Suppressing interference in electrocardiogram signals using trained neural networks

Abstract

To provide methods and systems for suppressing interference in electrocardiogram signals.SOLUTION: A method includes receiving a first electrocardiogram (ECG) signal, which is acquired in a heart of a patient and is distorted by interference. One or more external signals that sense the interference concurrently with the acquisition of the first ECG signal are received from one or more sources external to the heart. A second ECG signal, in which the interference is suppressed relative to the first ECG signal, is produced by applying a trained neural network (NN) to the first ECG signal and to the one or more external signals.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] (Cross-Reference to Related Application) This application claims the benefit of U.S. Provisional Patent Application No. 63 / 254,323 filed on October 11, 2021, the disclosure of which is incorporated herein by reference.

[0002] (Field of the Invention) The present invention relates generally to medical devices, and in particular to a method and system for suppressing interference in electrocardiogram signals.

Background Art

[0003] Various techniques, such as application of neural networks, are known in the art for improving the quality of signals acquired from an organ of a patient.

[0004] For example, U.S. Patent Application Publication No. 2020 / 0214597 describes sampling of high frequency (HF) QRS signals (or derived values or features) from a number of subjects, and the use of a deep learning convolutional neural network that enables finding and identifying features or values that, for example, (i) are sufficiently similar for the same subject across all samples, but (ii) are sufficiently different between different subjects. It is likewise disclosed that signatures are found that are sufficiently stable over a particular period of time such that these signatures are within a deviation threshold, followed by monitoring all subjects that are to be identified at least as frequently as the period of time used to establish the deviation threshold.

[0005] Indian Patent Application No. IN201841015767(A) describes the extraction of fetal ECGs from low-quality abdominal ECGs, which is an integration of multirate signal processing, the ANFIS algorithm, moving average filtering, and wavelet denoising techniques. The invention is carried out in two stages, in the first stage a multirate processing technique is used in which the majority of maternal components are identified and removed. In the second stage the remaining maternal components are identified and then removed by the ANFIS algorithm using a hybrid learning technique. The extracted signals are further post-processed to remove baseline wonder noise and other noise components to obtain a clean fetal ECG.

[0006] U.S. Patent Application Publication 2020 / 0260980 describes a self-learning dynamic electrocardiogram (ECG) examination analysis method using artificial intelligence. This method includes pre-processing data, performing cardiac activity feature detection, interference signal detection, and cardiac activity classification based on deep learning methods, performing signal quality assessment and main combinations, examining cardiac activity, performing analytical calculations regarding ECG events and parameters, and then automatically outputting reported data. This method realizes an automated analysis method for a rapid and comprehensive dynamic ECG examination process, and records information on changes in the automated analysis results. Simultaneously, it collects the changed data and feeds it back into the deep learning model for continuous training, thereby continuously improving and enhancing the accuracy of the automated analysis method. Similarly, a self-learning dynamic ECG examination analysis device using artificial intelligence is disclosed. [Overview of the Initiative] [Means for solving the problem]

[0007] Embodiments of the present invention described herein provide a method comprising receiving a first electrocardiogram (ECG) signal acquired in a patient's heart and distorted by interference. Simultaneously with the acquisition of the first ECG signal, one or more interference-sensing external signals are received from one or more sources outside the heart. By applying a trained neural network (NN) to the first ECG signal and the one or more external signals, a second ECG signal is generated in which interference is suppressed relative to the first ECG signal.

[0008] In some embodiments, the interference includes one or more spectral lines and one or more harmonics of one or more spectral lines. In other embodiments, the method includes training a neural network using (i) one or more training ECG signals that are not distorted by interference, and (ii) one or more training interference signals each having one or more individual spectral lines and one or more individual harmonics.

[0009] In one embodiment, training the NN involves training an autoencoder artificial NN having at least five layers. In another embodiment, at least one of the spectral lines includes the alternating current (AC) of the output signal.

[0010] According to embodiments of the present invention, a system including an interface and a processor is further provided. The interface is configured to receive (i) a first electrocardiogram (ECG) signal acquired in the patient's heart and distorted by interference, and (ii) one or more external signals received from one or more external sources outside the heart, and to sense interference simultaneously with the acquisition of the first ECG signal. The processor is configured to generate a second ECG signal in which interference is suppressed relative to the first ECG signal by applying a trained neural network (NN) to the first ECG signal and one or more external signals.

