difficult, particularly due to noise on the unipolar signal. Unlike bipolar signals, unipolar signals may contain an excessive amount of remote noise, for example. This makes the analysis of unipolar signals difficult.

[0006] As described above, a computer-assisted analysis of intracardiac signals using both bipolar and unipolar signals to determine the timing of cardiac activity was previously proposed in U.S. Patent Application Publication 2018 / 0042504. This patent application describes a rules-based algorithm and method comprising receiving a bipolar signal from a pair of electrodes adjacent to the myocardium of a human subject and receiving a unipolar signal from one of the selected electrodes. The method further comprises calculating the local unipolar minimum derivative of the unipolar signal and the time of occurrence of the unipolar minimum derivative. The method further comprises calculating the bipolar derivative of the bipolar signal, determining the ratio of the bipolar derivative to the local unipolar minimum derivative, and identifying the time of occurrence as myocardial activation time if the ratio exceeds a preset threshold ratio value. [Overview of the Initiative] [Means for solving the problem]

[0007] One embodiment of the present invention, described below, provides a method for collecting multiple bipolar electrophoresis diagrams and their respective unipolar electrophoresis diagrams from a patient, wherein the electrophoresis diagrams include annotations in which one or more reviewers identify and mark regions of interest and one or more activation times within the region of interest. A ground truth dataset is generated from the electrophoresis diagrams for training at least one electrophoresis diagram preprocessing step of a machine learning (ML) algorithm. To train at least one electrophoresis diagram preprocessing step, the ML algorithm is applied to the electrophoresis diagrams to detect the occurrence of activation in a given bipolar electrophoresis diagram within the region of interest.

[0008] In some embodiments, each bipolar potential diagram is acquired from a pair of electrodes positioned within the patient's heart, and each unipolar potential diagram is acquired from one of the electrodes in the pair.

[0009] In some embodiments, collecting bipolar and unipolar electrophoresis diagrams involves collecting multiple bipolar and unipolar electrophoresis diagrams from multiple electrode pairs of a multi-electrode catheter.

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

[0011] In another embodiment, at least one preprocessing step of the potentiometers includes point-by-point multiplication between the dipolar potentiometers filtered by one of the convolution kernels and the respective unipolar potentiometers filtered by another of the convolution kernels, and applying the ML algorithm includes inputting the multiplied signal obtained from the point-by-point multiplication into the ML algorithm.

[0012] In some embodiments, the method further includes presenting the user with a given bipolar potential diagram having annotations marking the time of detected activation.

[0013] In some embodiments, applying an ML algorithm includes applying an artificial neural network (ANN).

[0014] In some embodiments, the method further includes receiving the patient's bipolar and unipolar electrophoresis, as well as the region of interest, for inference. Using a trained ML algorithm, the occurrence of activation in the bipolar electrophoresis is detected, and the activation is associated with the location of the cardiac tissue in contact with the electrode from which the respective unipolar electrophoresis is obtained.

[0015] According to another embodiment of the present invention, a system is further provided comprising one or more displays and one or more processors. The one or more displays are configured to display annotated electrographs. The one or more processors are configured to (a) collect a plurality of bipolar electrographs and each unipolar electrograph of a patient, wherein the electrographs include annotations, and within the annotations, one or more reviewers identify and mark regions of interest and one or more activation times within the region of interest using one or more processors and each of the one or more displays; (b) in at least one of the one or more processors generate a ground truth dataset from the electrographs for training at least one electrograph preprocessing step of a machine learning (ML) algorithm; and (c) apply the ML algorithm to the electrographs to train at least one electrograph preprocessing step, thereby detecting the occurrence of activation in a given bipolar electrograph within the region of interest.

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

[0017] In some embodiments, training at least one electromagnet preprocessing step includes optimizing the coefficients of the convolved convolution kernel by electromagnetism during preprocessing.

[0018] In some embodiments, identifying an activation involves the ML model generating a probability for each of several possible activation values ​​within the region of interest.

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

[0020] According to another embodiment of the present invention, a computer software product is further provided, which includes a tangible, non-temporary computer-readable medium in which program instructions are stored, and when the instructions are read by a processor, the processor... A machine learning (ML) algorithm, which includes at least one trainable electrograph preprocessing step, is applied to a ground truth dataset of annotated electrographs to train at least one electrograph preprocessing step of the ML algorithm, thereby detecting the occurrence of activation in a given bipolar electrograph within a region of interest, the ground truth dataset of annotated electrographs being generated by collecting multiple bipolar electrographs and their respective unipolar electrographs from a patient, the electrographs including annotations, in which one or more reviewers identify and mark the region of interest and one or more activation times within the region of interest. [Brief explanation of the drawing]

