Noise reduction of intracardiac electrocardiograms using autoencoders, and utilization and refinement of intracardiac and surface electrocardiograms using deep learning training loss functions.

Abstract

To provide a system and a method for reducing noise of intracardiac electrocardiograms using an autoencoder and for utilizing and refining intracardiac and body surface electrocardiograms using one or more deep learning training loss functions.SOLUTION: A system and a method include a memory storing a processor-executable code for a denoised autoencoder, and one or more processors coupled to the memory. The processors execute the processor-executable code to receive raw signal data comprising signal noise, encode, by the denoised autoencoder, the raw signal data by performing a denoising autoencoder operation to produce a latent representation, and decode, by the denoised autoencoder, the latent representation to produce clean signal data reconstructed without the signal noise.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] (Cross-reference of related applications) This application claims the interests of U.S. Provisional Patent Application No. 63 / 034,694 (JNJBIO-6332USPSP1), filed on 4 June 2020, and U.S. Provisional Patent Application No. 63 / 061,929 (JNJBIO-6368USPSP1), filed on 6 August 2020, which are incorporated by reference as if they were fully described.

[0002] (Field of Invention) This invention relates to signal processing, artificial intelligence, and machine learning. More specifically, the invention relates to a system and method for reducing noise in intracardiac electrocardiograms using an autoencoder and for utilizing and refining intracardiac and surface electrocardiograms using one or more deep learning training loss functions. [Background technology]

[0003] Treatment for cardiac diseases such as cardiac arrhythmias often requires heart mapping (i.e., mapping the cardiac tissue, ventricles, veins, arteries, and / or pathways, also known as cardiac mapping). An electrocardiogram (ECG) is an example of cardiac mapping. An ECG is generated from electrical signals from the heart that describe the activity of the heart. ECGs are used during cardiac procedures to identify the potential origin of cardiac disease.

[0004] An autoencoder can be used to refine the electrical signals of the heart through encoding and decoding operations. This refined electrical signal of the heart can then be used to generate an ECG. This autoencoder utilizes artificial intelligence and machine learning to build and train its encoding and decoding operations. For example, a denoising autoencoder is trained to discover more robust features / representations within the electrical signal, preventing it from learning specific identifying information in the electrical signal. When attempting to extract or learn important features / representations during autoencoder training, current mean squared error (MSE) functions and / or other regression loss functions may fail to highlight clinically important zones or events (e.g., potential origins of cardiac disease). [Overview of the Initiative] [Means for solving the problem]

[0005] According to one embodiment, a method is provided. This method includes receiving raw signal data containing signal noise and encoding the raw signal data by a denoising autoencoder. This encoding includes performing a denoising autoencoder operation on the raw signal data to generate a potential representation. This method also includes decoding the potential representation by a denoising autoencoder to generate clean signal data restored without signal noise.

[0006] According to one embodiment, another method is provided. This method is carried out by a training algorithm executed by a memory-coupled processor. This training algorithm applies a first filter to a signal to highlight activity in the signal and generate a first modified signal. This training algorithm applies a rectifier and a second filter to the first modified signal to smooth out areas of the first modified signal that have clinical importance and generate a second modified signal. This training algorithm uses an energy threshold to automatically detect the high-frequency energy zone of the second modified signal and generate a weight vector.

[0007] According to one or more embodiments, embodiments of the above method can be implemented as a device, system, and / or computer program product. [Brief explanation of the drawing]

[0008] The following explanation, given as an example in conjunction with the attached drawings, can provide a more detailed understanding, and similar reference numbers in each drawing indicate similar elements. [Figure 1] A schematic diagram of a typical system in which one or more features of the subject matter of this disclosure may be implemented based on one or more embodiments is shown. [Figure 2] A block diagram of an exemplary system for remotely monitoring and communicating patient biometrics, based on one or more embodiments, is shown. [Figure 3] This shows a graphic representation of an artificial intelligence system based on one or more embodiments. [Figure 4] A block diagram of a method implemented in the artificial intelligence system of Figure 3, based on one or more embodiments, is shown. [Figure 5] A block diagram of a method based on one or more embodiments is shown. [Figure 6A] Examples of neural networks based on one or more embodiments are shown. [Figure 6B]A block diagram of the method implemented in the neural network of Figure 6A is shown, based on one or more embodiments. [Figure 7] Examples of contact noise based on one or more embodiments are shown. [Figure 8A] Examples of biased noise based on one or more embodiments are shown. [Figure 8B] Examples of the deflection noise in Figure 8A, with increasing x-axis values, are shown for each feature, based on one or more embodiments. [Figure 9] A graphical representation of a signal based on one or more embodiments is shown. [Figure 10] A graphical representation of a signal based on one or more embodiments is shown. [Figure 11] A graphical representation of a signal based on one or more embodiments is shown. [Figure 12] A block diagram of a method based on one or more embodiments is shown. [Figure 13] Multiple graphs illustrating signal processing by an autoencoder based on one or more embodiments are shown. [Figure 14] Multiple graphs based on one or more embodiments are shown. [Modes for carrying out the invention]

[0009] A system for utilizing and refining intracardiac and body surface electrocardiograms using one or more deep learning training loss functions (i.e., types of artificial intelligence and machine learning operations) is disclosed herein, and the functions are typically referred to herein as training algorithms. The training algorithm of the system for utilizing and refining intracardiac and body surface electrocardiograms is processor-executable code or software that is implemented by the processing hardware of a medical device instrument and necessarily within that processing hardware to provide improved ECGs and intracardiac ECGs for treating heart disease. According to an embodiment, the training algorithm provides a specific training method for a medical device instrument and an autoencoder therein. This specific training method includes multi-stage data manipulation of the electrical signals of the heart that emphasizes clinically important zones or events (e.g., potential origin locations of heart disease) within the electrical signals.

[0010] In this regard, also during operation, a training algorithm of a system for utilizing and refining intracardiac and body surface electrocardiograms receives a signal. This signal can be an electrical signal from the heart that represents the activity of the heart. The signal can be stored within a medical device apparatus or provided to the medical device apparatus. The training algorithm processes the signal to clean it. In this regard, the training algorithm applies a first filter to the signal to emphasize the activity within the signal and generates a first modified signal, applies a rectifier and a second filter to the first modified signal to smooth the region of the first modified signal having clinical significance and generates a second modified signal, uses an energy threshold to detect a high-frequency energy zone of the second modified signal and generates a weight vector. (For example, based on the weight vector, a weighted MSE loss function is defined and used to train a noise removal autoencoder). The training algorithm further constructs a training dataset including at least the weight vector and uses the training dataset to train an autoencoder (e.g., any autoencoder type). Next, the autoencoder can generate an improved electrocardiogram from one or more output intracardiac signals output by the autoencoder.

[0011] Advantages of the system for utilizing and refining intracardiac and body surface electrocardiograms (and the training algorithm therein) include reducing the overall training of the autoencoder by only a factor of 10 and increasing the accuracy and reliability of signal restoration / noise removal of the trained autoencoder. Additionally, due to the technical effects and advantages of this system, the generation of improved ECGs and intracardiac ECGs for physicians (such as during cardiac procedures) who study cardiac activity using the improved ECGs and intracardiac ECGs further enables identifying potential starting positions of heart diseases.

[0012] Figure 1 is a schematic diagram of a typical system 100 (e.g., a medical device) in which one or more features of the subject matter of this disclosure may be implemented. Using all or part of system 100, information for a training dataset can be collected and / or a training algorithm can be implemented using all or part of system 100 to train an autoencoder (e.g., a trained model). The autoencoder (e.g., an artificial intelligence and machine learning autoencoder) is processor executable code or software that is necessarily implemented by and in process operation within the processing hardware of a medical device to provide improved ECG and intracardiac ECG for treating cardiac disease.

[0013] According to one or more embodiments, an autoencoder can receive an input intracardiac signal (e.g., electrical signals of the heart including signal interference, signal artifacts, and signal noise). The intracardiac signal may be recorded and processed in real time by a monitoring and processing device (e.g., a catheter having an autoencoder therein) and / or recorded and transmitted by the monitoring and processing device to a computing device having an autoencoder therein. The autoencoder can encode the input intracardiac signal using an intracardiac dataset (e.g., a predetermined and approved electrical signal of the heart free from signal interference, signal artifacts, and signal noise). This encoding by the autoencoder generates a latent representation from the input intracardiac signal. The autoencoder can further decode its latent representation to generate an output intracardiac signal. The output intracardiac signal is the input intracardiac signal restored without signal interference, signal artifacts, and signal noise. Improved ECGs and intracardiac ECGs for treating cardiac disease can then be generated from the output intracardiac signal.

[0014] An autoencoder encodes raw signal data by performing a denoising autoencoder operation (for example, passing the raw signal data through a deep neural network such as a convolutional neural network (CNN) architecture) to reduce the dimensionality of the raw signal data and retain only the important information (for example, the denoising autoencoder operation cleans the raw signal data). In this encoding stage, the autoencoder generates a latent representation from the raw signal data that has reduced dimensionality and contains the important information.

[0015] The autoencoder further decodes its latent representation to produce a restored clean version of the raw signal data as output. That is, in this decoding stage, the autoencoder restores clean signal data from the latent representation. More specifically, the clean signal data (i.e., the output of the decoding stage) contains the input intracardiac signal restored without signal noise. The autoencoder then learns from the clean signal data to further denoise subsequent raw intracardiac signals and produce the corresponding final output (e.g., an improved ECG).

[0016] The technical effects and advantages of autoencoders include generating their output in real time, which, by generating an improved ECG, further enables physicians studying cardiac activity using that improved ECG (e.g., during cardiac procedures) to identify potential origins of cardiac disease. Note that the improved ECG is not obscured by signal noise in the raw signal data, as these issues are eliminated during encoding. Furthermore, the technical effects and advantages of autoencoders include generating an improved ECG with increased accuracy (due to the removal of signal noise), and as a result, the origins of the ventricles and atria can be provided separately in real time.

[0017] Figure 1 is a schematic diagram of a typical system 100 (e.g., a medical device) in which one or more features of the subject matter of this disclosure may be implemented. Using all or part of system 100, information can be collected for learning clean signal data (e.g., training an autoencoder), and / or using all or part of system 100, an improved ECG as described herein can be generated (e.g., implementing a trained autoencoder to denoise subsequent raw data input).

[0018] System 120 includes a probe 121, which has a shaft that allows a physician 130 to navigate through a body part of a patient 128 lying on a table 129, such as the heart 126. While multiple probes may be provided according to multiple embodiments, for brevity, this example describes a single probe 121. However, it will be understood that probe 121 may represent multiple probes. As shown in Figure 1, the physician 130 can insert the shaft 122 through the sheath 123 and simultaneously manipulate the distal end of the shaft 122 and / or deflect the sheath 123 using a manipulator near the proximal end of the catheter 140. As shown in inset 125, the catheter 140 can be attached at the distal end of the shaft 122. The catheter 140 can be inserted through the sheath 123 in a folded state and then expanded within the heart 126. The catheter 140 may include at least one ablation electrode 147 and a catheter needle, as further described herein.

[0019] System 100 may include components such as a catheter 105 configured to inflict damage on a tissue area of ​​an internal organ. This catheter 105 may also be further configured to acquire biometric data, including electrical signals of the heart (e.g., intracardiac signals). Although the catheter 105 is shown as a point catheter, it will be understood that embodiments disclosed herein can be carried out using a catheter of any shape including one or more elements (e.g., electrodes).

[0020] System 100 includes a probe 110, which has a shaft that allows a physician or medical professional 115 to navigate into body parts such as the heart 120 of a patient 125 lying on a bed (or table) 130. While multiple probes may be provided according to multiple embodiments, for brevity, only a single probe 110 is described herein. Furthermore, it is understood that probe 110 may represent multiple probes.

[0021] A typical system 100 can be used to detect, diagnose, and treat cardiac diseases (e.g., using intracardiac signals). Cardiac diseases, such as cardiac arrhythmias (especially atrial fibrillation), will never disappear, especially in the elderly population, as common and dangerous medical conditions. In a patient with normal sinus rhythm (e.g., patient 125), the heart (e.g., heart 120), including the atria, ventricles, and conduction tissue, is electrically stimulated and pulsates in a synchronously patterned manner (note that this electrical excitation can be detected as an intracardiac signal).

