Computer device for real-time analysis of electromotive force diagrams

Abstract

A computer device for analyzing electrograms in real time comprises a memory arranged to receive real-time electrogram signals respectively originating from a plurality of electrodes, an extractor and a first evaluator including a gradient boosting based machine learning module. The extractor is arranged to extract a set of features from each electrogram signal in a set of electrogram signals, the set of features including at least one temporal analysis feature and at least one morphological feature, and to provide the resulting set of features to the gradient boosting based machine learning module. The machine learning module is trained on data including a set of features labeled with a value indicating whether an associated electrogram signal exhibits variance and is arranged to provide an output for each set of electrogram signals.

Description

Technical Field

[0001] The present invention relates to a computer device for real-time analysis of potential diagrams.

Background Art

[0002] The field of atrial fibrillation has rapidly developed in the past five years.

[0003] Detection of the heart site that promotes atrial fibrillation includes that by an offline computing device using software such as Topera or Cardiolnsight, and that by a real-time computing device using software such as CARTO.

[0004] Topera software aims to reconstruct the electrical activation of the atrium of the heart. Signal acquisition is performed using a "basket catheter" (a catheter deployed over the entire atrium). Analysis is performed with a delay time of more than 2 minutes from the start of acquisition. This type of catheter is difficult to insert and electrode contact is not guaranteed, so this solution is inconvenient. The analysis time is very long, a high-definition map cannot be obtained in reconstruction, and in the case of complex electrical activation (accounting for 70% of atrial fibrillation), it may fail due to overly simple reconstruction.

[0005] Cardiolnsight software aims to reconstruct the electrical activity of the heart using a multi-electrode electrocardiogram measurement vest worn on the patient's skin. The patient wears the vest before surgery, and the data is analyzed and accessible during surgery. Analysis is performed with a delay time (more than 15 minutes) after the start of acquisition. This solution has the drawback of very long computation time, causing a delay time and making setup difficult (the patient needs to come several days earlier and wear the vest so that data can be extracted before surgery), and it is quite costly. Also, in reconstruction, a high-definition map cannot be obtained because the measurement setup is too far away, and in the case of complex electrical activation, it may fail due to overly simple reconstruction.

[0006] CARTO software by Biosense & Webster can implement algorithms to detect cardiac arrhythmias.

[0007] The CFAE (Complex Scattered Atrial Potential) algorithm calculates the number of deviation points (changes in the sign of the derivative) in the signal and creates a color map in real time. Physicians interpret these complex maps to determine the area of ​​interest. This algorithm is not particularly specialized and is relatively simple.

[0008] Ensite software from Abbott Laboratories can implement algorithms for detecting cardiac arrhythmias.

[0009] The Ripple algorithm (Ripple mapping) is a module in the CARTO software that can reproduce the electrical activity of the atria after the initial catheter passage. This module visualizes both the amplitude and propagation of electrical waves. As a result, it creates a map that is very complex for physicians to understand. While this may be effective in simple cases, in cases of complex electrical activation, the interpretation of the map becomes too difficult, and the analysis fails. [Overview of the project] [Problems that the invention aims to solve]

[0010] The applicant has filed French Republic Patent Application No. 1852850, describing an improved system that uses two interconnected methods to detect cardiac regions that promote atrial fibrillation in real time. More specifically, the invention describes the use of a classifier arranged to apply two classification models to electrophoretic data. One model is applied faster and less accurately than the other, and this is used to alert a physician to potential areas of concern in the cardiac region. [Means for solving the problem]

[0011] Further research by the applicant has made it possible to determine an improved computer device for real-time detection of cardiac regions that promote atrial fibrillation. The device comprises a memory configured to receive real-time electromorphic signals emitted from multiple electrodes, and a first evaluator comprising an extractor and a gradient boosting-based machine learning module. The extractor is configured to extract a set of features from each electromorphic signal in a set of electromorphic signals, including at least one temporal analysis feature and at least one morphological feature, and to feed the resulting set of features to the gradient boosting-based machine learning module. The machine learning module is trained on data including a set of features labeled with a value indicating whether the relevant electromorphic signal exhibits variance, and is configured to output, for each set of electromorphic signals, a first probability array indicating whether each electromorphic signal in the set exhibits variance. The device also comprises a second evaluator, which comprises a convolutional neural network. The convolutional neural network receives a set of real-time electromorphic signals and outputs a second probability array indicating that the input real-time electromorphic signals exhibit variance. The convolutional neural network is trained on a database of electromorphic signals, each labeled with a value indicating whether or not each electromorphic signal exhibits variance. The device also includes a predictor, which, based on a first and second probability array determined for a given set of electromorphic signals, returns a third probability array that is at least partially based on a weighted average of the values ​​in the first and second probability arrays.