[0021] A more complete understanding of the disclosure can be obtained by reading the following detailed description of embodiments in conjunction with the drawings. [Figure 1] This is a schematic diagram of a catheter-based electrophysiological (EP) mapping system configured to detect activation time in an electrogram (EGM) according to an exemplary embodiment of the present invention. [Figure 2] This is a schematic diagram illustrating a workflow for training and implementing an algorithm for detecting activation time in an electromagnetism (EGM) according to an exemplary embodiment of the present invention. [Figure 3]A block diagram schematically illustrating an algorithm for detecting activation time in an electrogram (EGM) according to an exemplary embodiment of the present invention. [Figure 4] A flowchart schematically showing a method and algorithm for detecting activation time, annotating activation, and generating an activation map in an electrogram (EGM) according to an exemplary embodiment of the present invention. [Figure 5] A flowchart schematically showing a training method of the algorithm of FIG. 4 according to an exemplary embodiment of the present invention. [Figure 6A] A diagram showing the selection of one or more local activation time (LAT) values from the probability distribution of the LAT values determined by the algorithm of FIG. 4 according to some exemplary embodiments of the present invention. [Figure 6B] A diagram showing the selection of one or more local activation time (LAT) values from the probability distribution of the LAT values determined by the algorithm of FIG. 4 according to some exemplary embodiments of the present invention. [Figure 6C] A diagram showing the selection of one or more local activation time (LAT) values from the probability distribution of the LAT values determined by the algorithm of FIG. 4 according to some exemplary embodiments of the present invention. **Modes for Carrying Out the Invention**

[0022] General Overview Intracardiac electrophysiological (EP) mapping is a catheter-based method that is sometimes applied to characterize cardiac EP wave propagation abnormalities, such as those that cause arrhythmias. In a typical catheter-based procedure, the distal end of a catheter equipped with multiple sensing electrodes is inserted into the heart to collect a set of data points that includes (i) the measurement locations on the ventricular wall tissue and (ii) each EP signal from which the EP mapping system can generate an EP map. An example of an EP map useful for diagnosing arrhythmias is an EP timing diagram map called a local activation time (LAT) map of the region of the ventricular wall tissue.

[0023] To generate the LAT map, the processor may need to analyze intracardiac ECG signals collected at various points within the heart, hereinafter referred to as electrograms (EGMs), thereby identifying the activation (i.e., activation in the waveform of the EGM) in each signal, annotating the activation, and calculating the LAT value.

[0024] The annotation time represents the time when a cardiac impulse has passed through a specific cardiac tissue location measured by the catheter. Since different EGM signals are collected at different times, a gating measurement technique is required to align the signals collected at different times. In the field of cardiac electrophysiology, the "gate" used for gating measurement is called a reference annotation. For example, the R peak of the QRS signal of the body surface ECG, or the activation detected in the coronary sinus, the activation detected in the high right atrium, or a more advanced method based on the input of multiple signals can be used as the reference annotation. The LAT value is defined as the difference between the time of the mapping annotation and the reference annotation time.

[0025] Some embodiments of the present invention described below provide machine learning (ML) techniques and algorithms configured to train both an algorithm preprocessing module and an artificial neural network (ANN) model. After pairs of bipolar signals and their respective unipolar EGM signals are preprocessed by the trained preprocessing module, the processor analyzes the preprocessed signals to determine the precise timing of activation under each of the catheter electrodes.

[0026] To prepare for training, multiple electrocardiograms are collected from multiple sites (e.g., hospitals), and each electrocardiogram includes annotations in which one or more reviewers identify and mark activation times using one or more processors. Typically, the activation annotations are verified and / or corrected by a number of experts skilled in analyzing diagnostic electrocardiograms.

[0027] One or more experts (e.g., clinical application engineers or algorithm engineers) compile a ground truth dataset from annotated electrophoresis data to be used to train the provided ML algorithm.

[0028] During training and inference, two components of the algorithm (i.e., preprocessing and ANN) work together to identify activations in the potentiometer input. To this end, the processor trains at least one potentiometer preprocessing step to apply the ML algorithm to the potentiometer to detect the occurrence of activations in a given bipolar potentiometer.

[0029] In some embodiments, during either training or inference, the disclosed model searches for activations present in both the bipolar signal and the respective unipolar EGM signals. Using the unipolar EGM, the model determines the activation time on the bipolar EGM and the tissue location of the activation (e.g., under the electrode from which the unipolar EGM was acquired). When applied to a large number of electrographs, the disclosed algorithm can generate a reliable database for producing electrophysiologically accurate LAT maps for at least a portion of the ventricle.

[0030] To clearly define the algorithm, an ML model (e.g., an ANN model) is applied to unipolar and bipolar signals via a sliding window. The output of the ANN model is the probability associated with the LAT value, or the set of probabilities associated with all possible LAT values ​​within the region of interest.

[0031] Training is typically performed using a loss function, such as the cross-entropy function, where the loss is defined as zero (0) when the calculated LAT value is equal to the ground truth LAT value. The arguments of the loss function may include a set of differences between the LAT values ​​estimated by the algorithm and their respective validated LAT values ​​(ground truth). The LAT differences are clearly defined only within a region of interest, such as a region of interest (WOI) with an effective width of cardiac cycle length. Arguments outside the WOI are ignored.

[0032] The loss function is defined such that a zero LAT difference contributes to the zero loss relative to the total loss, while the contributing loss asymptotically approaches a maximum value (e.g., a normalized value of 1) if the difference between the calculated LAT value and the expected LAT value exceeds a predetermined value such as 5 milliseconds. For the EP signals analyzed in this application, a loss function width of 5 milliseconds (e.g., at loss = 1 / 2) was found to be suitable for sufficiently fast convergence of the ML algorithm with sufficient estimation accuracy (e.g., ±1 ms) in the presence of noise.