[0022] In patients with cardiac arrhythmias (e.g., patient 125), abnormal areas of cardiac tissue do not follow a synchronous heartbeat cycle associated with normally transmitting tissue, as in patients with normal sinus rhythm. Conversely, abnormal areas of cardiac tissue transmit abnormally to adjacent tissues, thereby disrupting the cardiac cycle into an asynchronous cardiac rhythm (note that this asynchronous cardiac rhythm can also be detected as an intracardiac signal). Such abnormal conduction has been known and occurs in various regions of the heart (e.g., heart 120), for example, in the region of the sinoatrial (SA) node along the conduction pathway of the atrioventricular (AV) node, or in the myocardial tissue forming the walls of the ventricular and atrium chambers.

[0023] Furthermore, cardiac arrhythmias, including atrial arrhythmias, can be of the multiple low-wave reentrant type, characterized by multiple asynchronous loops of electrical impulses that are scattered around the atria and often self-propagating (e.g., another example of intracardiac signaling). Alternatively, or in addition to the multiple low-wave reentrant type, cardiac arrhythmias can also have a local origin, such as when an isolated area of ​​tissue within the atria is autonomously excited in a rapid, repetitive manner (e.g., another example of intracardiac signaling). Ventricular tachycardia (V-tach or VT) is a tachycardia or rapid heart rhythm that occurs in one of the ventricles. It is a potentially fatal arrhythmia because it can lead to ventricular fibrillation and sudden death.

[0024] Atrial fibrillation, a type of arrhythmia, occurs when normal electrical impulses generated by the sinoatrial node (e.g., another example of intracardiac signals) are overwhelmed by disordered electrical impulses (e.g., signal interference) that originate in the atria and pulmonary veins and conduct irregular impulses to the ventricles. This results in an irregular heartbeat, which can last from minutes to weeks or even years. Atrial fibrillation (AF) is a chronic condition that often leads to a slight increase in the risk of death, often due to stroke. The first-line treatment for AF is medication to reduce the heart rate or restore the rhythm to normal. Furthermore, patients with AF are often given anticoagulants to protect them from the risk of stroke. The use of such anticoagulants carries its own risk of internal bleeding. In some patients, drug therapy is insufficient, and their AF is deemed untreatable with medication, i.e., incurable with standard pharmacological treatment. Synchronized electrical defibrillation can also be used to convert AF back to a normal heart rhythm. Alternatively, patients with AF may be treated with catheter ablation.

[0025] Catheter ablation-based therapies may include mapping the electrical properties of cardiac tissue, particularly the endocardium and cardiac volume, and selectively ablating cardiac tissue by applying energy. Cardiac mapping (e.g., heart mapping) may include creating, for example, a potential map of wave propagation along cardiac tissue (e.g., a voltage map) or a map of arrival times to various tissue locations (e.g., a local excitation time (LAT) map). Cardiac mapping can be used to detect localized cardiac tissue dysfunction. Ablation, such as that based on cardiac mapping, can stop or modify the propagation of undesirable electrical signals from one part of the heart to another.

[0026] Ablation is a method that disrupts unwanted electrical pathways by forming non-conductive damaged areas. Various modes of energy delivery have been disclosed to date for the purpose of forming damaged areas, including the use of microwaves, lasers, and more generally, radiofrequency energy to create conduction blocks along the cardiac tissue wall. In a two-step procedure—ablation after mapping—electrical activity points within the heart are typically detected and measured by advancing a catheter (e.g., catheter 105) containing one or more electrical sensors (e.g., at least one ablation electrode 134 of catheter 105) into the heart (e.g., heart 120) and acquiring data at numerous points. This data (e.g., biodata including intracardiac signals) can then be used to select a target region of the endocardium where ablation may be performed. Note that due to the use of an autoencoder trained by a training algorithm (e.g., used by a typical system 100), this data is more accurate than the underlying electrical signals of the ECG, including signal interference, signal artifacts, and signal noise, and can better support the selection of a target region of the endocardium for ablation. Signal interference, signal artifacts, and signal noise may be collectively referred to as artifacts in this specification.

[0027] Cardiac ablation and other cardiac electrophysiological procedures are becoming increasingly complex when physicians treat difficult conditions such as atrial fibrillation and ventricular tachycardia. The treatment of complex arrhythmias can subsequently be based on the use of three-dimensional (3D) mapping systems to reconstruct the anatomical structure of the cardiac chambers of interest. In this regard, an autoencoder trained by a training algorithm (e.g., used by a typical system 100) as described herein can provide a basic output signal, thereby generating improved 3D maps and / or ECGs for treating cardiac diseases.

[0028] For example, a cardiologist might rely on software such as the Complex Fractionated Atrial Electrograms (CFAE) module of the CARTO® 3 3D mapping system from Biosense Webster (Diamond Bar, Calif.) to generate and analyze intracardiac electrocardiograms (EGMs). An autoencoder trained by a training algorithm (e.g., used by a typical system 100) enhances this software to generate and analyze improved intracardiac electrocardiograms (EGMs), as the ablation sites can then be determined for the treatment of a wide range of cardiac conditions, including irregular atrial flutter and ventricular tachycardia.

[0029] Improved 3D maps supported by training algorithms (e.g., those used by typical system 100) can provide multiple fragments of information regarding the electrophysiological properties of tissues that represent the anatomical and functional substrates of these challenging arrhythmias.

[0030] Cardiomyopathy with different etiologies (ischemic cardiomyopathy, dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), arrhythmogenic right ventricular dysplasia (ARVD), left ventricular noncompaction (LVNC), etc.) is characterized by unhealthy tissue regions surrounded by normally functioning cardiomyocyte regions that possess a identifiable matrix.

[0031] Abnormal tissue generally features low-voltage EGM. However, initial clinical experience in endocardial-epidermal mapping indicates that low-voltage regions are not always the sole arrhythmic mechanism in such patients. Indeed, low or intermediate-voltage regions may present EGM fragmentation and delayed activity during sinus rhythm, corresponding to significant stenosis identified during sustained and organized ventricular arrhythmias, such as intolerable ventricular tachycardia. Furthermore, fragmentation and delayed EGM activity are often observed in regions exhibiting normal or near-normal voltage amplitude (>1-1.5mV). While the latter regions can be evaluated according to voltage amplitude, they may not be considered normal in relation to intracardiac signals and therefore represent true arrhythmogenic substrates. 3D mapping may allow for the identification of arrhythmogenic substrates in the endocardial and / or epicardial layers of the right / left ventricle, where their location may vary in distribution depending on the extent of the primary disease.

[0032] These substrates associated with cardiac disease are linked to the fragmentation of the endocardial and / or epicardial layers of the ventricular chambers (right and left) and the presence of delayed EGM. 3D mapping systems such as CARTO® 3 can pinpoint the location of potential arrhythmogenic substrates in cardiomyopathy in relation to abnormal EGM detection.

[0033] Electrode catheters (e.g., catheter 105) are used during medical procedures. Electrode catheters are used to stimulate and map electrical activity within the heart and to ablate areas of abnormal electrical activity. During use, the electrode catheter is inserted into a major vein or artery, such as the femoral artery, and then guided into the target cardiac chamber. A typical ablation procedure requires the insertion of a catheter with at least one electrode at its distal end into the cardiac chamber. A reference electrode is typically provided by a second catheter, either taped to the patient's skin or positioned within or near the heart. A radiofrequency (RF) current is applied to the tip electrode of the ablation catheter, causing the current to flow through the surrounding medium, i.e., blood and tissue, toward the reference electrode. The distribution of the current depends on the amount of electrode surface in contact with tissue, compared to blood, which has higher conductivity than tissue. Tissue heating occurs due to the electrical resistance of the tissue. When the tissue is sufficiently heated, cell destruction is induced in the cardiac tissue, resulting in the formation of damaged areas within the non-conductive cardiac tissue. During this process, the electrode is also heated by conduction from the heated tissue to the electrode itself. If the electrode temperature becomes sufficiently high, sometimes exceeding 60 degrees Celsius, a thin, transparent film of dehydrated blood proteins may form on the electrode surface. As the temperature continues to rise, this dehydrated layer gradually thickens, and blood coagulates on the electrode surface. Since dehydrated biological material has a higher electrical resistance than endocardial tissue, the impedance to the flow of electrical energy into the tissue also increases. When the impedance becomes sufficiently high, an impedance surge occurs, the catheter must be removed from the body, and the tip electrode must be cleaned.

[0034] The treatment of cardiac diseases such as cardiac arrhythmias often requires obtaining detailed mapping of cardiac tissue, cardiac chambers, veins, arteries, and / or electrical pathways. For example, a prerequisite for successfully performing catheter ablation is accurately identifying the cause of the cardiac arrhythmia within the cardiac chambers. Such identification can be performed via electrophysiological investigation, during which spatially resolved potentials are detected using a mapping catheter introduced into the cardiac chambers. This electrophysiological investigation is a so-called electroanatomical mapping, and therefore provides 3D mapping data that can be displayed on a monitor. Often, the mapping function and the therapeutic function (e.g., ablation) are provided by a single catheter or a group of catheters, and as a result, the mapping catheter also acts as a therapeutic (e.g., ablation) catheter simultaneously. In this case, the training algorithm, and / or the autoencoder trained by the training algorithm, can be directly stored and executed by catheter 105.

[0035] Mapping of cardiac regions of the heart (e.g., 120), such as cardiac regions, tissues, veins, arteries, and / or electrical pathways, can consequently identify problem areas such as scar tissue, sources of arrhythmias (e.g., electrorotors), and healthy areas. As further disclosed herein, cardiac regions can be mapped such that a visual representation of the mapped cardiac regions is provided using a display. Furthermore, cardiac mapping may include mapping based on one or more forms, but is not limited to, local excitation time (LAT), electrical activity, topology, bipolar mapping, dominant frequency, or impedance. Data corresponding to multiple forms can be captured using a catheter inserted into the patient's body and can be provided for representation simultaneously or at different points in time based on corresponding settings and / or the preference of the medical professional.

[0036] Cardiac mapping can be performed using one or more techniques. As an example of a first technique, cardiac mapping can be performed by detecting the electrical properties of cardiac tissue, such as LAT, as a function of its precise location within the heart. The corresponding data can be acquired using one or more catheters advanced into the heart, each having electrical and positional sensors at its distal tip. As a specific example, initially, location and electrical activity may be measured at approximately 10 to 20 locations on the inner surface of the heart. These data points are usually sufficient to generate a preliminary reconstruction or map of the cardiac surface of satisfactory quality. This preliminary map can be combined with data acquired at additional points to generate a more comprehensive map of the cardiac electrical activity. In clinical settings, it is not uncommon to collect data from more than 100 sites to generate a detailed comprehensive map of the electrical activity of the cardiac chambers. The resulting detailed map can then serve as a guideline for therapeutic actions to modify the propagation of cardiac electrical activity and restore normal rhythm, such as making decisions regarding tissue ablation.

[0037] Returning to Figure 1, to perform the cardiac mapping of interest, a medical professional 115 can insert the shaft 137 through the sheath 136 while manipulating the distal end of the shaft 137 using a manipulator 138 near the proximal end of the catheter 105 and / or deflection from the sheath 136. As shown in insert 140, the catheter 105 can be fitted at the distal end of the shaft 137. The catheter 105 can be inserted through the sheath 136 in a folded state and then expanded within the heart 120. As further described herein, the catheter 105 may include at least one ablation electrode 134 and a catheter needle.

[0038] According to several embodiments, the catheter 105 can be configured to ablate tissue areas of the cardiac chambers of the heart 120. Inset 150 shows the catheter 105 in a magnified view, inside the cardiac chambers of the heart 120. As shown in the figure, the catheter 105 may include at least one ablation electrode 134 coupled to the body of the catheter. According to other embodiments, multiple elements may be connected via splines that form the shape of the catheter 105. One or more other elements (not shown) may be provided, which may be any elements configured to perform ablation or acquire biometric data, and may be electrodes, transducers or one or more other elements.