[0012] What makes this computer device particularly interesting is that the applicant was able to construct patient distribution maps in real time and use these maps to achieve an 88% termination rate for patients' atrial fibrillation. This means that the applicant's invention can be reliably used to assist physicians in real time in constructing distribution maps and treating atrial fibrillation.

[0013] In various embodiments, the computer device may include one or more of the following features: The extractor is configured to extract at least one temporal analysis feature from a group consisting of a first period length estimate, a second period length estimate, and the frequency in the fast Fourier transform of the potentiometer signal having the maximum amplitude. The extractor is configured to extract at least one morphological feature from a group consisting of the Euclidean norm of the electrographic signal and the integrated absolute derivative of the electrographic signal. Furthermore, the device is configured to divide the real-time electromagnetism signal into a series of electromagnetism signals having a selected duration. Furthermore, the device is arranged to provide a set of potentiometric signals having the same selected duration. Furthermore, the device is configured to divide the real-time electromagnetism signal into a series of electromagnetism signals having a selected duration. Furthermore, the predictor is configured to use the probability in the first probability array in the third probability array when the absolute value of the difference between the probability in the first probability array and the probability in the second probability array corresponding to the same potential diagram signal exceeds a threshold. Furthermore, the device further comprises a display arranged to determine a color associated with a value in a third array, and for each electrode, a display arranged to output a color associated with the probability determined for the corresponding electromorphic signal in the third probability array.

[0014] The present invention also relates to a computer program including instructions for implementing a first evaluator, a second evaluator, and a predictor; a data storage medium on which this computer program is recorded; and a computer-implemented method for receiving a real-time potential diagram signal, executing the first evaluator, the second evaluator, and the predictor according to the present invention, and returning a third probability array. [Brief explanation of the drawing]

[0015] Other features and advantages of the present invention will be described below with reference to the accompanying drawings illustrating exemplary embodiments of the present invention. [Figure 1] This is an overall diagram relating to one embodiment of a computer device according to the present invention. [Figure 2] This is a schematic diagram of the convolutional neural network used in the embodiment shown in Figure 1. [Modes for carrying out the invention]

[0016] The accompanying drawings and the following description primarily contain effective and well-defined features of the present invention. Therefore, they can be used not only to better understand the present invention but also to contribute appropriately to its definition.

[0017] Persistent atrial fibrillation (SFIVA) remains a therapeutic challenge. While the number of patients eligible for catheter ablation is increasing, the optimal ablation strategy for SFIVA remains unclear. STAR-AF2 (Verma A, Jiang Cy, Betts TR et al., "Approaches to catheter ablation for persistent atrial fibrillation," New England Journal of Medicine, 2015;372:1812-1822) concluded that additional ablation procedures beyond pulmonary vein isolation (PVI) do not improve long-term outcomes.

[0018] On the other hand, several observational studies (see Jadidi AS, Lehrmann H, Keyl C, et al., "Ablation of persistent atrial fibrillation targeting low-voltage areas with selective activation characteristics," Circulation: Arrhythmia and Electrophysiology, 2016;9:e002962, and Seitz J, Bars C, Theodore G, et al., "AF ablation guided by spatiotemporal electrogram dispersion without pulmonary vein isolation: a wholly patient-tailored approach," Journal of the American College of Cardiology, 2017;69:303-321) suggest that non-PVI lesions targeting areas with electrographic abnormalities are beneficial for patients with persistent atrial fibrillation.

[0019] In particular, the latter study showed results supporting the application of high-frequency energy to the atrial region, which exhibits dispersion in multipolar electrophoresis. Another study (Narayan SM, Krummen DE, Shivkumar K, Clopton P, Rappel WJ, Miller JM, "Treatment of atrial fibrillation by the ablation of localized sources: CONFIRM (Conventional Ablation for Atrial Fibrillation With or Without Focal Impulse and Rotor Modulation) trial," Journal of the American College of Cardiology 2012;60:628-636) instead employed advanced signal analysis and provided a mechanism-based representation for inducing drivers of atrial fibrillation.