[0033] Typically, during training, the disclosed technique optimizes the algorithm's preprocessing module, thereby including appropriate high-pass, low-pass, band-pass, and / or band-erasing convolutional filters (i.e., kernels) so that the preprocessing module can flatten and smooth the EGM signal and highlight the desired features of the EGM signal, and thus accurately determine the presence of features indicating activation that are simultaneously present in both the bipolar and unipolar signals.

[0034] Searching for features that indicate such activation, which are simultaneously present in both bipolar and unipolar EGMs, is roughly equivalent to searching for such patterns in signal multiplication. Therefore, in one embodiment, a preprocessing module multiplies the bipolar and unipolar EGMs acquired at the same location point by point. The resulting multiplied signal is concatenated with the signal itself and input to the ANN. In addition to using the signal itself, the advantage of using the multiplied signal in the input layer of the ANN is that the multiplication emphasizes the temporal matching of the amplitudes of the respective unipolar signals in the bipolar signals, thereby improving the ANN model's ability to detect such amplitudes that indicate activation.

[0035] Convolutional filters are optimized through training to overcome artifacts such as low-frequency baseline fluctuations and high-frequency noise artifacts. Initially, the algorithm learns a sufficiently optimal convolutional filter (by training the neural network with a large number of signals). At the very beginning of the training iterations, the convolutional kernel is typically created with random numbers or other numbers such as "all 1s" or "all zeros". Nevertheless, through training iterations, the kernel's numerical values ​​and the ANN module as a whole are optimized.

[0036] When the algorithm applies a convolutional filter that is well optimized for both the bipolar signal and its respective unipolar signal, the effectiveness of training through convolutional multiplication increases, and the preprocessing and ANN parameters are further optimized.

[0037] In some embodiments, multiple convolutional filters can be applied in parallel to the same signal. All convolutional filters may be trained to extract different features from a bipolar or unipolar signal.

[0038] Through training, the kernel parameter vector

[0039]

number

[0040]

number

[0041] Many types of ML models are available for use, and in addition to the ANN model used herein as an example, those skilled in the art can select different ML models, including decision tree learning, support vector machines (SVMs), and Bayesian networks. Examples of ANN models include convolutional NNs (CNNs), recursive NNs (RNNs), long short-term memory (LSTMs), enforcement learning, autoencoders, and probabilistic neural networks (PNNs). Typically, one or more processors used (collectively referred to as "processors") are programmed with software containing specific algorithms that enable the processors to perform each of the processor-related processes and functions outlined above. Typically, training is performed using a computing system comprising multiple processors, such as a graphics processing unit (GPU) or a tensor processing unit (TPU). However, any of these processors may also be a central processing unit (CPU).

[0042] In some embodiments, the disclosed technique detects a single activation (i.e., finds one LAT value per input electrographic waveform). In other embodiments, the technique is used to detect fragmented ECG signals, signals indicating local abnormal ventricular activity (LAVA), delayed potentials, dual potentials, and isolated potentials. Embodiments designed to detect regular activations can select the LAT value with the highest probability. Embodiments designed to detect dual potentials, isolated potentials, fragmented potentials, fragmented potentials, or multiple potentials of any kind can select all LAT values ​​that have a probability above a given constant threshold. Embodiments designed to detect delayed potentials can use a variable threshold (e.g., decreasing) such that the threshold increases as the LAT is later. The reverse may be true for embodiments to detect early potentials.

[0043] After training with 800,000 atrial ECG signals, the inventors achieved a 95 percent success rate when analyzing 3.1 million electrophoresis data that were not included in the training (i.e., by performing inference with the trained algorithm).

[0044] The disclosed ANN-based EP signal analysis technology could enhance the value of diagnostic catheterization procedures by providing physicians with reliable diagnostic methods.

[0045] System Description Figure 1 is a schematic diagram of a catheter-based electrophysiological (EP) mapping system 21 configured to detect activation time in an electrogeographic map (EGM) according to an exemplary embodiment of the present invention. Figure 1 illustrates a physician 27 using an EP mapping catheter 29 to perform EP mapping on the heart 23 of a patient 25. The mapping catheter 29 includes one or more arms 20 at its distal end, each of which is connected 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 adjacent electrodes 22a and 22b.

[0046] To acquire unipolar signals from electrodes 22a and / or 22b, the Wilson central terminal (WCT) grounding point is formed from three body surface electrodes 24 (attached to the patient's skin). For example, the three body surface electrodes 24 may be coupled to the patient's chest. (For illustrative purposes, only one external electrode 24 is shown in Figure 1).

[0047] Bipolar signals are easier to interpret but less accurate. Each unipolar signal is more accurate but more susceptible to far-field noise. If the user focuses only on bipolar signals, the user cannot distinguish between activation on the first electrode and activation on the second electrode. If the user focuses only on unipolar signals, the user may obtain a large number of false positives. To find the most accurate activation time, the disclosed technique applies an ML algorithm to the search for features that simultaneously predict electrical activation in pairs of unipolar and bipolar signals, as described below.