[0039] According to several embodiments disclosed herein, at least one ablation electrode, such as an ablation electrode 134, can be configured to deliver energy to a tissue area of ​​an internal organ, such as the heart 120. The energy may be thermal energy and may cause damage to the tissue area, starting from the surface of the tissue area and extending to the thickness of the tissue area.

[0040] According to several embodiments disclosed herein, biodata may include one or more of the following: LAT, electrical activity, topology, bipolar mapping, dominant frequency, impedance, etc. The LAT may be a point in time at a threshold activity corresponding to local activation calculated based on a normalized initial start point. Electrical activity may be any applicable electrical signal that can be measured based on one or more thresholds and may be detected and / or augmented based on a signal-to-noise ratio and / or other filters. For example, an autoencoder may detect the noise type and / or quality of an electrical signal and compare the noise type and / or quality to one or more thresholds. Topology may correspond to the physical structure of a body part or part of a body part, or to changes in the physical structure of different parts of a body part or different parts of a body part. Dominant frequency may be a frequency or range of frequencies that permeate a part of a body part and may differ in different parts of the same body part. For example, the dominant frequency of the pulmonary veins of the heart may differ from the dominant frequency of the right atrium of the same heart. Impedance may be a resistance measurement in a given region of a body part.

[0041] As shown in Figure 1, the probe 110 and catheter 105 can be connected to a console 160. This console 160 may include a computing device 161, which uses a training algorithm and / or an autoencoder trained by the training algorithm, as described herein. According to one embodiment, the console 160 and / or computing device 161 include at least a processor and memory, the processor executing computer instructions with respect to the training algorithm and / or autoencoder, and the memory storing instructions for execution by the processor.

[0042] The computing device 161 can be any computing device including software and / or hardware, such as a general-purpose computer, and includes a suitable front end and interface circuit 162 for transmitting and receiving signals between the catheters 105 and controlling other components of the system 100. The computing device 161 may typically include a real-time noise reduction circuit configured as a field-programmable gate array (FPGA), followed by an analog-to-digital (A / D) electrocardiograph or electromyography (EMG) signal conversion integrated circuit. The computing device 161 can pass signals from the A / D ECG or EMG circuit to another processor and / or can be programmed to perform one or more functions disclosed herein.

[0043] For example, one or more functions include applying a first filter to a signal, thereby highlighting activity in the signal to generate a first modified signal; applying a rectifier and a second filter to the first modified signal to smooth out areas of the first modified signal that have clinical importance to generate a second modified signal; and automatically detecting the high-frequency energy zone of the second modified signal using an energy threshold to generate a weight vector used to construct a training dataset. Furthermore, one or more functions include receiving an input intracardiac signal; using the intracardiac dataset to encode the input intracardiac signal to generate a latent representation; and decoding the latent representation to generate an output intracardiac signal. The front-end and interface circuitry 162 includes an input / output (I / O) communication interface that enables the console 160 to receive signals from and / or transmit signals to at least one ablation electrode 134.

[0044] In some embodiments, the computing device 161 may be further configured to receive biometric data such as electrical activity and to determine whether a given tissue region conducts electricity. According to one embodiment, the computing device 161 may be located outside the console 160 and may be, for example, located inside a catheter, in an external device, in a mobile device, in a cloud-based device, or as a standalone processor.

[0045] As described above, the computing device 161 may include a general-purpose computer, which is programmed in software to perform the functions of the training algorithms and / or autoencoders described herein. The software may be downloaded to the general-purpose computer in electronic form, for example, over a network, or it may be provided and / or stored on a non-temporary tangible medium such as magnetic memory, optical memory, or electronic memory (e.g., any suitable volatile and / or non-volatile memory such as random access memory or a hard disk drive). The exemplary configuration shown in Figure 1 may be modified to embody the embodiments disclosed herein. Embodiments of this disclosure can be similarly applied using other system components and settings. Furthermore, the system 100 may include additional components such as elements for sensing electrical activity, wired or wireless connectors, processing and display devices.

[0046] In one embodiment, the display 165 is connected to a computing device 161. During the procedure, the computing device 161 can facilitate the representation of the body parts to be drawn on the display 165 for the medical professional 115 and can store data representing the body parts to be drawn in memory. In some embodiments, the medical professional 115 may be able to manipulate the body parts to be drawn using one or more input devices such as a touchpad, mouse, keyboard, or gesture recognition device. For example, the position of the catheter 105 can be changed using the input device, resulting in an updated drawing. In an alternative embodiment, the display 165 may include a touchscreen, which, in addition to presenting the body parts to be drawn, can be configured to receive input from the medical professional 115. Note that the display 165 may be installed in the same location, a remote location such as a separate hospital, or within a separate healthcare provider network. Furthermore, the system 100 may be part of a surgical system, which is configured to acquire anatomical and electrical measurements of a patient's organs, such as the heart 120, and to perform a cardiac ablation procedure. One example of such a surgical system is the Carto® system, sold by Biosense Webster.

[0047] Console 160 can be connected by cable to body surface electrodes, which may include adhesive skin patches attached to the patient 125. The processor can work with a current tracking module to determine the position coordinates of the catheter 105 inside a body part of the patient 125 (e.g., heart 120). The position coordinates may be based on the impedance or electromagnetic field measured between the body surface electrodes and electrodes or other electromagnetic components of the catheter 105 (e.g., at least one ablation electrode 134). In addition, or otherwise, a position pad may be placed on the surface of the bed 130 or separated from the bed 130.

[0048] System 100 may also optionally acquire biometric data, such as anatomical measurements of the heart 120, using ultrasound, computed tomography (CT), magnetic resonance imaging (MRI), or other medical imaging techniques known in the art. System 100 may acquire ECG or electrical measurements using a catheter or other sensors that measure the electrical properties of the heart 120. The biometric data, including anatomical and electrical measurements, can then be stored in a non-transient tangible medium of console 160. The biometric data can be transmitted from the non-transient tangible medium to a computing device 161. Alternatively, the biometric data can be transmitted to a server, which may be local or remote using a network, as further described herein.

[0049] According to one or more embodiments, a catheter including a position sensor can be used to determine the trajectory of a point on the surface of the heart. These trajectories can be used to infer kinetic properties such as the contractile force of the tissue. A map illustrating such kinetic properties can be constructed when trajectory information is sampled at a sufficient number of locations within the heart 120.

[0050] Electrical activity at a point within the heart 120 can typically be measured by advancing a catheter 105, which houses an electrical sensor at or near its distal tip (e.g., at least one ablation electrode 134), toward that point within the heart 120, bringing the tissue into contact with the sensor to acquire data at that point. One drawback of mapping the cardiac chambers using a catheter 105 housing only a single distal tip electrode is the long time required to accumulate data point by point over the number of essential points needed for a detailed map of the cardiac chambers as a whole. Therefore, multi-electrode catheters have been developed, as shown by the balloon catheter 135, to simultaneously measure electrical activity at multiple points within the cardiac chambers.

[0051] Multi-electrode catheters can be implemented using a linear catheter with multiple electrodes, a balloon catheter 135 including electrodes arranged on multiple aggregates forming a balloon, a lasso catheter or loop catheter with multiple electrodes, or any other applicable shape. A linear catheter may be elastic in whole or in part, and as a result, it can twist, bend, or change its shape based on received signals and / or on the application of external force (e.g., cardiac tissue) to the linear catheter. A balloon catheter 135 can be designed so that its electrodes can be held in close contact with the surface of the endocardium when deployed in the patient's body. As an example, a balloon catheter 135 may be inserted into a lumen such as a pulmonary vein (PV). A balloon catheter 135 may be inserted into a PV in a deflated state, and as a result, the balloon catheter does not occupy its maximum volume while inserted into the PV. The balloon catheter 135 can be inflated while such electrodes on the balloon catheter 135 are in contact with the entire circular area of ​​the PV inside the PV. Such contact of the entire circular portion of the PV, or any other lumen, can enable efficient mapping and / or ablation.

[0052] In one example, a multi-electrode catheter can be advanced into the cardiac chambers of the heart 120. Dorsal-ventral (AP) and transverse fluorescence maps can be obtained to establish the position and orientation of each electrode. EGM can be recorded from each electrode in contact with the cardiac surface against transient criteria such as the generation of P waves in sinus rhythm from a body surface ECG. As further disclosed herein, the system can distinguish between those electrodes that register electrical activity and those that do not due to proximity to the endocardial wall. After the initial EGM is recorded, the catheter can be repositioned, and fluorescence maps and EGM can be recorded again. An electrical map can then be constructed from repeated steps of the above process.

[0053] In one example, cardiac mapping can be generated based on the detection of intracardiac potential fields. Non-contact techniques can be implemented to simultaneously acquire vast amounts of cardiac electrical information. For example, a catheter having a distal end portion may be provided with a series of sensor electrodes, which may be distributed across the surface of the catheter and connected to an insulated conductor for connection to signal detection and processing means. The size and shape of the end portion may be such that the electrodes are substantially spaced apart from the walls of the cardiac chambers. Intracardiac potential fields may be detected during a single heartbeat. In one example, the sensor electrodes may be distributed on a series of peripheral portions in planes spaced apart from each other. These planes may be perpendicular to the long axis of the end portion of the catheter. At least two additional electrodes may be provided adjacent to the end portion along the long axis of the end portion. In a more specific example, the catheter may include four peripheral portions, each having eight electrodes spaced equally apart on each peripheral portion. Thus, in this particular embodiment, the catheter may include at least 34 electrodes (32 circumferential electrodes and 2 end electrodes).

[0054] In another example, electrophysiological cardiac mapping systems and techniques based on non-contact and non-expandable multi-electrode catheters can be implemented. EGM can be acquired using a catheter having multiple electrodes (e.g., 42 to 122 electrodes). According to this embodiment, knowledge of the relative geometric shapes of the probe and endocardium can be acquired by independent imaging methods such as transesophageal echocardiography. After independent imaging, the surface potential of the heart can be measured using non-contact electrodes, and a map can be constructed from there. This technique may include the following steps (after the independent imaging step): (a) measuring potentials using multiple electrodes arranged on a probe positioned within the heart 120; (b) determining the geometric relationship between the probe surface and the endocardial surface; (c) generating a matrix of coefficients representing the geometric relationship between the probe surface and the endocardial surface; and (d) determining the endocardial potential based on the electrode potentials and the coefficient matrix.

[0055] In another example, techniques and apparatus for mapping the potential distribution of cardiac chambers can be implemented. An intracardiac multi-electrode mapping catheter assembly may be inserted into the patient's heart 120. This mapping catheter assembly may include a multi-electrode array with an integrated reference electrode, or preferably a companion reference catheter. These electrodes can be used in the form of a substantially spherical array. This electrode array may be spatially referenced to a point on the endocardial surface by a reference electrode or reference catheter that comes into contact with the endocardial surface. A preferred electrode array catheter can carry a number of individual electrode sites (e.g., at least 24). Furthermore, this exemplary technique can be implemented in conjunction with knowing the location of each electrode site on the array, as well as knowing the geometric shape of the heart. These locations are preferably determined by impedance hemodynamic techniques.

[0056] In another example, a cardiac mapping catheter assembly may include an electrode array that defines a number of electrode sites. This mapping catheter assembly may also include a lumen for receiving a reference catheter having a distal tip electrode assembly that can be used to examine the cardiac wall. This mapping catheter may include a braid of insulated wires (e.g., having 24 to 64 wires in the braid), each of which can be used to form an electrode site. This catheter may be easily positionable within the heart 120 so as to be used to acquire electrical activity information from a first set of non-contact electrode sites and / or a second set of contact electrode sites.

[0057] In another example, a different catheter can be implemented to map electrophysiological activity within the heart. The catheter body may include a distal tip adapted for delivering stimulating pulses to regulate the heart's pace, or an ablation electrode for ablating tissue in contact with the tip. The catheter may further include at least a pair of orthogonal electrodes, which can generate a difference signal indicating local cardiac electrical activity adjacent to the orthogonal electrodes.

[0058] In another example, a process can be carried out to measure electrophysiological data within the cardiac chambers. This method may partially include inserting a pair of active and inactive electrodes into the heart 120, generating an electric field within the cardiac chambers by supplying current to the active electrodes, and measuring the electric field at the site of the inactive electrodes. The inactive electrodes are housed in an array positioned on the inflatable balloon of a balloon catheter. In a preferred embodiment, the array is said to have 60 to 64 electrodes.