[0011] This invention will be more fully understood by considering the following "Modes for Carrying Out the Invention" in conjunction with the drawings. [Brief explanation of the drawing]

[0012] [Figure 1] This is a schematic diagram of a catheter-based tracking and ablation system according to an embodiment of the present invention. [Figure 2A] This is a schematic block diagram illustrating the training of a neural network (NN) to suppress interference in a distorted ECG signal according to an embodiment of the present invention. [Figure 2B] This schematic block diagram illustrates the application of a trained neural network (NN) to generate an ECG signal with reduced interference from a distorted ECG signal received from a patient's heart, according to an embodiment of the present invention. [Figure 3] This flowchart schematically illustrates a method for suppressing interference in a distorted real-time ECG signal received from a patient's heart, according to an embodiment of the present invention. [Modes for carrying out the invention]

[0013] Overview During medical procedures, such as those involving electrophysiological (EP) mapping, electrocardiogram (ECG) signals obtained from a patient's heart may be distorted by interference from other signals received from one or more sources outside the patient's heart. Examples of such signal sources may include spectral lines of a power grid with frequencies of approximately 50 Hz or 60 Hz, and at least one or more harmonics of these spectral lines.

[0014] In principle, various types of algorithms or other techniques can be used to suppress at least some of the interference and generate an undistorted ECG signal. However, interference suppression in such techniques can take a long time, and the generation of an undistorted ECG signal must be performed in real time, for example, when acquiring the ECG signal. Furthermore, it is important to present the undistorted ECG signal to the physician in real time, for example, simultaneously with the performance of the medical procedure.

[0015] Embodiments of the present invention described herein hereafter provide techniques for improving the quality of distorted ECG signals obtained from a patient's heart, wherein the quality improvement is performed in real time, for example, within less than one second of receiving the distorted ECG signal. In some cases, the distortion may be caused by spectral lines of the power grid signal and its individual harmonics, as described above.

[0016] In some embodiments, a system for improving the quality of an ECG signal by suppressing spectral lines and harmonics of a power grid signal comprises an interface and a processor.

[0017] In some embodiments, the processor is configured to accommodate a neural network (NN) model, such as an autoencoder artificial NN having at least five layers, typically about ten layers, an example of an NN model is shown in Figure 2A below.

[0018] In some embodiments, the processor is configured to train the neural network using one or more training ECG signals that are not distorted by interference from the output signals described above. The training ECG signals may be prepared in advance, for example, by suppressing distortion in ordinary ECG signals, or by generating synthetic ECG signals with features of interest, such as those relating to a specific heart disease.

[0019] In some embodiments, the processor is configured to train the neural network (NN) using one or more training interference signals, each having one or more individual spectral lines and one or more individual harmonics in addition to the training ECG signal. For example, the spectral lines and harmonics of the training interference signals may have one or more frequencies, typical of power grid output signals used in Europe-Asia or North America, e.g., around 50 Hz or 60 Hz. After training is complete, the processor has the trained NN, which can be implemented in software, hardware, or a preferred combination thereof.

[0020] In some embodiments, after training the neural network and during electrophysiology mapping, the interface is configured to receive an ECG signal acquired from the patient's heart, also referred to herein as a first ECG signal. In some cases, the first ECG signal is distorted by interference caused by spectral lines and harmonics of power grid signals.

[0021] In some embodiments, the interface may also be configured to receive one or more external signals that sense interference from one or more sources external to the heart, simultaneously with the acquisition of the first ECG signal (e.g., during electrophysiology mapping).

[0022] In some embodiments, the processor is configured to apply the trained neural network to the first ECG signal and the one or more external signals to generate a second ECG signal in which interference is suppressed with respect to the first ECG signal. It should be noted that by applying the trained neural network to the first ECG signal, the processor can generate the second ECG signal in real time (e.g., within less than one second after receiving the first ECG signal). Furthermore, the processor is configured to immediately present the second ECG signal to a physician while performing electrophysiology mapping, instead of or in addition to presenting the first ECG signal.

[0023] In other embodiments, the above techniques, with the appropriate modifications, can be applied to improve the quality of signals other than ECG signals acquired from the patient's heart or any other organ of the patient.

[0024] The techniques of the present disclosure improve the accuracy of electroanatomical (EA) mapping, and therefore improve the efficiency and quality of medical procedures based on EA mapping.