[0048] During the mapping procedure, the positions of the electrodes 22 are tracked while they are inside the patient's heart 23. For this purpose, electrical signals are transmitted between the electrodes 22 and the body surface electrodes 24 (typically, three additional body surface electrodes may be coupled to the patient's back). Based on these signals, and given the known positions of the electrodes 24 on the patient's body, the processor 28 calculates the estimated position of each electrode 22 within the patient's heart. Such tracking may be performed, for example, using the 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.

[0049] The processor associates signals acquired by electrodes 22, such as a unipolar EGM or a bipolar EGM, with the location where these signals were acquired. The processor 28 receives the signals via the electrical interface 35 and uses the information contained in these signals to construct an EP map 31 and an EGM 40, which are then displayed on the display 26.

[0050] In the illustrated embodiment, the processor 28 calculates local activation time (LAT) values ​​and generates a LAT map using an algorithm that includes an ML algorithm as disclosed in Figure 3. The ML algorithm is trained to find LAT values ​​at specific tissue locations.

[0051] The training, as shown in Figure 3, is largely based on the following approach, which is likely to have been performed by a human expert reviewer: observing the bipolar EGM signals (V22a-V22b) and understanding the approximate timing of cardiac wavefront activation from their location; then searching for the activation pattern in each unipolar EGM, V22a, to find the precise timing of cardiac wavefront activation. As can be seen under the magnifying callout, the irregularity 101 (a very small feature) identified in the unipolar signal is located precisely at the activation time (102), and the LAT value is calculated as the time difference between time 102 and baseline time 103. By repeating the training process, the aforementioned ML algorithm is expected to mimic the decisions of a skilled reviewer by the following: • Reshape the preprocessing parameters of the bipolar signal so that they emphasize the activation features observed on the bipolar EGM. • Reshape the preprocessing parameters of the monopolar signal to remove baseline fluctuations and emphasize the activation features observed on the monopolar EGM. Interpreting these features and shaping the machine learning parameters to calculate the probability of activation per LAT value within the region of interest.

[0052] The signal shown is a simplified example for conceptual purposes only. In reality, signals can be noisier than what is shown in this diagram.

[0053] The processor 28 typically comprises a general-purpose computer along with software programmed to perform the functions described herein. The software can be downloaded to the computer in electronic form, for example, over a network, or alternatively or additionally, can be provided and / or stored on a non-temporary physical medium, such as magnetic memory, optical memory, or electronic memory. Specifically, the processor 28 executes dedicated algorithms, including those disclosed in Figures 3, 4, and 5, which enable the processor 28 to perform the disclosed steps, as will be further described below. In some cases, the processor 28 has enhanced computing power, for example, by using the aforementioned GPU or TPU.

[0054] The example shown in Figure 1 is selected purely for the purpose of clarifying the concept. Other types of electrophysiological sensing catheter geometry may be employed, such as the Lasso® catheter (manufactured by Biosense Webster, Inc., Irvine, California). In addition, a contact sensor may be fitted to the distal end of the mapping catheter 29 and transmit data indicating the physical quality of electrode contact with tissue. In one embodiment, measurements from one or more electrodes 22 may be discarded if they indicate poor physical contact quality, while measurements may be considered valid if the contact quality of other electrodes is indicated to be sufficient.

[0055] Training and implementation of trained ML models Figure 2 is a schematic diagram illustrating the workflow for training and implementing an algorithm (e.g., including an ANN and a trainable preprocessing model) for detecting activation times in electromagnetisms (EGMs) according to one embodiment of the present invention. As can be seen, the collection of the training set of EGMs is performed at numerous sites 202, where a physician collects a large number of correlated bipolar and unipolar EGMs and verifies that they are accurately annotated with activation times (e.g., annotation 102 in Figure 1). The physician can manually correct any incorrect annotations of activations. This set of verified annotations is used during an ANN training session at training station 204 to minimize the ANN loss function and determine the ANN parameters.

[0056]

number

[0057]

number

[0058] The resulting trained ANN LAT activation detection algorithm may be stored in a memory stick (206) and provided to the user of system 21 in Figure 1, or to the user of other related diagnostic systems.

[0059] The examples shown in Figure 2 are selected simply to illustrate the concepts. For example, the trained model may be transmitted via the web instead of using a memory stick. Training may be performed using distributed computing instead of a training station.

[0060] Preprocessing trained on a neural network of electrophoresis Figure 3 is a block diagram illustrating a schematic algorithm for detecting activation time in an electromagnetism (EGM) according to one embodiment of the present invention.

[0061] As can be seen, the disclosed deep learning model is divided into two parts: a preprocessing module 300 and an ANN module 301. The preprocessing module 300 optimizes the parameters

[0062]

number

[0063]

number

[0064] The raw EGM input to the model includes correlated portions of the raw unipolar and bipolar signals (93, 95), which are successively selected by applying sliding windows 303 and 305, respectively.