[0059] In another example, cardiac mapping can be performed using one or more ultrasound transducers. These ultrasound transducers can be inserted into the patient's heart 120 and can collect multiple ultrasound slices (e.g., two-dimensional or three-dimensional slices) at various locations and orientations within the heart 120. The location and orientation of a given ultrasound transducer may be known, and its collected ultrasound slices may be stored and later displayed. Later, one or more ultrasound slices corresponding to the position of a probe (e.g., a therapeutic catheter) may be displayed, and the probe may be superimposed on one or more ultrasound slices.

[0060] In other examples, body patches and / or body surface electrodes can be positioned on or adjacent to the patient's body. A catheter with one or more electrodes can be positioned within the patient's body (e.g., within the patient's heart 120), and the catheter's position can be determined by the system based on signals transmitted between the catheter's one or more electrodes and the body patches and / or body surface electrodes. Furthermore, the catheter electrodes can detect biometric data (e.g., LAT values) from within the patient's body (e.g., within the heart 120). The biometric data can be associated with the determined position of the catheter, and as a result, a drawing of the patient's body part (e.g., the heart 120) can be displayed, showing the biometric data superimposed on the body shape.

[0061] Now, turning to Figure 2, a block diagram of an exemplary system 200 for remotely monitoring and transmitting biometric data (i.e., patient biometric information, patient data, or patient biometric data) is shown. In the example shown in Figure 2, the system 200 includes a monitoring and processing unit 202 associated with patient 204 (i.e., a patient data monitoring and processing unit), a local computing device 206, a remote computing system 208, a first network 210, and a second network 211. According to one or more embodiments, the monitoring and processing unit 202 may be an example of the catheter 105 in Figure 1, patient 204 may be an example of patient 125 in Figure 1, and the local computing device 206 may be an example of the console 160 in Figure 1.

[0062] The monitoring and processing unit 202 includes a patient biosensor 212, a processor 214, a user input (UI) sensor 216, a memory 218, and a transmitter-receiver (i.e., transceiver) 222. During surgery, the monitoring and processing unit 202 acquires and / or receives at least a portion of the biosensor data representing any acquired patient biosensor data, as well as additional information associated with any acquired patient biosensor data from one or more other patient biosensor monitoring and processing units. The additional information may be, for example, diagnostic information and / or additional information obtained from additional devices such as wearable devices.

[0063] The monitoring and processing device 202 may use the training algorithm described herein, apply a first filter to the signal to highlight activity in the signal and generate a first modified signal, apply a rectifier and a second filter to the first modified signal to smooth out areas of the first modified signal that are of clinical importance and generate a second modified signal, and use an energy threshold to automatically detect high-frequency energy zones of the second modified signal and generate a weight vector, such that larger weights correspond to areas of clinical importance and smaller weights are associated with other areas.

[0064] The monitoring and processing device 202 may use the autoencoder described herein and may process data including acquired biometric data, as well as any biometric data received from one or more other patient biometric monitoring and processing devices. For example, when processing data in this regard, the autoencoder includes a neural network used to learn latent representations (or data encodings) from the biometric data in an unsupervised manner. Furthermore, the autoencoder learns to detect specific data by training the neural network (using a training algorithm) and to ignore signal interference, signal artifacts, and signal noise by considering a clean dataset without being pre-programmed with specific rules.

[0065] The monitoring and processing device 202 can continuously or periodically monitor, store, process, and transmit any number of diverse patient biometric data (e.g., acquired biometric data) via the network 210. Examples of patient biometric data, as described herein, include electrical signals (e.g., ECG signals and brain biometric data), blood pressure data, blood glucose data, and temperature data. This patient biometric data may be monitored and transmitted for treatment across any number of diverse diseases, such as cardiovascular diseases (e.g., arrhythmias, cardiomyopathy, and coronary artery disease) and autoimmune diseases (e.g., type 1 and type 2 diabetes).

[0066] The patient biosensor 212 includes, for example, one or more transducers configured to convert one or more environmental conditions into electrical signals, thereby acquiring different types of biodata. For example, the patient biosensor 212 may include one or more electrodes configured to acquire electrical signals (e.g., cardiac signals, brain signals, or other bioelectrical signals), a temperature sensor (e.g., a thermocouple), a blood pressure sensor, a blood glucose sensor, a blood oxygen sensor, a pH sensor, an accelerometer, and a microphone.

[0067] As will be described in more detail herein, the monitoring and processing device 202 may be an ECG monitor for monitoring the ECG signal of the heart (e.g., heart 120 in Figure 1). In this regard, the patient biosensor 212 of the ECG monitor may include one or more electrodes (e.g., electrodes of catheter 105 in Figure 1) for acquiring the ECG signal. The ECG signal may be used for the treatment of various cardiovascular diseases.

[0068] In another example, the monitoring and processing device 202 may be a continuous glucose monitor (CGM) for continuously monitoring a patient's blood glucose levels based on the ongoing treatment of various diseases such as type 1 and type 2 diabetes. In this regard, the patient biosensor 212 of the CGM may include a subcutaneously placed electrode (e.g., the electrode of the catheter 105 in Figure 1) which can monitor blood glucose levels from the patient's interstitial fluid. The CGM may also be a component of a closed-loop system in which blood glucose data is sent to an insulin pump for calculated insulin delivery without user intervention, for example.

[0069] The processor 214 can be configured to receive, process, and manage biometric data acquired by the patient biosensor 212, and to transmit that biometric data to a memory 218 for storage and / or across the network 210 via the transceiver 222. As described in further detail herein, data from one or more other monitoring and processing devices 202 can also be received by the processor 214 via the transceiver 222. Also, as described in further detail herein, the processor 214 can be configured to selectively respond to different tapping patterns (e.g., single taps or double taps) received from the UI sensor 216 (e.g., a capacitive sensor therein), so that different tasks of the patch (e.g., data acquisition, storage, or transmission) can be triggered based on the detected pattern. In some embodiments, the processor 214 can generate audible feedback with respect to the detection of gestures.

[0070] The UI sensor 216 includes, for example, a piezoelectric or capacitive sensor configured to receive user input such as tapping or touching. For example, the UI sensor 216 may be configured to achieve capacitive coupling in response to a patient 204 tapping or touching the surface of the monitoring and processing device 202. Gesture recognition can be implemented via any one of various capacitive types, such as resistive capacitive, surface capacitive, projected capacitive, surface acoustic wave, piezoelectric, and infrared touch. The capacitive sensor can be placed over a small area of ​​its surface or along its length, so that tapping or touching of its surface triggers the monitoring device.

[0071] The memory 218 is any non-temporary tangible medium such as magnetic, optical, or electronic memory (e.g., any suitable volatile and / or non-volatile memory such as random access memory or a hard disk drive). According to one or more embodiments, the memory 218 stores the processor executable code, software, or instructions for the training algorithm and autoencoder.

[0072] The transceiver 222 may include a separate transmitter and a separate receiver. Alternatively, the transceiver 222 may include a transceiver integrated into a single device.

[0073] According to one embodiment, the monitoring and processing device 202 may be a device located inside the body of the patient 204 (for example, implantable subcutaneously). The monitoring and processing device 202 can be inserted into the patient 204 by any applicable method, including oral infusion, surgical insertion via vein or artery, endoscopic procedures, or laparoscopic procedures.

[0074] According to one embodiment, the monitoring and processing device 202 may be a device located outside the patient 204. For example, as will be described in more detail herein, the monitoring and processing device 202 may include a patch that can be attached (e.g., attached to the patient's skin). The monitoring and processing device 202 may also include a catheter, probe, blood pressure cuff, wristband or smartwatch biometric tracker having one or more electrodes, a glucose monitor, a continuous positive airway pressure (CPAP) machine, or virtually any device that can provide input regarding the patient's health or biometric information.

[0075] According to one embodiment, the monitoring and processing device 202 may include components that are inside the patient and components that are outside the patient.

[0076] Although a single monitoring and processing unit 202 is shown in Figure 2, the exemplary system may include multiple patient vital signs monitoring and processing units. For example, the monitoring and processing unit 202 may communicate with one or more other patient vital signs monitoring and processing units. Furthermore, or otherwise, one or more other patient vital signs monitoring and processing units may communicate with the network 210 and other components of the system 200.

[0077] The local computing device 206 and / or remote computing system 208, together with the monitoring and processing unit 202, may be any combination of software and / or hardware that separately or collectively store, execute, and implement training algorithms, autoencoders, and their functions. Furthermore, as described herein, the local computing device 206 and / or remote computing system 208, together with the monitoring and processing unit 202, may be an electronic computer framework that encompasses and / or uses any number and combination of computing devices and networks utilizing various communication technologies. The local computing device 206 and / or remote computing system 208, together with the monitoring and processing unit 202, may be readily scalable, expandable, and modular, and have the ability to be changed to different services or to reconfigure some features independently of other features.

[0078] According to one embodiment, the local computing device 206 and the remote computing system 208, together with the monitoring and processing unit 202, include at least a processor and memory, wherein the processor executes computer instructions relating to a training algorithm and an autoencoder, and the memory stores these computer instructions for execution by the processor.

[0079] The local computing device 206 of system 200 can communicate with the monitoring and processing unit 202 and be configured to act as a gateway to the remote computing system 208 via a second network 211. The local computing device 206 may be, for example, a smartphone, smartwatch, tablet, or other portable smart device configured to communicate with other devices via network 211. Alternatively, the local computing device 206 may be a fixed or standalone device, such as a fixed base station including modem and / or router functionality, a desktop or laptop computer using an executable program to transmit information between the processing unit 202 and the remote computing system 208 via a wireless module of a PC, or a USB dongle. Biometric data can be transmitted between the local computing device 206 and the monitoring and processing unit 202 using short-range wireless technology standards (e.g., Bluetooth, Wi-Fi, ZigBee, Z-wave, and other short-range wireless standards) via a short-range wireless network 210, such as a local area network (LAN) (e.g., a personal area network (PAN)). In some embodiments, as will be described in more detail herein, the local computing device 206 may also be configured to display acquired patient electrical signals and information associated with those acquired patient electrical signals.

[0080] In some embodiments, the remote computing system 208 can be configured to receive at least one of monitored patient biometric information and information associated with the monitored patient via a long-range network, such as a network 211. For example, if the local computing device 206 is a mobile phone, the network 211 may be a wireless cellular network, and information can be transmitted between the local computing device 206 and the remote computing system 208 via a wireless technology standard, such as one of the wireless technologies described above. As will be described in more detail herein, the remote computing system 208 can be configured to provide (e.g., visually and / or audibly) the patient biometric information and at least one of the associated information to medical professionals, physicians, healthcare professionals, etc.

[0081] In Figure 2, network 210 is an example of a short-range network (e.g., a local area network (LAN) or personal area network (PAN)). Information can be transmitted between the monitoring and processing unit 202 and the local computing device 206 via the short-range network 210 using one of various short-range wireless communication protocols such as Bluetooth, Wi-Fi, ZigBee, Z-wave, near-field communication (NFC), ultra-wideband wireless, Zigbee, or infrared (IR).

[0082] Network 211 may be a wired network, a wireless network, or may include one or more wired and wireless networks such as an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a direct or serial connection, a mobile phone network, or any other network or medium that facilitates communication between the local computing device 206 and the remote computing system 208. Information can be transmitted via Network 211 using any one of various long-range wireless communication protocols (e.g., TCP / IP, HTTP, 3G, 4G / LTE, or 5G / New Radio). Wired connections can be implemented using Ethernet, Universal Serial Bus (USB), RJ-11, or any other wired connection commonly known in the art. Wireless connections can be implemented using Wi-Fi, WiMAX, Bluetooth, infrared, cellular networks, satellite, or any other wireless connection technique. Furthermore, several networks can function independently or communicate with each other to facilitate communication within Network 211. In some cases, the remote computing system 208 can be implemented as a physical server on network 211. In other cases, the remote computing system 208 can be implemented as a virtual server on a public cloud computing provider (e.g., Amazon Web Services (AWS)) on network 211.