[0020] Among facilities attempting to implement a mapping approach based on similar electrocardiograms, regardless of electrocardiogram abnormalities or analysis, there are significant differences in the level of experience and the learning curve stage.

[0021] FIG. 1 is an overall view of one embodiment of a computer device according to the present invention.

[0022] The computer device 2 includes a memory 4, a first evaluator 6 including an extractor 8 and a machine learning module 10 based on gradient boosting, a second evaluator 12, and a predictor 14.

[0023] From the perspective of use, an operator uses electrodes inserted into a patient's heart. The electrodes are indicated by reference numeral 16 and are displayed on the display of the computer device 2. As shown in FIG. 1, 16 various types of electrodes may be provided.

[0024] The sensing results by the electrodes are a set of real-time electrocardiogram signals 18 supplied to the memory 4. These electrocardiogram signals are analyzed, and the predictor 14 returns a corrected map 20 of the electrodes. Thereby, a doctor can identify which electrodes sensed regions potentially related to dispersion. In return, the doctor can mark (or automatically return relevant positions) the corresponding positions on the patient's heart. Thereby, a map 22 of the heart having a region to be treated from the perspective of atrial fibrillation indicated by dots is obtained.

[0025] Memory 4 stores real-time electrophysical crystal signals (for example, after they have been digitized and divided at regular time intervals so that all electrophysical crystal signals have the same duration, e.g., 3 seconds) and a database of more than 275,000 electrophysical crystal signals labeled with a value indicating whether they show dispersion, i.e., whether they are related to atrial fibrillation. Electrophysical crystal signals having the same duration may be consecutive or overlapping. Memory 4 may also store any temporary data that may be generated in the process of carrying out the present invention. Memory 4 may also store maps obtained as a result of the operation of the predictor 14, optionally combined with annotations and catheter shapes created by a physician.

[0026] In the embodiments described herein, memory 4 may be implemented in an appropriate manner, such as a hard disk drive, a solid-state drive, flash memory, memory integrated into a processor, or cloud-accessible remote storage.

[0027] In the embodiments described herein, the first evaluator 6, the second evaluator 12, and the predictor 14, which include the extractor 8 and the machine learning module 10 based on gradient boosting, are computer programs that run on one or more processors. Such processors include any means known to perform automated computation, such as a CPU, GPU, CPU and / or GPU grid, telecomputing grid, specially configured FPGA, specially configured ASIC, dedicated chip such as SOC or NOC, or AI dedicated chip.

[0028] In the embodiments described herein, the first evaluator 6 is a machine learning module. The machine learning module is configured to extract features from the electromorphic signal and pass them through a gradient boosting method to determine the variance probability.

[0029] Extractor 8 is configured to determine a set of features for each real-time potentiometer signal. The applicant has identified that this can achieve excellent variance detection when machine learning techniques are applied. These features may be classified into two main feature categories.

[0030] The first category relates to features that are essentially relevant to the temporal analysis of a potentiometer signal, i.e., the time-domain or frequency-domain analysis of a potentiometer signal. In the embodiments described herein, the extractor 8 may extract a first period length estimate, a second period length estimate, and / or the frequency in the fast Fourier transform of the potentiometer signal having the highest amplitude. In the latter case, the amplitude may also be retained as a feature.

[0031] The first and second period length estimates may be determined as follows:

[0032] Real-time electromorphism signal data is stored in a vector, where each element corresponds to its sample. Next, using the autocorrelation parameter T, two vectors of size T are defined, a first vector containing the first T vector samples, and a second vector containing the last T vector samples.

[0033] In the embodiments described herein, the first estimate is calculated by measuring the normalized autocorrelation by dividing the scalar product of the first vector and the second vector by the product of the Euclidean norms of the first vector and the second vector. By varying T, the maximum and minimum values ​​of the first estimate are determined, and a value is selected for the first estimate, which is the value provided by T that results in the first estimate calculated with this T value being greater than the maximum value of the first estimate minus 0.3 times the difference between the maximum and minimum values ​​of the first estimate.

[0034] Alternatively, the first estimate can be determined by changing the coefficient, for example, 0.3, or by finding a T-value that optimizes the estimate empirically, exhaustively, or in other ways.