[0065] During training, intermediately optimized parameters

[0066]

number

[0067]

number

[0068]

number

[0069] In the illustrated embodiment, the parameter

[0070]

number

[0071] As described above, a central feature of the disclosed technology (e.g., algorithm) is the training of a preprocessing step by optimizing the convolution kernel applied to the time-windowed (303, 305) EGM signals (93, 95), i.e., the unipolar signal 302 (e.g., V22a in Figure 1) and the bipolar signal 304 (e.g., V22a-V22b in Figure 1).

[0072] The preprocessing module 300 may include any number of convolutional kernels, typically 4 to 8, to increase training flexibility in finding the optimized kernel. Each convolutional kernel can be "learned" to emphasize different features of the signal. The preprocessing stage and the ML algorithm are optimized through iterative training to identify which features need to be emphasized. For example, it is reasonable to assume that a convolutional kernel associated with a unipolar signal learns to remove baseline fluctuations, but this should not be defined by the algorithm designer and should occur as a natural result of training.

[0073] Generally, a convolutional filter acts as a finite impulse response (FIR) filter. Convolution may function as a low-pass, high-pass, band-canceling, band-canceling, or band-pass filter type.

[0074] As mentioned above, optimized parameters such as the type and sharpness of the optimized kernel

[0075]

number

[0076] A key feature of the preprocessing module 300 is that it accepts a multiplying signal as input. As can be seen, the convolved unipolar and bipolar signals are multiplied point by point (313), and the multiplying signal 306 is input along with other convolved signals. As mentioned above, any number of convolution kernels "A" may be used, and the same number of multiplying signals with different results are input to the ANN module 301 in the form of a concatenated input vector.

[0077] The motivation for using multiplying signals is to utilize the time correlation between unipolar and bipolar signals for training modules 300 and 301, as illustrated by signals V22a-V22b and V22a in Figure 1.

[0078] During inference, the same deep learning model is used, and the optimized parameters are applied.

[0079]

number

[0080] Some embodiments of the invention described herein receive a certain number of signal samples as input (e.g., receive digitized signals of a given duration), but the physician is typically interested in a specific region of interest that depends on the period length of the patient's current arrhythmia. To overcome this problem, in both the training and inference phases, the probabilities of LAT values ​​outside the region of interest are set to zero immediately before normalizing the distribution, for example, using a Softmax function.

[0081] By setting all probability values ​​outside the region of interest to zero, we ensure that the sum of probabilities is still 1 after the normalization process, and that the region of interest is precisely cut out, taking into account both the training and inference stages.

[0082] The following pseudocode example represents a convolution preprocessing and signal multiplication process having k convolutions of a unipolar signal and k convolutions of a bipolar signal, where each bipolar signal p ≤ k is multiplied point by point with its respective unipolar signal to generate p multiplied signals. In this notation,

[0083]

number

[0084]

number

[0085] The example block diagram shown in Figure 3 is selected simply for the purpose of illustrating the concept. ANN module 301 is a schematic module shown conceptually only, and represents a selection from a number of possible ANN embodiments, including library-coded ANN functions.

[0086] A method for detecting activation in EGM using a preprocessing module trained with ANN. Figure 4 is a flowchart illustrating a method and algorithm according to one embodiment of the present invention for detecting activations in an electromagnetism (EGM), annotating activations, and generating an activation map. The algorithm according to the presented embodiment is divided into two parts: preparation of the algorithm and training dataset 1101 and use of the algorithm 1102.

[0087] Algorithm preparation begins in ANN modeling step 1201, which involves generating an ANN algorithm for activation detection in EGM, and the algorithm includes training of preprocessing of correlated unipolar and bipolar signals, including an ANN trainable preprocessing module 300, such as the algorithm shown in Figure 3.

[0088] Independently, the preparation phase of the training database set begins with a ground truth acquisition step 1202, which involves collecting electrocardiograms from multiple sites (e.g., hospitals) that include activation annotations verified and / or corrected by a number of specialists skilled in analyzing diagnostic electrocardiograms.

[0089] In the ground truth dataset generation step 1203, a person skilled in the art (e.g., a clinical application engineer or an algorithm engineer) compiles a ground truth dataset from the collection of annotated electrographs to be used to train the algorithm prepared in step 1201.

[0090] Next, using the ground truth dataset of annotated electrophoresis maps, whose annotation accuracy was verified in step 1202, the processor trains the algorithm (e.g., the ANN and preprocessor) in ML algorithm training step 1204.

[0091] Algorithm preparation is completed in the trained model storage step 1206 by storing the trained model on a non-temporary computer-readable medium such as a disk-on key (memory stick). In an alternative embodiment, the model is pre-transmitted and its optimized parameters

[0092]

number

[0093] Algorithm usage 1102 executes the process that starts in algorithm upload step 1208, during which the user can access the entire ML model or (for example,

[0094]

number

[0095] Next, using the trained ANN model, the processor generates a set of probability distributions for activation times (typically one distribution per bipolar EGM) in the EGM inference process 1212.

[0096] In some embodiments, the peaks of the distribution may be selected in subsequent steps, i.e., beyond the steps included in the ANN model, to determine the annotations and their respective LAT values. Thus, in the LAT extraction step 1214, the processor annotates the bipolar EGM and extracts (e.g., calculates) the corresponding LAT values ​​according to the criteria provided, for example, in Figure 6. Note that any annotations outside the WOI are ignored.