[0083] In exemplary embodiments, network 220 may be a wired network, a wireless network, or include one or more wired and wireless networks. For example, network 220 may be a long-range network (e.g., a wide area network (WAN), the Internet, or a cellular network). Information can be transmitted via network 220 using any one of various long-range wireless communication protocols (e.g., TCP / IP, HTTP, 3G, 4G / LTE, or 5G / New Radio).

[0084] In exemplary embodiments, the patient monitoring and processing unit 202 may include a patient biosensor 212, a processor 214, a user input (UI) sensor 216, a memory 218, and a transmitter-receiver (i.e., transceiver) 222. The patient monitoring and processing unit 202 can continuously or periodically monitor, store, process, and transmit any number of diverse patient biometric data via the network 210. Examples of patient biometrics include electrical signals (e.g., ECG signals and brain biometrics), blood pressure data, blood glucose data, and body temperature data. This patient biometric data may be monitored and transmitted for treatment across any number of diverse diseases, such as cardiovascular diseases (e.g., arrhythmias, cardiomyopathy, and coronary artery disease) and autoimmune diseases (e.g., type 1 and type 2 diabetes).

[0085] Figure 3 shows an artificial intelligence system 300 based on one or more embodiments. This artificial intelligence system 300 includes data 310, a machine 320, a model 330, multiple results 340, and basic hardware 350. Figure 4 shows a block diagram of a method 400 implemented in the artificial intelligence system of Figure 3. The explanation of Figures 3 and 4 will be given with reference to Figure 2 for ease of understanding.

[0086] Generally, the artificial intelligence system 300 operates method 400 by using data 310 to train a machine 320 (e.g., a local computing device 206 in Figure 2 having a training algorithm and an autoencoder on top of it) while building a model 330 to enable multiple outcomes 340 (predicted). In such a configuration, the artificial intelligence system 300 can operate with respect to hardware 350 (e.g., a monitoring and processing unit 202 in Figure 2) to train the machine 320, build the model 330, and predict outcomes using algorithms. Using these algorithms, the trained model 330 can be solved and the outcomes 340 associated with the hardware 350 can be predicted. These algorithms can typically be categorized into classification algorithms, regression algorithms, and clustering algorithms.

[0087] In block 410, this method 400 includes collecting data 310 from hardware 350. Machine 320 operates as a controller or data acquisition associated with hardware 350 and / or is associated with that hardware. Data 310 (e.g., biometric data that may be generated using the monitoring and processing device 202 in Figure 2) may be related to hardware 350. For example, data 310 may be ongoing data or output data associated with hardware 350. Data 310 may also include currently collected data, historical data, or other data from hardware 350. For example, data 310 may include measurements taken during a surgical procedure and may be associated with the outcome of the surgical procedure. For example, cardiac temperature (e.g., of patient 204) may be collected and correlated with the outcome of a cardiac procedure.

[0088] In block 420, this method 400 includes training a machine 320, such as with respect to hardware 350. This training may include analyzing and correlating data 310 collected in block 410. For example, in the case of this heart, temperature and outcome data 310 can be used to train the machine to determine whether there is a correlation or relationship between the temperature of the heart (e.g., of patient 204) during a cardiac procedure and its outcome.

[0089] Turning to Figure 5, a block diagram of Method 500 based on one or more embodiments is shown. Method 500 is an example of a training operation that may be used in block 420 of Figure 4. This training operation may include analyzing and correlating the data 310 collected in block 410 using a training algorithm stored on machine 320.

[0090] Method 500 begins in block 522, where the training algorithm applies a first filter to a signal to highlight activity within the signal and generate a first modified signal. In block 524, the training algorithm applies a rectifier and a second filter to the first modified signal to smooth out areas of the first modified signal that have clinical significance and generate a second modified signal. In block 526, the training algorithm uses an energy threshold to automatically detect the high-frequency energy zone of the second modified signal and generate a weight vector.

[0091] Next, this weight vector is used as the loss function for training the algorithm. Then, for example, this weight vector is used to build model 330 and / or build the training dataset.

[0092] According to one or more embodiments, the training algorithm applies a weighted mean squared error as the loss function. In this point of the optimization function, the neural network in its final layer reconstructs a "clean" version of the intracardiac ECG signal, where y is the clean version of the intracardiac ECG.

[0093]

number

[0094]

number

[0095]

number

[0096]

number

[0097]

number

[0098] In block 430, this method 400 includes constructing a model 330 on data 310 associated with hardware 350. Constructing the model 330 may include modeling physical hardware or software, modeling algorithms, and / or similar. This modeling may endeavor to represent the collected and trained data 310. According to one embodiment, the model 330 may be configured to model the operation of hardware 350 and model the data 310 collected from hardware 350 to predict the results achieved by hardware 350. According to one or more embodiments, with respect to an autoencoder, the model 330 receives raw signal data including signal noise, encodes the raw signal data to generate a potential representation, and decodes the potential representation to generate clean signal data restored without signal noise.

[0099] According to one or more embodiments, Model 330 highlights clinically significant zones or events within electrical signals (e.g., potential origin locations of cardiac disease) with respect to the training algorithm. According to one or more embodiments, Model 330 separates ventricular distal view and atrial activation with respect to the autoencoder, generating separate maps for atrial and ventricular activation.

[0100] In block 440, this method 400 includes predicting multiple outcomes 340 of a model 330 associated with hardware 350. This prediction for multiple outcomes 340 may be based on a trained model 330. For example, to enhance understanding of this disclosure, in the case of this heart, if the temperature during the procedure is between 36.5°C and 37.89°C (i.e., 97.7°F and 100.2°F) and produces better results from the cardiac procedure, then the outcome can be predicted in a given procedure based on the temperature of the heart during the procedure. Thus, using the predicted outcomes 340, hardware 350 can be configured to provide a specific desired outcome 340 from the hardware 350.

[0101] Figures 6A and 6B, collectively referred to as Figure 6, show an example of a neural network 600 (Figure 6A) and a block diagram of a method 601 (Figure 6B) performed within the neural network, based on one or more embodiments. This neural network 600 operates as an embodiment of an autoencoder. This neural network 600 can be implemented in hardware such as a machine 320 (e.g., local computing device 206 in Figure 2) and / or hardware 350 (e.g., monitoring and processing device 202 in Figure 2). Generally, a neural network is a network or circuit of neurons, or, in the modern sense, an artificial neural network (ANN) consisting of artificial neurons or nodes or cells.

[0102] For example, an ANN (Analytic Network) contains a network of processing elements (artificial neurons) that can exhibit complex and comprehensive behavior, determined by the connections between the processing elements and element parameters. These connections in the network or circuit of neurons are modeled as weights. Positive weights reflect excitatory connections, while negative values ​​signify inhibitory connections. Inputs are modified by the weights and summed using linear combinations. The activation function can control the amplitude of the output. For example, the acceptable range of the output is usually 0 to 1, but it can also be -1 to 1.

[0103] In most cases, an ANN is an adaptive system that changes its structure based on external or internal information flowing through the network. In more practical terms, a neural network is a nonlinear statistical data modeling or decision-making tool that can be used to model complex relationships between inputs and outputs, or to find patterns in data. Thus, ANNs can be used for predictive modeling and adaptive control applications, while being trained through datasets. Note that self-learning derived from experience can occur within an ANN, allowing it to draw conclusions from a complex set of seemingly unrelated information. The usefulness of artificial neural network models lies in the fact that they can be used to estimate functions from observations and then use them. Unsupervised neural networks can be used to learn representations of inputs that capture prominent features of the input distribution, and more recently, in deep learning algorithms that can implicitly learn the distribution functions of observed data. Learning in neural networks is particularly useful in applications where the complexity of the data or task makes it impossible to design such functions manually.

[0104] Neural networks can be used in a variety of fields. Tasks to which ANNs are applied tend to span a broad category, including regression analysis with function approximation or time series forecasting and modeling, pattern and sequence recognition, classification with novelty detection and sequential decision-making, filtering, clustering, and data processing with blind signal separation and compression.

[0105] Applications of ANN include identification and control of nonlinear systems (vehicle control, process control), game play and decision-making (backgammon, chess, racing), pattern recognition (radar systems, facial recognition, object recognition), sequence recognition (gesture, speech, handwritten text recognition), medical diagnostics, financial applications, data mining (or knowledge discovery in databases, i.e., "KDD"), visualization, and email spam filtering. For example, it is possible to create a semantic profile of a user's interests arising from photographs trained for object recognition.

[0106] Returning to Figure 6, a block diagram of Method 601 is shown based on one or more embodiments. Method 601 illustrates the operation of a neural network 600 (e.g., an autoencoder). In this neural network 600, the input layer 610 is represented by a plurality of inputs, such as 612 and 614. With respect to block 620 of Method 601, the input layer 610 receives a plurality of inputs (e.g., input intracardiac signals) as an initial operation. These plurality of inputs may be an ultrasonic signal, a radio signal, an acoustic signal, or a two-dimensional image. More specifically, these plurality of inputs may be represented as input data (X), which is raw data recorded from the atrium. Note that the desired information is within the high-frequency zone of the heart (e.g., the atrium), and the autoencoder provides a better structure of the input intracardiac signals.

[0107] In block 625 of this method 601, the neural network 600 uses an intracardiac dataset (e.g., generated as a "clean" signal by a training algorithm) to encode the input intracardiac signal and generate a latent representation. The latent representation includes one or more intermediate images introduced from the input intracardiac signal. According to one or more embodiments, the latent representation is generated by an elemental activation function of the autoencoder (e.g., an S-shaped function or a rectified linear unit), which applies a weight matrix to the input intracardiac signal and adds a bias vector to the result. Note that the weights and biases of the weight matrix, as well as the bias vector, are randomly initialized and then iteratively updated during training.

[0108] An intracardiac dataset can be clean data containing predetermined approved signals free from interference, artifacts, and noise (e.g., an example of a clean dataset). In one embodiment, the training algorithm generates one or more signals. In another embodiment, an expert medical professional, such as a physician, can scrutinize and edit the data to remove signal interference, signal artifacts, and signal noise and approve each electrical signal in the intracardiac dataset. In another embodiment, the intracardiac data may have an order of thousands or more electrical signals, but the signal morphology of each electrical signal is examined using template matching and erasure. Given the volume of electrical signals and the complexity of scrutiny, editing, and approval, the creation of the intracardiac dataset can be considered a data training portion of multi-stage data manipulation by the autoencoder.

[0109] As shown in Figure 6, inputs 612 and 614 are provided to a hidden layer 630, which is illustrated as containing nodes 632, 634, 636, and 638. Thus, layers 610, 630, and 650 can be considered encoder stages, which take the multiple inputs 612 and 614 and transfer them to a deep neural network illustrated in 630 to learn some smaller representation of the inputs (e.g., the resulting latent representation or data coding 652). The deep neural network can be a convolutional neural network, a long-short-term memory neural network, a fully connected neural network, or a combination thereof. Inputs 612 and 614 can be intracardiac ECG, ECG, or intracardiac ECG and ECG. This coding gives a dimensionality reduction of the input intracardiac signals. Dimensionality reduction is the process of reducing the number of random variables (of multiple inputs) under consideration by taking a set of primary variables. For example, dimensionality reduction can be a feature extraction process that transforms data (e.g., multiple inputs) from a high-dimensional space (e.g., more than 10 dimensions) to a low-dimensional space (e.g., 2-3 dimensions). Technical effects and benefits of dimensionality reduction include reducing the time and memory space required for the data, improving data visualization, and improving parameter interpretation for machine learning. This data transformation can be linear or nonlinear. The receiving (block 620) and encoding (block 625) operations can be considered as the data preparation portion of multi-stage data manipulation by the autoencoder.

[0110] According to one embodiment, data preparation may further include the acquisition of intracardiac electrocardiogram (IC-ECG) data from the atria (the upper cardiac chambers into which blood enters the ventricles) with simultaneous recording from the ventricles (the two lower cardiac chambers).