[0035] In the embodiments described herein, the second estimate is calculated for each T-value by dividing the square of the Euclidean norm of the difference between the first vector and the second vector by the product of the sample that maximizes the absolute value of the first vector and the sample that maximizes the absolute value of the second vector. Similar to the first estimate, the maximum and minimum values ​​of the second estimate are determined, and the values ​​selected for the second estimate are those provided by the T-value such that the second estimate calculated with these T-values ​​is smaller than the minimum value of the second estimate plus 0.2 times the difference between the maximum and minimum values ​​of the second estimate. Alternatively, the first estimate can also be determined by changing the coefficient, for example, 0.2, or by finding a T-value that optimizes the estimate empirically, exhaustively, or in other ways.

[0036] The second category relates to morphological analysis of signals, i.e., analysis based on or related to mathematical morphology. In the embodiments described herein, extractor 8 may extract the Euclidean norm of the electromorphic signal and the integrated absolute derivative of the electromorphic signal.

[0037] By using all the features described above, the gradient boosting-based machine learning module 10 can use up to six different features for each real-time electromorphic signal. Needless to say, in other embodiments, the extractor 8 may extract more features. These features can relate to combinations or comparisons of electromorphic signals associated with different electrodes by sharing the same timecode. In some embodiments, the applicant has utilized up to 70 features in the gradient boosting-based machine learning module 10.

[0038] The gradient boosting-based machine learning module 10 is trained using 275,000 electromorphic signals labeled with values ​​indicating whether they exhibit variance. First, features are extracted for each of these electromorphic signals, and then the gradient boosting-based machine learning module 10 is trained using logarithmic loss in the embodiments described herein. After training, the gradient boosting-based machine learning module 10 may be used to receive a set of features extracted from the set of electromorphic signals as input, and to return an array of values ​​each indicating the probability that one corresponding electromorphic signal in the set exhibits variance. Optionally, the training data may include catheter shape information and / or related 3D information.

[0039] In the embodiments described herein, the second evaluator 12 is a convolutional neural network. The convolutional neural network receives a set of electromorphic signals and outputs an array of values ​​that each represent the probability that one corresponding electromorphic signal in the set exhibits variance.

[0040] Figure 2 shows an exemplary structure for the second evaluator 12's convolutional neural network.

[0041] In the embodiments described herein, the neural network is a convolutional neural network having five convolutional layers. Thus, a set of electromorphic signals 100 (e.g., including 16 electromorphic signals) is processed by a first convolutional layer 110 that extracts 200 features, then by a second convolutional layer 120 that extracts 3000 features from layer 110, a third convolutional layer 130 that extracts 20000 features from layer 120, a fourth convolutional layer 140 that extracts 8000 features from the maximum pooling of layer 130, and a fifth convolutional layer 150 that extracts 1000 features from layer 140.

[0042] The fifth convolutional layer 150 is linked to the fully connected layer 160 of the neural network. The fully connected layer 160 contains two layers of neurons, each containing 50 neurons and 10 neurons. The fully connected layer 160 returns a probability array 170 to the output layer.

[0043] The training of this convolutional neural network is performed using 275,000 electrologic signals labeled with values ​​indicating whether or not they exhibit variance. The electrologic signals are used in groups of 16. Each electrologic signal has a corresponding value of 0 or 1 indicating the presence or absence of variance. In the embodiments described herein, training is performed using the Adam optimizer and binary cross-entropy loss. Other optimizers and losses may be used, or the optimizer may be bypassed.

[0044] Predictor 14 is used to reconcile the predictions of the first evaluator 6 with those of the second evaluator 12. More specifically, the second evaluator 12 is often more robust to overfitting than the first evaluator 6, and as a result, the first evaluator 6 is usually more accurate. Therefore, predictor 14 can perform a weighted average of the values ​​of the probability arrays output by the first evaluator 6 and the second evaluator 12, respectively. In one embodiment, the weights can be 0.7 for the first evaluator 6 and 0.3 for the second evaluator 12. Because convolutional neural networks are more prone to drift than machine learning methods based on gradient boosting, if the difference between the predictions of the first evaluator 6 and the second evaluator 12 is too large, predictor 14 may, at its discretion, be configured to retain only the value of the first evaluator 6. Thus, the output of predictor 14 is another probability array in which a given real-time potentiometer signal shows its variance.