[0097] Finally, using the set of LAT values, the processor 28 generates an LAT map of at least a portion of the ventricle in the LAT map generation step 1216.

[0098] The exemplary flowchart shown in Figure 4 has been selected purely for the purpose of clarifying the concept. This embodiment may also include additional steps in the algorithm, such as receiving multiple bipolar and unipolar ECM signals, and receiving an indicator of the degree of physical contact between the electrode and the tissue being diagnosed from a contact force sensor. These steps and other possible steps have been intentionally omitted from the disclosure herein in order to provide a more simplified flowchart.

[0099] Figure 5 is a flowchart illustrating a training method for the algorithm shown in Figure 4 according to one embodiment of the present invention. The algorithm according to the presented embodiment starts with random parameters, and in step 502, random or pseudo-random parameters are added to the algorithm.

[0100]

number

[0101] Next, in training process 506, parameters

[0102]

number

[0103] In the gradient descent calculation step 508, the processor calculates the gradient of the sum of losses for each parameter. By examining the gradient, each parameter is adjusted in the direction of reducing the losses.

[0104] In one embodiment, training may be performed using stochastic gradient descent. During training,

[0105]

number

[0106]

number

[0107] Parameters can be converged by using gradient values ​​through various methods such as stochastic gradient descent, batch gradient descent, small batch gradient descent, gradient descent, and Newton's method, also known as the Newton-Raphson method.

[0108] The termination conditions can be defined as convergence of the sumOfTheLoss variable, convergence of parameters, achievement of a sumOfTheLoss value below a certain threshold, timeout, or a combination thereof.

[0109] Analysis of activation values As described above, in some embodiments, the disclosed technique detects a single activation (i.e., finds one LAT value per input electrographic waveform). In other embodiments, the technique is used to detect fragmented ECG signals, signals indicating localized abnormal ventricular activity (LAVA), delayed potentials, dual potentials, and isolated potentials. Embodiments designed to detect regular activations can select the LAT value with the highest probability. Embodiments designed to detect dual potentials, isolated potentials, fragmented potentials, fragmented potentials, or multiple potentials of any kind can select all LAT values ​​that have a probability above some threshold. Embodiments designed to detect delayed potentials can use a variable threshold such that the threshold increases as the LAT is later. The reverse may be true for embodiments to detect early potentials.

[0110] Figures 6A to 6C show the selection of one or more local activation time (LAT) values ​​from a probability distribution of LAT values ​​determined by the algorithm in Figure 4, according to several embodiments of the present invention.

[0111] Figure 6A shows the selection of the LAT value (602) with the highest probability. This method can be used, for example, to find regular activations.

[0112] Figure 6B shows the selection of all LAT values ​​(604, 606) that have a probability of exceeding a given threshold of 608. This method can be used, for example, to find dual potentials, separated potentials, subdivided potentials, fragmented potentials, or any kind of multiple potentials. Dual potentials play an important role in detecting block lines or gaps within the ablation line. Subdivided potentials play an important role in various arrhythmias, including atypical flutter and VT cases, where isthmuses in reentry circuits or slow zones may exhibit subdivided potentials.

[0113] Figure 6C shows the threshold 616, which changes so that the threshold is lower than the later LAT value 614 and higher than the initial values ​​(610, 612). This method prioritizes finding delayed potentials. Finding delayed potentials plays an important role in ablation procedures for ischemic VT cases. Delayed potentials may be an indicator of slow conduction pathways in reentry circuits in scar-associated VT cases and other cardiac pathologies.

[0114] To clarify Figures 6A to 6C, only LAT values ​​between 10 msec and 25 msec are shown in all the above figures. In a real system, ANN should calculate the probability of all LAT values ​​within the region of interest. The width of the region of interest is typically similar to the period length of the patient's arrhythmia. For example, if the patient has an arrhythmia of 200 beats per minute, the period length is 300 milliseconds. In this case, the physician may define the region of interest as, for example, -150 to +150 milliseconds centered on the reference annotation. In this particular embodiment, our ML method generates the probability of all LAT values ​​between -150 and +150 milliseconds.

[0115] While the embodiments described herein primarily address cardiac EP mapping systems, the methods and systems described herein can also be used in any medical application where cardiac activation needs to be detected based on bipolar and monopolar intracardiac EGM signals, such as intracardiac defibrillators (ICDs) and artificial heart pacemakers.

[0116] The embodiments described above are illustrative examples, and it should be understood that the present invention is not limited to those specifically illustrated and described above. Rather, the scope of the present invention includes both combinations and partial combinations of the various features described above, as well as variations and modifications thereof that a person skilled in the art would likely conceive upon reading the foregoing description, and which are not disclosed in the prior art.