[0025] Description of the System Figure 1 is a schematic diagram of a catheter-based tracking and ablation system 20, according to an embodiment of the present invention.

[0026] In some embodiments, the system 20 comprises a catheter 22, which in this embodiment is a cardiac catheter, and a control console 24. In the embodiments described herein, the catheter 22 may be used for any preferred therapeutic and / or diagnostic purposes, such as sensing electroanatomical signals within the heart 26.

[0027] In some embodiments, the console 24 includes a processor 34, which is typically a general-purpose computer, having suitable front-end and interface circuits for receiving signals via the catheter 22 and controlling other components of the system 20 described herein. The console 24 further includes a user display 35 configured to receive a map 27 of the heart 26 from the processor 34 and to display the map 27.

[0028] In some embodiments, map 27 may include any suitable type of three-dimensional (3D) anatomical map generated using any suitable technique. For example, the anatomical map may be generated using anatomical images produced by using a suitable medical imaging system, or using a technique called fast anatomical mapping (FAM) available in the CARTO® system supplied by Biosense Webster Inc. (Irvine, Calif.), or using any other suitable technique, or any suitable combination of the above.

[0029] Refer to inset 23. In some embodiments, before performing the ablation procedure, physician 30 inserts a catheter 22 from the vascular system of patient 28 lying on a table 29 to perform electroanatomical (EA) mapping of the target tissue of the heart 26. During EA mapping, physician 30 controls the catheter 22 to sense one or more electrocardiogram (ECG) signals acquired within the tissue of the heart 26, as described herein.

[0030] In some embodiments, the catheter 22 comprises a distal end assembly 40 having a plurality of sensing electrodes (not shown). For example, the distal end assembly 40 may comprise (i) a basket-type catheter having a plurality of splines, each spline having a plurality of sensing electrodes, (ii) a balloon catheter having a plurality of sensing electrodes disposed on the surface of a balloon, or (iii) a focal catheter having a plurality of sensing electrodes (as shown in the example in Figure 1).

[0031] In some embodiments, each sensing electrode is configured to generate one or more signals indicating a sensed ECG signal in response to sensing an electrophysiological (EP) signal, such as an ECG signal, in the tissue of the heart 26.

[0032] In some embodiments, the proximal end of the catheter 22 is connected, among other things, to an interface circuit referred to herein as interface 38, or to an interface circuit of the processor 34, to transfer ECG signals to the processor 34 in order to perform EA mapping. In some embodiments, during EA mapping, the signals generated by the sensing electrodes of the distal end assembly 40 may include thousands of data points, e.g., about 50,000 or more, and these data points may be stored in the memory (not shown) of the console 24. Based on the data points, the processor 34 is configured to present wave vectors, also referred to herein as vectors, on the map 27, representing the electrical signals propagating across the surface of the heart 26.

[0033] In the context of this disclosure and in the claims, the terms “about” or “approximately” used with respect to any number or range of numbers indicate a suitable dimensional tolerance that enables a part or set of components to function in accordance with its intended purpose as described herein.

[0034] In some cases, one or more of the acquired ECG signals may be distorted by interference that may be undesirably received from one or more sources outside the heart 26. For example, the alternating current (AC) output signals of the power grid 43 have typical frequencies of approximately 60 Hz (used in North American and Japanese grids) or 50 Hz (used in European, most Asian countries, and grids on other continents), and one or more individual harmonics may undesirably interfere with the acquired ECG signals. The output signals originate outside the heart 26 and are led to the system 20, for example, to interface 38, via cable 42, and are therefore referred to herein as “external signals.” Such external signals may be perceived simultaneously with the acquisition of ECG signals sensed by electrodes of the distal end assembly 40. For example, external signals may be transmitted to the processor 34 via cable 42 and catheter 22 and enter interface 38, and thus may distort the ECG signals. Techniques for suppressing interference are disclosed in detail in Figures 2A, 2B, and 3 below.

[0035] In other embodiments, the catheter 22 may include one or more ablation electrodes (not shown) connected to a distal end assembly 40. The ablation electrodes are configured to ablate tissue at a target location in the heart 26, determined based on an analysis of the EA mapping of the target tissue in the heart 26. After determining the ablation plan, the physician 30 navigates the distal end assembly 40 to a location very close to the target location in the heart 26, for example, by using a manipulator 32 for manipulating the catheter 22. The physician 30 then places one or more of the ablation electrodes in contact with the target tissue and applies one or more ablation signals to the tissue. Additionally or alternatively, the physician 30 may use any different type of suitable catheter for ablating tissue in the heart 26 to carry out the ablation plan described above.