[0111] In block 660 of method 610, the neural network 600 decodes the latent representation to generate output intracardiac signals. The decoding stage takes the encoder output (e.g., the resulting latent representation or data encoding 652) and attempts to reconstruct several forms of the inputs 612 and 614 using another deep neural network 660. In this regard, nodes 672, 674, 676, and 678 are combined to generate outputs 692 and 694 in the output layer 690, as shown in block 699. That is, the output layer 690 reconstructs the inputs 612 and 614 in reduced dimensions, but without signal interference, signal artifacts, and signal noise. The output intracardiac signals 692 and 694 could be ventricular distal view estimates in the case of IC-ECG. Other examples of outputs 692 and 694 include intracardiac ECG, a clean version of intracardiac ECG (denoised version), ECG, and denoised ECG. A denoised version of the intracardiac ECG may not contain one or more of the following: far-field reduction, power line noise, contact noise, deflection noise, baseline fluctuation, respiratory noise, and fluro noise.

[0112] The neural network 600 performs processing through hidden layers 630 of nodes 632, 634, 636, and 638, exhibiting complex and comprehensive behavior determined by the connections between processing elements and element parameters. The target data in the output layer 650 includes ventricular activity (Y1) of target data type 1, and input data (Y2) of target data type 2 after far field reduction. Note that the far field poses problems with generating and navigating the 3D map (e.g., ventricular far field may interfere with atrial activation). Therefore, the technical effects and advantages of an autoencoder using the neural network 600 include improving the accuracy of the 3D map by removing artifacts (related to the far field).

[0113] According to one or more embodiments, a model of an autoencoder using a neural network 600 can separate between ventricular far-field views and atrial-based activations and generate separate maps for atrial and ventricular activations.

[0114] According to an embodiment, the autoencoder can be a denoising autoencoder for finding mapping functions (f, g) such that f(X)=Y1 and g(X)=Y2. Any combination of software and / or hardware (e.g., local computing device 206 and remote computing system 208 cooperating with monitoring and processing device 202) can store, execute, and implement the denoising autoencoder and its functionality separately or collectively. The denoising autoencoder trains the autoencoder to reconstruct an input from its degraded version, discovers more robust features in the hidden layer (e.g., hidden layer 630 in FIG. 6), i.e., useful features that will constitute a better and higher-level representation of the input, and prevents the input from learning a particular uniqueness (i.e., always returning to the same value). In this regard, the denoising autoencoder encodes the input (e.g., to preserve information about the input) and reverses the effect of a corruption process probabilistically applied to the input of the autoencoder.

[0115] FIG. 7 shows an example of contact noise 700 recorded within a controlled aquarium environment. Maps 1 - 4 710 (collectively and individually 7101, 7102, 7103, 7104) can be provided as mapping catheters in this example. Catheters P1 - P20 720 (collectively and individually 7201, 7202, 7203, 7204, 7205, 7206, 7207, 7208, 7209, 720 10 、720 11 、720 12 、720 13 、720 14 、720 15 、720 16 、720 17 、720 18 、720 19, 720 20 These are provided as different Pentaray catheters. In this explanatory diagram of Example 700, map 1 7101 (mapping catheter electrode 1) is in contact with Pentaray electrodes 5 7205 and 6 7206 (P5, P6), and map 2 7102 is in contact with P8 7208, which shows contact noise.

[0116] Figure 8A shows an example of deflection noise 800 recorded in a controlled tank environment. These data samples may have random start times and random durations. Figure 8B shows the deflection noise example 800 from Figure 8A, magnified in terms of features as the x-axis increases. In particular, the bottom plot 830 represents the deflection noise recorded in the tank. The bottom plot 830 shows three high-frequency bursts 8301, 8302, and 8303, indicating three points in time when the catheter was deflected. The top plot 810 represents a signal that does not contain deflection noise or contact noise. The middle plot 820 shows a signal that is the sum of deflection noise and contact noise.

[0117] These examples illustrate another method for using a denoised autoencoder. Interference (contact noise, deflection noise, power line noise, etc.) is recorded and then added to a clean version of the intracardiac ECG. Thus, denoised and clean versions of the signal are created for training purposes.

[0118] According to one or more embodiments, the denoising autoencoder implements a long-short-term memory neural network architecture, a convolutional neural network architecture, and the like. The architecture of the denoising autoencoder may be configurable with respect to the number of layers, the number of connections (e.g., encoder / decoder connections), the regularization technique (e.g., dropout or BN), and the optimization features.

[0119] Long-Short-Term Memory (TSM) neural network architectures include feedback connections and can process a single data point (e.g., an image) in cooperation with the overall order of data (e.g., audio or video). A unit of a TSM neural network architecture can consist of cells, input gates, output gates, and forget gates, where cells remember values ​​over arbitrary time intervals, and gates coordinate the flow of information in and out of the cells.

[0120] A convolutional neural network architecture is a shared weighted architecture with transformation invariance, where each neuron in one layer is connected to all neurons in the next layer. Regularization techniques for convolutional neural network architectures leverage hierarchical patterns in the data, allowing smaller and simpler patterns to be used to construct more complex patterns. If a denoising autoencoder implements a convolutional neural network architecture, other configurable aspects of that architecture may include the number of filters at each stage, kernel size, and the number of kernels per layer.

[0121] In exemplary operation according to one or more embodiments, a denoising autoencoder receives a “clean and approved” intracardiac dataset from multiple electrical signals. As shown herein, an expert medical professional, physician, etc., can scrutinize and edit the dataset to remove signal interference, signal artifacts, and signal noise and approve each electrical signal in the intracardiac dataset. The denoising autoencoder then constructs a model (e.g., Model 330 in Figure 3) from the clean and approved intracardiac dataset.

[0122] A noise-reducing autoencoder receives an input intracardiac signal, but this signal contains at least far-field artifacts. The input intracardiac signal can be recorded by one or more monitoring and processing devices (e.g., a Pentaray catheter with 20 electrodes, a basket catheter with 64 electrodes, multiple body surface leads, etc.). Note that the far field may cause problems with generating and navigating a 3D map (i.e., the ventricular far field may interfere with atrial activation).

[0123] A denoising autoencoder uses a model to encode the input intracardiac signal. This encoding provides a dimensionality reduction of the input intracardiac signal, depending on how the model, which at least removes far-field artifacts, instructs its reduction. The result of the encoding is the generation of a latent representation. This encoding provides a dimensionality reduction of the input intracardiac signal.

[0124] A denoising autoencoder decodes the latent representation to generate an output intracardiac signal and then maps the output intracardiac signal. For example, the denoising autoencoder (using its underlying architecture) finds a mapping function (f,g) such that f(X)=Y1 and g(X)=Y2.

[0125] An ECG is generated from mapped output intracardiac signals. This ECG can be generated by a computing device running a denoising autoencoder, or by another device. This ECG is improved to remove signal interference, signal noise, and signal artifacts, and then can be displayed to a physician. Improved ECGs and intracardiac ECGs dramatically reduce the time spent for each heart case.

[0126] As described herein, during intracardiac electrocardiographic mapping, the mapping catheter records both atrial and ventricular activation. In some cases, the ventricular distal view may interfere with atrial activation (e.g., signal interference), which can affect the clinical understanding and interpretation of the Carto map. According to one or more embodiments, the technical effects and benefits of a denoising autoencoder include separating the ventricular distal view from atrial-based activation and generating separate maps for atrial and ventricular activation (e.g., the denoising autoencoder uses a model during decoding to separate the ventricular distal view from atrial-based activation in one or more output intracardiac signals).

[0127] Turning to Figure 9, a graphical depiction of signal 900 based on one or more embodiments is shown. As shown by signal 900, the ECG signal includes a P wave 910 (resulting from atrial depolarization), a QRS complex 920 (resulting from atrial and ventricular repolarization), and a T wave 930 (resulting from ventricular repolarization). The ECG signal is generated by the contraction (depolarization) and relaxation (repolarization) of the atrial and ventricular muscles of the heart. To record the ECG signal, electrodes may be placed at specific locations on the human body or placed inside the human body via a catheter. Artifacts (e.g., noise) are undesirable signals that combine with electrical signals such as the ECG signal and, in some cases, create an impediment to the diagnosis and / or treatment of cardiac disease. Artifacts in electrical signals may include baseline fluctuations, power line interference, EMG noise, power line noise, etc. In other words, examples of artifacts include, but are not limited to, power noise (e.g., electrostatic and electromagnetic coupling between the circuit and 50 or 60 Hz power lines), Fluro noise (e.g., fluorescent lights), contact noise (e.g., collisions between catheter electrodes), and deflection noise (e.g., electrostatic discharge during catheter deflection).

[0128] Baseline fluctuation or drift occurs when the base axis (x-axis) of a signal appears to "fluctuate" or move up and down rather than being a straight line. This can cause the entire signal to shift from its normal base. In ECG signals, baseline fluctuation is caused by improper electrode contact (e.g., electrode skin impedance), patient movement, and periodic movement (e.g., breathing).

[0129] Figure 10 shows a graphical representation of signal 1010 shown in plot 1000 based on one or more embodiments. In this regard, signal 1010 is a typical ECG signal affected by baseline fluctuation 1020. The frequency components of baseline fluctuation 1020 are in the range of 0.5 Hz. As body movement increases during exercise or stress testing, the frequency components of baseline fluctuation increase. According to several embodiments, given that the baseline signal is a low-frequency signal, a finite impulse response (FIR) with high-pass zero-phase forward-backward filtering using a cutoff frequency of 0.5 Hz may be used to evaluate and remove baseline fluctuation 1020 in the ECG signal 1010.

[0130] Electromagnetic fields generated by power lines represent a common noise source in electrical signals such as ECGs, as well as in any other bioelectrical signals recorded from a patient's body. Such noise is characterized, for example, by 50 or 60 Hz sinusoidal interference, sometimes accompanied by many harmonics. Such narrowband noise makes the analysis and interpretation of ECGs more difficult because it compromises the reliability of low-amplitude waveform representations and can lead to spurious waveforms. When superimposed on low-frequency ECG waves such as P-waves (910) and T-waves (930), it may be necessary to remove power line interference from the ECG signal.

[0131] The presence of muscle noise can interfere in many electrical signal fields, such as the ECG field, which can cause low-amplitude waveforms to be obscured. Muscle noise, in contrast to baseline fluctuations 1020 and 50 / 60Hz interference, is not removed by narrowband filtering, but presents a different filtering problem because the spectral components of muscle activity overlap considerably with the spectral components of the PQRST complex 920. Since the ECG signal 1010 is a repeating signal, techniques for reducing muscle noise can be used in a similar manner to those used for processing evoked potentials. Figure 11 shows a graphical depiction 1100 of signal 1110 as shown based on one or more embodiments. In this regard, signal 1110 is an ECG signal interfered with by EMG noise 1120. Instruments for measuring electrical signals such as ECG signals often detect electrical interference corresponding to line, or power supply, frequency. Line frequencies in most countries are nominally set to 50Hz or 60Hz, but can vary by only a few percent from these nominal values.

[0132] Various techniques can be implemented to remove electrical interference from the electrical signal 1110. Some of these techniques use one or more low-pass filters or notch filters. For example, a system for variable filtering noise in an ECG signal can be implemented. This system may have multiple low-pass filters, for example, one filter having a 3dB point at approximately 50 Hz and a second low-pass filter having a 3dB point at approximately 5 Hz.

[0133] In another example, a system for blocking the line frequency components of an electrical signal 1110 can be realized by passing the signal through two series-connected notch filters. A system can be realized that includes a notch filter having either a low-pass coefficient, a high-pass coefficient, or both, for removing line frequency components from an ECG signal. This system can also support the removal of burst noise and can calculate the heart rate from the notch filter output.

[0134] In another example, a system comprising several units for removing interference can be realized. These units may include an averaging unit for generating an average signal over several cardiac cycles, a subtraction unit for subtracting the average signal from the input signal to generate a residual signal, a filtering unit for providing a filtered signal from the residual signal, and / or an addition unit for adding the filtered signal to the average signal.

[0135] In another example, an analog-to-digital (A / D) converter can provide noise suppression by synchronizing its clock with a phase-locked loop set to line frequency.

[0136] Furthermore, patient monitors for biometric information (e.g., biopotential) can use surface electrodes to measure biopotentials such as ECG or electroencephalography (EEG). The fidelity of these measurements is limited by the effectiveness of the connection between the electrodes and the patient. The resistance of the electrode system to the flow of current is known as electrical impedance and characterizes the effectiveness of the connection. Typically, higher impedance results in lower measurement fidelity. Several mechanisms can contribute to lower fidelity.