[0117] [Implementation Method] (1) A method, The collection of multiple bipolar electrophoresis diagrams and their respective unipolar electrophoresis diagrams from a patient, wherein the electrophoresis diagrams include annotations, and within the annotations, one or more reviewers identify and mark areas of interest (windows of interest) and one or more activation times within those areas of interest. From the aforementioned electromagnetism, a ground truth dataset is generated for training at least one electromagnetism preprocessing step of a machine learning (ML) algorithm, To train at least one of the above-mentioned potential diagram preprocessing steps, the ML algorithm is applied to the potential diagram to detect the occurrence of activation in a given bipolar potential diagram within the region of interest, including, method. (2) The method of step 1, wherein each bipolar potential diagram is obtained from a pair of electrodes positioned within the patient's heart, and each unipolar potential diagram is obtained from one of the electrodes of the pair. (3) The method according to Embodiment 1, wherein collecting the bipolar potential diagrams and the respective unipolar potential diagrams includes collecting multiple bipolar potential diagrams and the respective unipolar potential diagrams from multiple electrode pairs of a multi-electrode catheter. (4) The method according to Embodiment 1, wherein the at least one potentiometer preprocessing step includes performing one or more convolutions of the potentiometer using a set of convolution kernels, and training the at least one potentiometer preprocessing step includes specifying the coefficients of the convolution kernels. (5) The method according to Embodiment 4, wherein the at least one preprocessing step of the potential diagram includes point-by-point multiplication between the bipolar potential diagram filtered by one of the convolution kernels and each unipolar potential diagram filtered by another of the convolution kernels, and the application of the ML algorithm includes inputting the multiplication signal obtained from the point-by-point multiplication to the ML algorithm.

[0118] (6) The method according to Embodiment 1, comprising presenting to the user a given bipolar potential diagram having annotations marking the time of detected activation. (7) The method according to Embodiment 1, wherein applying the ML algorithm includes applying an artificial neural network (ANN). (8) To receive the patient's bipolar potential diagram and its corresponding unipolar potential diagram, as well as the region of interest, Using the trained ML algorithm, the occurrence of activation in the bipolar potential diagram is detected, and the activation is associated with the location of cardiac tissue in contact with the electrode from which the respective unipolar potential diagrams are obtained. including, The method according to Embodiment 1. (9) A system, One or more displays configured to display annotated electrophoresis diagrams, One or more processors, The collection of multiple bipolar electrophoresis diagrams and their respective unipolar electrophoresis diagrams of a patient, wherein the electrophoresis diagrams include annotations, and within the annotations, one or more reviewers identify and mark regions of interest and one or more activation times within those regions of interest using one or more processors and each of the one or more displays. In at least one of the one or more processors, a ground truth dataset is generated from the electromagnet for training at least one electromagnet preprocessing step of a machine learning (ML) algorithm. To train at least one of the above-mentioned potential diagram preprocessing steps, the ML algorithm is applied to the potential diagram to detect the occurrence of activation in a given bipolar potential diagram within the region of interest, Configured to perform, One or more processors, Equipped with, system. (10) The system according to step 9, wherein each bipolar potential diagram is obtained from a pair of electrodes positioned within the patient's heart, and each unipolar potential diagram is obtained from one of the electrodes of the pair.

[0119] (11) The system according to Embodiment 9, wherein one or more processors are configured to collect the bipolar electrophoresis diagrams and the respective unipolar electrophoresis diagrams, the system comprising collecting multiple bipolar electrophoresis diagrams and the respective unipolar electrophoresis diagrams from multiple electrode pairs of a multi-electrode catheter. (12) The system according to Embodiment 9, wherein the at least one potential diagram preprocessing step includes performing one or more convolutions of the potential diagram using a set of convolution kernels, and the one or more processors are configured to train a model by optimizing the coefficients of the convolution kernels. (13) The system according to Embodiment 12, wherein the at least one preprocessing step of the potential diagram includes point-by-point multiplication between a bipolar potential diagram filtered by one of the convolution kernels and each unipolar potential diagram filtered by another of the convolution kernels, and the one or more processors are configured to apply the ML algorithm by inputting the multiplication signal obtained from the point-by-point multiplication to the ML algorithm. (14) The system according to Embodiment 9, wherein one or more processors are further configured to present to the user a given bipolar potential diagram having annotations marking the detected activation times on one or more displays. (15) The system according to embodiment 9, wherein one or more processors are configured to apply the ML algorithm by applying an artificial neural network (ANN).

[0120] (16) The one or more processors For inference purposes, the patient's bipolar potential diagram and its corresponding unipolar potential diagram, as well as the region of interest, are received. Using the trained ML algorithm, the occurrence of activation is detected in the bipolar potential diagram, and the activation is associated with the location of cardiac tissue in contact with the electrode from which the respective unipolar potential diagrams are obtained. It is further structured in such a way. The system described in Embodiment 9. (17) A method for identifying activation in an electrocardiogram using a machine learning (ML) model having at least one electrocardiogram preprocessing step, Training the ML model on at least one potential diagram preprocessing step, Using the aforementioned ML model, activation is identified in the bipolar potential diagram using the respective unipolar potential diagrams within the region of interest. including, method. (18) The method according to Embodiment 17, wherein training the at least one potentiometer preprocessing step includes optimizing the coefficients of the convolved convolution kernel by the potentiometer during preprocessing. (19) The method of Embodiment 17, wherein identifying the activations includes generating a probability for each of the multiple possible activation values ​​within the region of interest using the ML model. (20) The method of Embodiment 17, wherein identifying the activations includes selecting one of (i) the activation having the highest probability, (ii) one or more activations having a probability above a given threshold, and (iii) one or more activations having a probability above a variable threshold.