[0036] In some embodiments, the position of the distal end assembly 40 within the cardiac chamber is measured using a position sensor (not shown) of a magnetic position tracking system connected to the distal end assembly 40. In this embodiment, the console 24 includes a drive circuit 41 configured to drive a magnetic field generator 36 located at a known location outside the patient 28 lying on the table 29, for example, under the patient's torso. The position sensor is connected to the distal end and is configured to generate a position signal in response to a sensed external magnetic field from the magnetic field generator 36. The position signal indicates the position of the distal end of the catheter 22 in the coordinate system of the position tracking system.

[0037] This position sensing method has been implemented in various medical applications, for example, in the CARTO® system manufactured by Biosense Webster Inc. (Irvine, Calif), and is described in detail in U.S. Patents 5,391,199, 6,690,963, 6,484,118, 6,239,724, 6,618,612 and 6,332,089, International Publication 96 / 05768, and U.S. Patent Application Publications 2002 / 0065455(A1), 2003 / 0120150(A1) and 2004 / 0068178(A1), all of which are incorporated herein by reference.

[0038] In some embodiments, the coordinate system of the position tracking system is aligned with the coordinate systems of system 20 and map 27, so that the processor 34 is configured to display the position of the distal end assembly 40 on an anatomical or EA map (e.g., map 27).

[0039] In some embodiments, the processor 34 typically comprises a general-purpose computer, which is programmed with software that performs the functions described herein. This software can be downloaded electronically to the computer, for example, over a network, or, alternatively or additionally, provided and / or stored on a non-transient tangible medium such as magnetic memory, optical memory, or electronic memory.

[0040] This particular configuration of System 20 is shown as an example to illustrate the specific problems addressed by embodiments of the present invention and to demonstrate the applicability of these embodiments in improving the performance of such systems. However, embodiments of the present invention are by no means limited to this particular type of exemplary system, and the principles described herein may also be applied to other types of medical systems.

[0041] Training a neural network to suppress interference in ECG signals Figure 2A is a schematic block diagram illustrating the training of a neural network (NN), also referred to herein as NN Model 55, for suppressing interference in a distorted ECG signal, according to an embodiment of the present invention.

[0042] In the examples of Figures 2A and 2B, the NN model 55 is implemented in software, which is processed in the processor 34 or any other suitable type of processing device. In other embodiments, the NN model 55 may be implemented in hardware, for example as a module of the processor 34, or at least in part in another electronic device (not shown) configured to exchange signals with the processor 34.

[0043] In some embodiments, the NN model 55 may comprise any suitable type of NN, such as autoencoder artificial NNs, e.g., regularized autoencoders, concrete autoencoders, variable autoencoders (VAEs), or other suitable types of NNs used to suppress interference (e.g., denoising). For example, the NN model 55 may have about 10 layers (or any other suitable number of layers, e.g., at least 5 layers) and may be based on the AlexNet NN architecture supplied by Alex Krizhevsky (Toronto, Canada), or the TensorFlow open-source machine learning platform supplied by Google AI (Mountain View, California, USA), a subsidiary of Google, or any other suitable type of NN architecture.

[0044] In some embodiments, the processor 34 is configured to receive, for example, one or more training ECG signals that are not distorted by various types of interference, such as the interference shown in Figure 1 above, via interface 38. In the example in Figure 2A, the training ECG signal is also referred to and shown herein as a clean ECG signal (CES) 33.

[0045] In some embodiments, CES33 may include an undistorted ECG signal received from the aforementioned CARTO® system having data arranged in 2.5-second blocks. For example, the signal from the sensing electrode of the distal end assembly 40 may be sampled at a frequency of about 1 kHz, so that each unnoised ECG signal 33 contains about 2,500 points used to train the NN model 55. Further or alternatively, one or more CES33 may include any other suitable unnoised ECG signals received from other sources or from a synthesized CES generated by using any suitable model.