[0137] Signals from electrodes with high impedance are affected by thermal noise (or so-called Johnson noise) and voltage, which increases with the square root of the impedance value. Furthermore, biopotential electrodes tend to exceed the voltage noise predicted by Johnson. Amplifier systems that perform measurements from biopotential electrodes may also degrade at higher electrode impedances. These degradations are characterized by low-quality common-mode blocking, but noise sources such as patient movement and electronic equipment that may be used on or around the patient tend to increase contamination of bioelectrical signals. These noise sources are particularly prevalent in operating rooms and may include electrosurgical units (ESUs), cardiopulmonary bypass pumps (CPBs), electric motor-driven surgical saws, lasers, and other sources.

[0138] During cardiac procedures, it is often desirable to continuously measure electrode impedance in real time while monitoring the patient. To do this, a very small current is typically injected into the electrode, and the resulting voltage is measured, thereby establishing the impedance using Ohm's law. This current can be injected using a DC or AC power supply. Often, it is impossible to isolate the voltage attributable to electrode impedance from voltage artifacts resulting from interference. Interference tends to increase the measured voltage, and therefore the apparent measured impedance can cause the bioelectronic measurement system to incorrectly detect a higher impedance than the one actually presented. Often, such monitoring systems have a maximum impedance threshold limit, which can be programmed to prevent the monitoring system from operating when it detects an impedance exceeding these limits. This is especially true for systems that perform measurements of very small voltages, such as EEG. Such systems require very low electrode impedance.

[0139] Figure 12 shows a block diagram of Method 1200 based on one or more embodiments. According to one embodiment, Method 1200 is carried out by an autoencoder, which includes a training algorithm such as those described herein. According to one or more embodiments, this training algorithm can be considered a deep learning training loss function, more specifically, an autoencoder clinical weighted MSE loss function. Furthermore, the autoencoder can be operated in one or more modes. For example, a first mode may include constructing a dedicated graphical user interface that includes one or more filters. In the first method, the dedicated GUI is constructed with the ability to manually filter out noise components based on a predefined set of filters (note that the raw data signal is usually recorded with noise). Thus, the autoencoder experiences at least two sets of signals, such as the original signal with noise and the denoised signal used to train the autoencoder. A second mode may include recording noise in a controlled environment and adding the noise to clean signal data (e.g., a clean version from a previous stage). The noise may include at least one of the following: power line noise, contact noise, deflection noise, Fluro noise, and ventricular distal view.

[0140] Any combination of software and / or hardware (e.g., a local computing device 206 and a remote computing system 208 working with the monitoring and processing device 202) can store, run, and implement the autoencoder and its functions, separately or collectively. Figure 13 shows several graphs 1322, 1324, 1326, and 1328, which illustrate signal processing by the autoencoder (e.g., stages for the autoencoder clinical weighted MSE loss function) and / or signal processing by a training algorithm (e.g., stages for the autoencoder clinical weighted MSE loss function) based on one or more embodiments. The explanation of Figures 12-13 is given collectively for ease of understanding.

[0141] An autoencoder automatically trains itself to reconstruct input from a degraded version of itself. In this regard, the autoencoder forces a hidden layer (e.g., the hidden layer in Figure 6) into it to discover more robust features (i.e., useful features that constitute a better, higher-level representation of the input) and prevents the input from learning specific uniqueness (i.e., always returning to the same value). Furthermore, the autoencoder encodes the input (e.g., to store information about the input) and reverses the effects of the corruption process that was probabilistically applied to the autoencoder's input.

[0142] In general, in deep learning, the backpropagation algorithm is used when training feedforward neural networks for supervised learning. The adjective "deep" in deep learning derives from the use of multiple layers in networks such as those described herein, as shown in Figure 6. The backpropagation algorithm works by calculating the gradient of the loss function with respect to each weight according to a chain rule. These weights are updated to suppress the loss function to the minimum. Furthermore, gradient descent or variants such as stochastic gradient descent can be used.

[0143] For example, a deep learning neural network learns to map a set of inputs to a set of outputs from training data. The learning problem is assigned as a search or optimization problem, and an algorithm is used to navigate a space of possible model parameters that can be used to make a reasonable prediction of the desired output. Typically, a neural network model is trained using a stochastic gradient descent optimization algorithm and learns a set of model parameters using backpropagation of the error algorithm. The gradient descent algorithm tries to change the model parameters so that the next evaluation reduces the error, meaning that the optimization algorithm navigates to reduce the gradient (or slope) of the loss function. In this regard, the loss function is

[0144]

number

[0145]

number

[0146] As described herein, when attempting to extract or learn important features / representations (e.g., classification or autocoding of IC-ECG or surface electrocardiogram (BS-ECG)), highlighting clinically significant zones or events is important. When MSE functions, MAE functions, and / or other regression loss functions fail to highlight clinically significant zones or events, training algorithms provide improved methods for highlighting clinically significant zones or events for use with intracardiac and surface electrocardiograms. For example, the technical effects and benefits of the autoencoder clinically weighted MSE loss function include reducing the overall training of the IC-ECG denoising autoencoder by a factor of 10, and increasing the accuracy and reliability of signal recovery / denoising of the IC-ECG denoising autoencoder.

[0147] Method 1200 begins in block 1215, where the training algorithm receives a signal (see ECG signal 1342 in graph 1322 of Figure 13). This signal may be an electrical signal from the heart describing cardiac activity. The signal may be stored in a medical device or provided to a medical device on which the training algorithm is performed. In the case of a surface body ECG, signal noise in the signal (e.g., raw signal data) may include, but is not limited to, baseline fluctuations, power line interference, motion artifacts, muscle artifacts, and additional white Gaussian noise. In the case of an IcECG, signal noise in the signal (e.g., raw signal data) may include, but is not limited to, power line noise, deflection noise, contact noise, ventricular far-field noise, Fluro noise, and additional white Gaussian noise.

[0148] In block 1220, the training algorithm applies a first filter to a signal to highlight the activity within the signal and generate a first modified signal. The first filter may be a high-pass filter (implemented, for example, through hardware and / or software in a medical device), which allows portions of the ECG signal 1342 with frequencies higher than a predetermined cutoff frequency to pass through and attenuates portions of the ECG signal 1342 with frequencies lower than a predetermined cutoff frequency. The activity highlighted by the training algorithm includes atrial or ventricular activity, or zones along the ECG signal 1342 that have (atrial / ventricular) activity. According to one or more embodiments, the predetermined cutoff frequency may be set to 40 Hz (see, for example, the high-pass 40 Hz signal 1344 in graph 1324 of Figure 13 as the first modified signal).

[0149] In block 1230, the training algorithm applies a rectifier to the first modified signal to generate a second modified signal. This rectifier (implemented, for example, through hardware and / or software in a medical device) converts the alternating current of the first modified signal directly into a single current by allowing the current to flow through it only in one direction. According to one or more embodiments, the second modified signal is shown as the rectified signal 1346 in graph 1326 of Figure 13.

[0150] In block 1235, the training algorithm applies a second filter to the second modified signal to generate a third modified signal. The second filter may be a low-pass filter (implemented, for example, through hardware and / or software within a medical device), which uses a frequency lower than a predetermined cutoff frequency to pass portions of the rectified signal 1346 and a frequency higher than a predetermined cutoff frequency to attenuate portions of the rectified signal 1346. The second filter (first) of the training algorithm smooths out clinically significant regions (see, for example, the normalized low-pass signal 1348 in graph 1328 of Figure 13 as the third modified signal). Clinical significance may include the origin of cardiac disease.

[0151] In block 1240, the training algorithm automatically detects the high-frequency energy zone of the third modified signal using an energy threshold to generate a fourth modified signal (see, for example, signal 1350 in graph 1328 of Figure 13). In other words, zones with high energy are clinically detected using an energy threshold, and as a result, atrial activity is present when the energy exceeds a certain value. The energy threshold, energy threshold, and / or a specific value may be predetermined data settings of the training algorithm. According to one or more embodiments, the energy threshold may be set to 10%.

[0152] In block 1250, the training algorithm applies a weighted vector derived from a certain ratio. According to one or more embodiments, the ratio is 1:P (P>1), and as a result, the weights in zones with “atrial or ventricular” activity are P times higher than those in the “quiet” zone of the BS-ECG or IC ECG. For example, if 1:P=1:3, then 1 is high energy and 3 is low energy, and the error in signs around atrial activity would be 3 times higher than in the atrial zone. When using weighted MSE for learning, the weights are based on the signal itself.

[0153] In block 1260, method 1200 includes adding clean signals to a training dataset. The training algorithm constructs a training dataset containing at least clean signals. This training dataset may be constructed in any data structure (e.g., a data organization, management, and storage format) that enables effective access, modification, and use of the clean signals. This data structure may be stored in the memory of a medical device, as described herein. According to one or more embodiments, the training dataset may include a clean data version with an optimal filter. The training algorithm may separately record a portion of the signal noise and randomly add the signal noise back into the training dataset.

[0154] In block 1270, method 1200 includes training an autoencoder using a training dataset. The training algorithm directly instructs the autoencoder on how to find cardiac disease in the intracardiac signals using the training dataset. In block 1280, method 1200 includes generating an electrocardiogram from one or more output intracardiac signals output by the trained autoencoder using the intracardiac dataset, as described herein.

[0155] According to one or more embodiments, the autoencoder can overcome “artificial” signals, including an intrinsic cost function (with clinical significance). A stack of LSTM layers connected to a dense layer is used to detect the type of noise and the quality of the intracardiac signal. The autoencoder then “denoises” the intracardiac signal if the quality of the intracardiac signal exceeds a certain threshold. The autoencoder can denoise all types of noise. In a multi-electrode approach (e.g., using different types of diagnostic and therapeutic catheters), embodiments herein can use a neural network to “denoise” the signals from microelectrodes. The neural network may be based on a deep learning autoencoder as described herein.

[0156] According to one or more embodiments, the autoencoder comprises two components: an encoder and a decoder. The encoder maps an input (e.g., an IcECG signal) to a hidden representation (u) via a nonlinear transformation. The decoder then demaps the hidden representation back to recovered data via another nonlinear transformation, as shown in Equations 4 and 5. u = f(noisy IcECG, encoder) Equation 4 clean IcECG=g(u,decoder) Equation 5

[0157] According to one or more embodiments, a denoised autoencoder can be trained based on a loss function that emphasizes atrial activity. This loss function utilizes a neural network in which the final layer reconstructs clean signal data. The loss function may be a weighted mean squared error loss function L(θ) obeying Equation 6, where θ represents a model parameter evaluated during the training procedure.

[0158]

number

[0159]

number

[0160] Hidden representations (u) are taught using a noisy IcECG signal. When teaching a neural network, a clean IcECG signal is initially generated by filtering out power line interference, baseline fluctuations due to respiration, and high-frequency noise typically due to muscle contractions from the actual IcECG signal. A noisy IcECG signal can be generated by adding these types of noise to the clean signal in a randomized manner. In some cases, white Gaussian noise may be added. Figure 14 shows several graphs 1410 and 1440 based on one or more embodiments. Graph 1410 shows the signal-to-noise (SNR) ratio for the input, while graph 1440 shows the SNR ratio for the output. If the SNR input is low at point 14101 in graph 1410, the resulting SNR output is low at point 14401 in graph 1440. Despite the low SNR output, there is still improvement in the signal because the autoencoder cleans the signal and provides improvement. This improvement is shown within the smaller SNR of point 14401 when compared to point 14101. If the SNR input is higher at point 14106 in graph 1410 (compared to point 14101), the SNR output is higher at point 14406 in graph 1440 (compared to point 14401). Despite the higher SNR output, there is still improvement in the signal because the autoencoder cleans up the signal and provides improvement. This improvement is shown within the smaller SNR of point 14406 when compared to point 14106. The SNR improvement (imp) is calculated by subtracting the SNR input from the SNR output.

[0161] The technical effects and benefits include denoising an autoencoder with a cost function having clinical averaging, detecting different types of IcECG noise, and evaluating the quality of the desired denoised signal to mark it as a low-quality noise signal.