[0045] Furthermore, the computer 14 may determine the color based on the predictive sequence. More specifically, this determination may be based on the predictive sequence associated with consecutive time codes. If the patient's atrial fibrillation is determined to be slow, two predictive sequences corresponding to consecutive time codes can be averaged. The values ​​of this sequence can be associated as follows: blue if the value is less than 0.35, orange if the value is 0.35 or greater and 0.65 or less, and red otherwise. If the patient's atrial fibrillation is determined to be fast, four predictive sequences corresponding to consecutive time codes can be averaged. The values ​​of this sequence can be associated as follows: blue if the value is less than 0.5, orange if the value is 0.5 or greater and 0.8 or less, and red otherwise. The determination of whether the atrial fibrillation is fast or slow is made by a physician and entered into the device. This allows for a visual response on the display (as shown by reference numeral 20 in Figure 1). This can be used by a physician to automatically, manually, or semi-manually tag cardiac regions resulting in dispersion areas shown by the dots in reference numeral 22 in Figure 1.

[0046] The above describes the use of two evaluators, but an apparatus or method using two or more evaluators, including the first and second evaluators according to the present invention, is also included in the scope of the appended claims.

Claims

1. A computer device for real-time analysis of electrophysics related to the heart, A memory (4) is arranged to receive multiple real-time electromagnetism signals transmitted from multiple electrodes, A first evaluator (6) comprising an extractor (8) and a gradient boosting-based machine learning module (10), wherein the extractor (8) is configured to extract from each electrographic signal in a set of electrographic signals, which are signals relating to the heart, a set of features comprising at least one temporal analysis feature relating to a feature essentially relevant to a time-domain or frequency-domain analysis of the electrographic signal and at least one morphological feature relating to an analysis based on or related to the mathematical morphology of the electrographic signal, and to supply the obtained set of features to the gradient boosting-based machine learning module (10), the machine learning module (10) being trained on data comprising a set of features labeled with a value indicating whether the relevant electrographic signal exhibits dispersion, and configured to output a first probability array for each set of electrographic signals indicating whether each electrographic signal in the set exhibits dispersion, A second evaluator (12) including a convolutional neural network, wherein the convolutional neural network receives a set of real-time electromorphic diagram signals, outputs a second probability array in which the input real-time electromorphic diagram signals represent variance, and the convolutional neural network is trained on a database of the electromorphic diagram signals, each labeled with a value indicating whether each of the electromorphic diagram signals represents variance. A predictor (14) that returns a third probability array which is at least partially based on a weighted average of the values ​​of the first probability array and the second probability array, based on the first probability array and the second probability array determined for a given set of the potential diagram signals, A computer device equipped with the following features.

2. The computer device according to claim 1, wherein the extractor (8) is configured to extract at least one temporal analysis feature from a group consisting of a first period length estimate, a second period length estimate having a different period length from the first period length estimate, and frequencies in the fast Fourier transform of the potentiometer signal having the maximum amplitude.

3. The computer device according to claim 1 or 2, wherein the extractor (8) is configured to extract at least one morphological feature from a group consisting of the Euclidean norm of the electromagnetism signal, which is a mathematical value derived from the electromagnetism signal, and the integrated absolute derivative of the electromagnetism signal.

4. The computer device according to any one of claims 1 to 3, wherein a real-time electromagnetism signal is arranged to be divided into a series of electromagnetism signals having a selected duration.

5. The computer device according to claim 4, arranged to provide a set of potentiometric signals having the same selected duration.

6. The computer device according to any one of claims 1 to 5, wherein the predictor (14) is configured to use the probability in the first probability array in the third probability array when the absolute value of the difference between the probability in the first probability array and the probability in the second probability array corresponding to the same potential diagram signal is greater than a threshold.

7. The computer device is configured to determine the color associated with the value in the third probability array, Each electrode further comprises a display arranged to output a color related to the probability determined for the corresponding electromagnetism signal in the third probability array. The computer device according to any one of claims 1 to 6.

8. A computer program comprising instructions for implementing a first evaluator (6), a second evaluator (12), and a predictor (14) according to any one of claims 1 to 7.

9. A data storage medium for recording the computer program described in claim 8.

10. A computer-implemented method for receiving a real-time potential diagram signal, running a first evaluator (6), a second evaluator (12), and a predictor (14) according to any one of claims 1 to 7, and returning a third probability array.

Images