[0046] In some embodiments, while training the NN model 55, the processor 34 is configured to receive one or more training interference signals (TIS) 44, for example, via interface 38, each training interference signal 44 having one or more spectral lines (e.g., 50 Hz or 60 Hz received from the power grid 43, as shown in Figure 1 above), and one or more individual harmonics of each spectral line. It should be noted that the TIS 44 may include one or more output signals sampled from the power grid 43, or other signals (sampled and / or synthetically generated) having other suitable spectral lines and their harmonics.

[0047] In some embodiments, the processor 34 is configured to train the NN model 55 using a sufficient number of samples of CES33 and TIS44 (e.g., about 100,000 samples, or any other preferred number of samples), and then output the trained NN66.

[0048] Note that in this embodiment, the trained NN66 is implemented in software. In other embodiments, for example, if at least a portion of the NN model 55 is implemented in hardware, then at least a portion of the trained NN66 is also implemented in hardware.

[0049] Application of neural networks to suppress interference in ECG signals Figure 2B is a schematic block diagram illustrating the application of a trained NN66 to suppress interference in a distorted ECG signal sensed in the heart 26, according to an embodiment of the present invention.

[0050] In some embodiments, during EA mapping, the processor 34 is configured to receive a real-time ECG signal (RTES) 77, which is typically distorted by the interference described in Figures 1 and 2A above, from the sensing electrodes of the distal end assembly 40 placed in contact with the heart 26, via interface 38 (for example, via interface 38).

[0051] In some embodiments, the processor 34 is configured to receive one or more real-time external signals (RTEXs) 88, also referred to herein as real-time spectral lines and harmonics, from one or more sources outside the heart 26 (e.g., a power grid 43 received via cable 42 and interface 38). In the context of this disclosure, the term “real-time” refers to, for example, the time interval of the EA mapping procedure when the ECG signal is sensed in the heart 26 by the sensing electrodes of the distal end assembly 40.

[0052] In some embodiments, the processor 34 is configured to apply a trained NN66 to generate a noise-free real-time clean ECG signal (RTCES) 99, which is an ECG signal with interference suppressed relative to the distorted ECG signal received from the heart 26 of the patient 28, RTES 77, during the EA mapping procedure. In other words, the ECG signal received from the heart 26 is distorted by "linear noise" interference, for example, by interference from the output signal received from the power grid 43. The trained neural network 66 is applied to the distorted ECG signal to remove (e.g., subtract) the linear noise and to generate a "less noise-free" ECG signal, such as RTCES 99, from which at least the "linear noise" interference has been removed.

[0053] In some embodiments, when a trained NN66 is applied to the acquired ECG signal, the process of "de-noising" the signal acquired from the heart 26 (e.g., RTES77) is performed immediately (e.g., within less than 1 second), and as a result, the processor 34 can display the corresponding "noise-free" ECG signal, e.g., RTCES99, to the physician 30. It should be noted that the characteristics of the signal acquired from the heart 26 (e.g., RTES77) are preserved in the corresponding "noise-free" ECG signal (e.g., RTCES99), and only distortion-causing interferences are removed.

[0054] Figure 3 is a flowchart illustrating a schematic method for suppressing interference in a distorted real-time ECG signal 77 received from the heart 26 of patient 28, according to an embodiment of the present invention.

[0055] This method, in a neural network, begins with training step 100, where the trained NN model 55 uses undistorted ECG signals, such as training interference signals with spectral lines and individual harmonics, including CES33 and TIS44, in the processor 34. Step 100 ends with acquiring the trained NN66, as shown in Figure 2A above.

[0056] In the ECG sensing step 102, as shown in Figure 2B above, the distal end assembly 40 of the catheter 22 is inserted into the heart 26 to perform EA mapping, and the sensing electrodes of the distal end assembly 40 are used to acquire a first ECG signal, such as RTES 77, which is distorted by interference and received by the processor 34.

[0057] In the external signal reception step 104, the processor 34 receives an external signal (e.g., RTEX88) from one or more external sources of information outside the heart 26, which includes interference that is sensed at the same time as the acquisition of the first ECG signal (e.g., RTES77). In this embodiment, RTEX88 is based on a signal received from the power grid 43 via the cable 42 and interface 38, as shown in Figures 1 and 2B above.

[0058] In step 106, which completes this method by generating a noise-free ECG signal, the processor 34 applies a trained NN66 to the first ECG signal (e.g., RTES77) and an external signal (e.g., RTEX88) to generate a second ECG signal (e.g., RTCES99) in which interference in RTES77 is suppressed, as shown in Figure 2B above.