[0121] (21) A computer software product comprising a tangible, non-temporary computer-readable medium in which program instructions are stored, wherein when the instructions are read by a processor, A computer software product wherein the processor applies a machine learning (ML) algorithm, which includes at least one trainable electrograph preprocessing step, to a ground truth dataset of annotated electrographs in order to train the at least one electrograph preprocessing step of the ML algorithm, thereby detecting the occurrence of activation in a given bipolar electrograph within a region of interest, the ground truth dataset of annotated electrographs being generated by collecting multiple bipolar electrographs and their respective unipolar electrographs from a patient, the electrographs including annotations, and within the annotations, one or more reviewers have identified and marked the region of interest and one or more activation times within the region of interest.

Claims

1. It is a method, The collection of multiple bipolar electrophoresis diagrams and their respective unipolar electrophoresis diagrams from a patient, wherein the electrophoresis diagrams include annotations, and within the annotations, one or more reviewers identify and mark regions of interest and one or more activation times within those regions of interest. From the aforementioned electromagnetism, a ground truth dataset is generated for training at least one electromagnetism preprocessing step of a machine learning (ML) algorithm, To train at least one of the above-mentioned potential diagram preprocessing steps, the ML algorithm is applied to the potential diagram to detect the occurrence of activation in a given bipolar potential diagram within the region of interest, Includes, The at least one potentiometer preprocessing step includes performing one or more convolutions of the potentiometer using a set of convolution kernels, and training the at least one potentiometer preprocessing step includes specifying the coefficients of the convolution kernels. The at least one preprocessing step of the potential diagram includes point-by-point multiplication between the bipolar potential diagram filtered by one of the convolution kernels and each unipolar potential diagram filtered by another of the convolution kernels, and the application of the ML algorithm includes inputting the multiplication signal obtained from the point-by-point multiplication into the ML algorithm. method.

2. The method according to claim 1, wherein each bipolar potential diagram is obtained from a pair of electrodes positioned within the patient's heart, and each unipolar potential diagram is obtained from one of the electrodes of the pair.

3. The method according to claim 1, wherein collecting the bipolar potential diagrams and the respective unipolar potential diagrams includes collecting multiple bipolar potential diagrams and the respective unipolar potential diagrams from multiple electrode pairs of a multi-electrode catheter.

4. The method according to claim 1, comprising presenting to the user a given bipolar potential diagram having annotations marking the time of detected activation.

5. The method according to claim 1, wherein applying the ML algorithm includes applying an artificial neural network (ANN).

6. For inference purposes, the patient's bipolar potential diagram and its corresponding unipolar potential diagram, as well as the region of interest, Using the trained ML algorithm, the occurrence of activation in the bipolar potential diagram is detected, and the activation is associated with the location of cardiac tissue in contact with the electrode that acquires the respective unipolar potential diagrams. including, The method according to claim 1.

7. It is a system, One or more displays configured to display annotated electrophoresis diagrams, One or more processors, The collection of multiple bipolar electrophoresis diagrams and their respective unipolar electrophoresis diagrams of a patient, wherein the electrophoresis diagrams include annotations, and within the annotations, one or more reviewers identify and mark regions of interest and one or more activation times within those regions of interest using one or more processors and one or more displays. In at least one of the one or more processors, a ground truth dataset is generated from the electromagnet for training at least one electromagnet preprocessing step of a machine learning (ML) algorithm. To train at least one of the above-mentioned potential diagram preprocessing steps, the ML algorithm is applied to the potential diagram to detect the occurrence of activation in a given bipolar potential diagram within the region of interest, Configured to perform, One or more processors, Equipped with, The at least one potential diagram preprocessing step includes performing one or more convolutions of the potential diagram using a set of convolution kernels, and the one or more processors are configured to train a model by optimizing the coefficients of the convolution kernels. The at least one preprocessing step of the potential diagram includes point-by-point multiplication between the bipolar potential diagram filtered by one of the convolution kernels and each unipolar potential diagram filtered by another of the convolution kernels, and the one or more processors are configured to apply the ML algorithm by inputting the multiplication signal obtained from the point-by-point multiplication to the ML algorithm. system.

8. The system according to claim 7, wherein each bipolar potential diagram is obtained from a pair of electrodes positioned within the patient's heart, and each unipolar potential diagram is obtained from one of the electrodes of the pair.

9. The system according to claim 7, wherein the one or more processors are configured to collect the bipolar potential diagrams and the respective unipolar potential diagrams, which includes collecting multiple bipolar potential diagrams and the respective unipolar potential diagrams from multiple electrode pairs of a multi-electrode catheter.

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

11. The system according to claim 7, wherein one or more processors are configured to apply the ML algorithm by applying an artificial neural network (ANN).

12. The one or more processors described above For inference purposes, the patient's bipolar potential diagram and its corresponding unipolar potential diagram, as well as the region of interest, are received. Using the trained ML algorithm, the occurrence of activation is detected in the bipolar potential diagram, and the activation is associated with the location of cardiac tissue in contact with the electrode that acquires the respective unipolar potential diagrams. It is further structured in such a way. The system according to claim 7.

Images