[0162] The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible embodiments of systems, methods, and computer program products based on various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, segment, or instruction portion, the instruction portion containing one or more executable instructions for implementing a particular logical function(s). In some alternative embodiments, the functions described within a block may be triggered in an order other than that shown in the figure. For example, two consecutively shown blocks may actually be executed substantially simultaneously, or they may sometimes be executed in reverse order depending on the relevant functionality. It should also be noted that each block in the block diagram and / or flowchart description, and combinations of blocks within the block diagram and / or flowchart description, may be implemented by a dedicated hardware-based system that performs a particular function or operation, or executes a combination of dedicated hardware and computer instructions.

[0163] While features and elements are described above in specific combinations, those skilled in the art will understand that each feature or element can be used individually or in combination with other features and elements. In addition, the methods described herein may be implemented in computer programs, software, or firmware incorporated into a computer-readable medium for execution on a computer or processor. When used herein, the computer-readable medium should not be construed as being a transient signal in itself, such as radio waves or other free-propagating electromagnetic waves, waveguides or other transmission media (e.g., optical pulses passing through optical fiber cables), or electrical signals transmitted through wires.

[0164] Examples of computer-readable media include electrical signals (transmitted via wired or wireless connections) and computer-readable storage media. Examples of computer-readable storage media include, but are not limited to, registers, cache memory, semiconductor memory devices, magnetic media such as internal hard disks and removable disks, magneto-optical media, optical media such as compact discs (CDs) and digital video discs (DVDs), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), and memory sticks. A software-related processor can be used to implement a radio frequency transceiver for use in a WTRU, UE, terminal, base station, RNC, or any host computer.

[0165] The technical terms used herein are for the purpose of describing specific embodiments only and are not intended to limit them. Where used herein, unless otherwise specified in the context, the singular forms "a," "an," and "the" also include the plural forms. Where used herein, the terms "comprises" and / or "comprising" specify the presence of the described features, integers, steps, actions, elements, and / or components, and do not exclude the presence or addition of one or more other features, integers, steps, actions, elemental parts, and / or groups thereof.

[0166] The descriptions of various embodiments herein are presented for illustrative purposes only and should not be intended to be exhaustive or limiting to the disclosed embodiments. It will be apparent to those skilled in the art that many modifications and variations are possible without departing from the scope and spirit of the embodiments described herein. The terminology used herein has been selected to best describe the principles of the embodiments, their practical applications or technical improvements across market technologies, or to enable those skilled in the art to understand the embodiments disclosed herein.

[0167] [Implementation Method] (1) A method, Receiving raw signal data containing signal noise, The raw signal data is encoded by a denoised autoencoder to generate a potential representation by performing a denoised autoencoder operation. A method comprising decoding the potential representation using the denoised autoencoder to produce clean signal data restored without the signal noise. (2) The method according to Embodiment 1, wherein the raw signal data includes at least one of power line noise, contact noise, deflection noise, fluro noise, and ventricular distal field. (3) The method according to Embodiment 1, wherein the denoised autoencoder runs in a mode for constructing a dedicated graphic user interface including one or more filters. (4) The method according to Embodiment 1, wherein the denoised autoencoder operates in a mode for recording noise in a controlled environment and adding the noise to the clean signal data. (5) The method according to Embodiment 1, wherein the raw signal data includes N channels of an intracardiac electrocardiogram, an intracardiac electrocardiogram and a surface electrocardiogram, or electrical signals of the heart from an intracardiac electrocardiogram having the position of each electrode and anatomical tissue information.

[0168] (6) The method according to Embodiment 1, wherein the denoised autoencoder is trained based on a loss function that enhances atrial activity. (7) The method according to embodiment 6, wherein the loss function utilizes a neural network in which the final layer restores clean signal data. (8) The method according to Embodiment 6, wherein the loss function includes a weighted mean squared error loss function. (9) The method according to Embodiment 1, wherein the patient biosensor of the monitoring and processing device records the raw signal data. (10) The method according to Embodiment 1, wherein the denoising autoencoder operation passes the raw signal data through a deep neural network in such a way as to reduce the dimensionality of the raw signal data and to preserve important information.

[0169] (11) The method according to Embodiment 1, wherein the potential representation includes the reduced number of dimensions from the raw signal data and important information. (12) The method according to Embodiment 1, wherein the clean signal data includes a restored input intracardiac signal that minimizes the signal noise. (13) The method according to Embodiment 1, wherein the denoised autoencoder learns the clean signal data so as to denoise subsequent raw intracardiac signals during the denoised autoencoder operation. (14) The method according to embodiment 13, wherein the denoised autoencoder detects the noise type and quality of the subsequent raw intracardiac signal. (15) The method according to embodiment 14, wherein the denoised autoencoder performs the denoised autoencoder operation on the subsequent raw intracardiac signal based on whether the quality exceeds a threshold.

[0170] (16) The method according to Embodiment 1, further comprising generating an electrocardiogram from the clean signal data, wherein the electrocardiogram substantially does not contain the signal noise. (17) A system, Memory for storing processor executable code for a denoised autoencoder, One or more processors coupled to the memory, wherein the one or more processors are configured to execute the processor executable code, and the processor executable code is Receiving raw signal data containing signal noise, The raw signal data is encoded by the denoised autoencoder to generate a potential representation by performing a denoised autoencoder operation. A system that causes the denoised autoencoder to decode the potential representation so as to generate clean signal data restored without the signal noise. (18) A method, A training algorithm executed by a memory-coupled processor applies a first filter to the signal in such a way that it enhances the activity in the signal and generates a first modified signal. The training algorithm applies the rectifier and the second filter to the first modified signal in such a way that it smooths the region of the first modified signal that is of clinical importance and generates a second modified signal. A method comprising: automatically detecting one or more high-frequency energy zones of the second modified signal using an energy threshold, such that the training algorithm generates a weight vector. (19) The method according to Embodiment 18, wherein the activity includes atrial activity or ventricular activity. (20) The method according to Embodiment 18, wherein the first filter includes a high-pass filter and the second filter includes a low-pass filter.

[0171] (21) The method according to embodiment 20, wherein the high-pass filter is set to 40 Hz. (22) The method according to Embodiment 18, wherein the clinical importance includes the location of the onset of cardiac disease, and the one or more high-frequency energy zones include at least specific values ​​indicating atrial activity. (23) The method according to embodiment 18, wherein the energy threshold is 10%. (24) The method according to Embodiment 18, wherein the method includes applying a weighted vector derived from a ratio. (25) The method according to Embodiment 24, wherein the ratio includes 1:P, where 1 is high energy and P is low energy.

[0172] (26) The method according to embodiment 18, further comprising constructing a training dataset that includes at least the weight vectors. (27) The above method The training algorithm trains an autoencoder using the training dataset, The method according to embodiment 26, further comprising generating an electrocardiogram from one or more output intracardiac signals. (28) A system, Memory for storing processor-executable instructions for the training algorithm, A processor configured to execute the processor-executable instructions of the training algorithm, wherein the processor-executable instructions are configured in the system Applying a first filter to the signal to enhance the activity in the signal and to generate a first modified signal, The rectifier and the second filter are applied to the first modified signal to smooth the region of the first modified signal that is of clinical importance and to generate a second modified signal. A system including a processor that causes the system to automatically detect one or more high-frequency energy zones of the second modified signal using an energy threshold to generate a weight vector. (29) The system according to embodiment 28, wherein the processor is further configured to execute the processor-executable instructions of the training algorithm so that the system constructs a training dataset including at least the weight vectors. (30) The processor is further configured to execute the processor-executable instructions of the training algorithm, and the processor-executable instructions are configured to the system Training an autoencoder using the aforementioned training dataset, The system according to embodiment 29, which generates an electrocardiogram from one or more output intracardiac signals.

Claims

1. It is a system, Memory for storing processor executable code for a noise-reducing autoencoder, The system comprises one or more processors coupled to the memory, wherein the one or more processors are configured to execute processor-executable instructions of a training algorithm and to execute the processor-executable code. The aforementioned processor-executable instruction is A first filter having a high-pass filter is applied to the electrical signal of the heart in such a way as to enhance the activity in the electrical signal of the heart and to generate a first modified signal. The rectifier and the second filter are applied to the first modified signal to smooth the region of the first modified signal that is of clinical importance and to generate a second modified signal. Automatically detect one or more high-frequency energy zones of the second modified signal using an energy threshold to generate a weight vector, The noise reduction autoencoder is trained using a loss function that utilizes the aforementioned weight vector, and the following is performed: The aforementioned processor-executable code, Receiving raw signal data, which is the electrical signal of the heart containing signal noise, The noise reduction autoencoder performs a noise reduction autoencoder operation to encode the raw signal data in such a way that it removes the signal noise. A system that causes the noise-reducing autoencoder to output an output intracardiac signal obtained by decoding the encoded raw signal data so as to generate clean signal data restored without signal noise.

2. The system according to claim 1, wherein the processor generates an electrocardiogram from the output intracardiac signal.

3. The system according to claim 1, wherein the second filter includes a low-pass filter.

4. The system according to claim 1, wherein the clinical importance includes the location of the onset of cardiac disease, and the one or more high-frequency energy zones include at least specific values ​​indicating atrial activity.

5. A method for operating a system comprising a memory for storing processor-executable code for a noise-reducing autoencoder, and one or more processors coupled to the memory, The aforementioned processor, A first filter having a high-pass filter is applied to the electrical signal of the heart in such a way as to enhance the activity in the electrical signal of the heart and to generate a first modified signal. The rectifier and the second filter are applied to the first modified signal to smooth the region of the first modified signal that is of clinical importance and to generate a second modified signal. Automatically detect one or more high-frequency energy zones of the second modified signal using an energy threshold to generate a weight vector, The noise reduction autoencoder is trained using a loss function that utilizes the aforementioned weight vector, Receiving raw signal data, which is the electrical signal of the heart containing signal noise, The noise reduction autoencoder performs a noise reduction autoencoder operation to encode the raw signal data in such a way that it removes the signal noise. An operating method comprising: outputting an output intracardiac signal, which is the encoded raw signal data decoded by the noise-reducing autoencoder, in order to generate clean signal data restored without signal noise.

6. The operating method according to claim 5, wherein the raw signal data includes at least one of power line noise, contact noise, deflection noise, Fluoro noise, and ventricular distal field.

7. The operating method according to claim 5, wherein the noise reduction autoencoder is operated in a mode for constructing a dedicated graphic user interface including one or more filters.

8. The operating method according to claim 5, wherein the noise-reducing autoencoder is operated in a mode for recording noise in a controlled environment and adding the noise to the clean signal data.

9. The operating method according to claim 5, wherein the raw signal data includes N channels of an intracardiac electrocardiogram, an intracardiac electrocardiogram and a surface electrocardiogram, or electrical signals of the heart from an intracardiac electrocardiogram having the position of each electrode and anatomical tissue information.

10. The operating method according to claim 5, wherein the noise-reducing autoencoder is trained based on a loss function that enhances atrial activity.

11. The operating method according to claim 10, wherein the loss function utilizes a neural network in which the final layer restores clean signal data.

12. The operating method according to claim 10, wherein the loss function includes a weighted mean squared error loss function.

13. The operating method according to claim 5, wherein the patient biosensor of the monitoring and processing device records the raw signal data.

14. The operation method according to claim 5, wherein the noise reduction autoencoder operation includes passing the raw signal data through a deep neural network in such a way as to reduce the dimensionality of the raw signal data and to preserve important information.

15. The operating method according to claim 5, wherein the encoded raw signal data includes the reduced number of dimensions from the raw signal data and important information.

16. The operating method according to claim 5, wherein the clean signal data includes a restored input intracardiac signal that minimizes the signal noise.

17. The operating method according to claim 5, wherein the noise-reducing autoencoder learns the clean signal data so as to remove noise from subsequent raw intracardiac signals during the operation of the noise-reducing autoencoder.

18. The operating method according to claim 17, wherein the noise-reducing autoencoder detects the noise type and quality of the subsequent raw intracardiac signal.

19. The operation method according to claim 18, wherein the noise reduction autoencoder performs the noise reduction autoencoder operation on the subsequent raw intracardiac signal based on whether the quality exceeds a threshold.

20. The operating method according to claim 5, further comprising the processor generating an electrocardiogram from the clean signal data, wherein the electrocardiogram substantially does not contain the signal noise.

Images