[0059] While the embodiments described herein primarily address the improvement of the quality of ECG signals perceived in a patient's heart, the methods and systems described herein can also be used in other applications, such as electroencephalography (EEG) procedures.

[0060] Accordingly, it will be understood that the embodiments described above are cited as examples, and that the present invention is not limited to those specifically shown and described above. Rather, the scope of the present invention includes both combinations and partial combinations of the various features described herein, as well as variations and modifications thereof not disclosed in the prior art, which would be conceivable to those skilled in the art upon reading the foregoing description. Documents incorporated by reference in this patent application shall be deemed to be part of this application, except that if any term is defined in such incorporated documents in a manner that contradicts the definitions expressed or implied herein, only the definitions herein shall be considered.

[0061] [Implementation Method] (1) Receiving a first electrocardiogram (ECG) signal acquired in the patient's heart and distorted by interference, Simultaneously with acquiring the first ECG signal, one or more external signals that detect the interference are received from one or more sources outside the heart, Applying a trained neural network (NN) to the first ECG signal and one or more external signals generates a second ECG signal in which the interference is suppressed with respect to the first ECG signal. Methods that include... (2) The method according to Embodiment 1, wherein the interference includes one or more spectral lines and one or more harmonics of the one or more spectral lines. (3) The method of Embodiment 2, comprising training the NN using (i) one or more training ECG signals that are not distorted by the interference, and (ii) one or more training interference signals each having one or more individual spectral lines and one or more individual harmonics. (4) The method according to Embodiment 2, wherein training the NN includes training an autoencoder artificial NN having at least five layers. (5) The method according to Embodiment 2, wherein at least one of the spectral lines includes an alternating current (AC) of the output signal.

[0062] (6) A system, (i) a first electrocardiogram (ECG) signal acquired in the patient's heart and distorted by interference, and (ii) an interface configured to receive one or more external signals received from one or more external sources outside the heart, which simultaneously acquire the first ECG signal and detect the interference. The system comprises a processor configured to generate a second ECG signal in which interference is suppressed with respect to the first ECG signal by applying a trained neural network (NN) to the first ECG signal and one or more external signals. (7) The system according to Embodiment 6, wherein the interference includes one or more spectral lines and one or more harmonics of the one or more spectral lines. (8) The system according to Embodiment 7, wherein the processor is configured to train the NN using (i) one or more training ECG signals that are not distorted by the interference, and (ii) one or more training interference signals each having one or more individual spectral lines and one or more individual harmonics. (9) The system according to embodiment 7, wherein the processor is configured to train an autoencoder artificial neural network having at least five layers. (10) The system according to Embodiment 7, wherein at least one of the spectral lines includes an alternating current (AC) of the output signal.

Claims

1. It is a system, (i) a first electrocardiogram (ECG) signal acquired by a sensing electrode placed in contact with the patient's heart and distorted by interference, and (ii) an interface configured to receive one or more external signals, which are power outputs causing the interference, received from one or more power grids outside the heart connected to the system, simultaneously with the acquisition of the first ECG signal. A processor is configured to generate a second ECG signal in which interference is suppressed with respect to the first ECG signal by applying a trained neural network (NN) to the first ECG signal and one or more external signals. A system in which the processor is configured to generate the trained NN by training the NN using (i) one or more training ECG signals that are not distorted by the interference, and (ii) one or more training interference signals in the external signal, which include an output wave having a power frequency and one or more harmonics.

2. The system according to claim 1, wherein the processor is configured to train an autoencoder artificial NN having at least five layers.

3. A program, wherein when the program is read by a processor, the processor: The system receives a first electrocardiogram (ECG) signal, which is acquired by a sensing electrode placed in contact with the patient's heart and distorted by interference. Simultaneously with acquiring the first ECG signal, one or more external signals, which are power output signals causing the interference, are received from one or more power grids outside the heart connected to the processor. By applying a trained neural network (NN) to the first ECG signal and one or more external signals, a second ECG signal is generated in which the interference is suppressed with respect to the first ECG signal. A program that performs the following: (i) generating a trained NN by training the NN using one or more training ECG signals that are not distorted by the interference, and (ii) one or more training interference signals in the external signal, which include an output wave having a power frequency and one or more harmonics.

4. The program according to claim 3, wherein training the NN includes training an autoencoder artificial NN having at least five layers.

Images