Automated detection of cardiac structures in cardiac mapping

Abstract

To provide methods, apparatuses, systems and models for automatically detecting a location of specific structures within a heart.SOLUTION: There are provided a system and a method for training a neural network to automatically detect a cardiac structure of an interest object including a processor comprising a neural network training model that receives training data. The training data comprises first input comprising first electrophysiological data regarding a first cardiac structure received by an electrode of a first catheter positioned within a heart, and second input comprising second data relating to the first cardiac structure. The neural network training model generates, as output, determination of whether the first cardiac structure is the cardiac structure of the interest object on the basis of the training data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to artificial intelligence and machine learning related to automatically detecting the location of specific structures within the heart, and more particularly, to automatically detecting the location of the His bundle within the heart. [Background technology]

[0002] It is well known to use ablation catheters to cause tissue necrosis in cardiac tissue and correct cardiac arrhythmias (including, but not limited to, atrial fibrillation, atrial flutter, atrial tachycardia, and ventricular tachycardia). Arrhythmias can result in a variety of dangerous conditions, including irregular heart rate, loss of synchronous atrioventricular contractions, and congestion of blood flow, which can lead to a variety of illnesses and even death. Stray electrical signals within one or more heart chambers are thought to be the primary cause of many arrhythmias.

[0003] During cardiac ablation, lesions are created in the tissue of a patient's heart by inserting a catheter into the heart so that the catheter contacts the tissue and electromagnetic radiofrequency (RF) energy is injected into the tissue from the catheter electrodes to effect ablation and lesion creation.

[0004] The bundle of His, otherwise known as the bundle of His, is a portion of the myocardium that originates near the orifice of the coronary sinus (CS). The bundle of His is an important part of the cardiac electrical conduction system in that it is responsible for transmitting electrical impulses from the atrioventricular (AV) node, located between the atria and ventricles, to the ventricles of the heart.

[0005] The His bundle is located in a vulnerable position within the heart, and if incorrectly ablated, it can cause harmful and undesirable effects on the cardiac electrical conduction system. During conventional ablation procedures, physicians often manually tag the His bundle to identify its location within the heart so that the His bundle can be avoided during the ablation procedure. Such manual tagging of the His bundle is tedious and time-consuming. In addition, manual tagging can result in false-positive readings, in which an electrocardiogram (ECG) signal appears to look like a His bundle impulse, but in fact is not. Summary of the Invention [Problem to be solved by the invention]

[0006] There is a need for an automated and reliable system and method that utilizes artificial intelligence and / or machine learning to accurately and automatically detect the His bundle. [Means for solving the problem]

[0007] SUMMARY Described herein are methods, devices, systems, and models for automatically detecting the location of specific structures within the heart.

[0008] According to one aspect, the subject matter disclosed herein relates to a system for automatically detecting a cardiac structure. The system preferably includes a first catheter positioned within a heart to receive electrophysiological data related to a first cardiac structure, a second catheter positioned at a predetermined location within the heart, and a processor including a neural network. The neural network receives the electrophysiological data from the first catheter, receives distance data related to a distance between the first and second catheters, determines whether the electrophysiological data related to the first cardiac structure matches predetermined electrophysiological data of a cardiac structure of interest, determines whether the distance between the first and second catheters is less than a predetermined threshold, and determines whether the first cardiac structure is a cardiac structure of interest based on the electrophysiological data and the distance data.

[0009] According to another aspect, the subject matter disclosed herein relates to a system for training a neural network to automatically detect cardiac structures. The system includes a processor including a neural network training model that receives training data. The training data includes a previously mapped location of a cardiac structure of interest, electrophysiological data for the first cardiac structure received by a first catheter positioned within the heart, predetermined electrophysiological data for the cardiac structure of interest, distance data for a distance between the first catheter and a second catheter, and a predetermined threshold for a distance between a point on the first catheter and a point on the second catheter. The neural network training model is trained to determine whether the electrophysiological data is consistent with the predetermined electrophysiological data for the cardiac structure of interest, determine whether the distance data is less than a predetermined value, and determine whether the first cardiac structure is the cardiac structure of interest based on the training data.

[0010] According to yet another aspect, the subject matter disclosed herein relates to a method for training a neural network model to automatically detect cardiac structures. The method includes receiving training data by a processor including a neural network model. The training data includes a previously mapped location of a cardiac structure of interest, electrophysiological data for the first cardiac structure received by a first catheter positioned within the heart, predetermined electrophysiological data for the cardiac structure of interest, distance data for a distance between the first catheter and a second catheter, and a predetermined threshold for a distance between a point on the first catheter and a point on the second catheter. The method further includes training the neural network model using the training data. The training includes determining whether the electrophysiological data is consistent with the predetermined electrophysiological data for the cardiac structure of interest, determining whether the distance data is less than a predetermined value, and determining whether the first cardiac structure is the cardiac structure of interest based on the training data.

[0011] According to yet another aspect, the cardiac structure of interest is the bundle of His.

[0012] According to yet another aspect, the first catheter includes a His bundle mapping catheter.

[0013] According to yet another aspect, the electrophysiological data related to the first cardiac structure received by the first catheter includes an electrogram, more specifically, a His bundle electrogram.

[0014] According to yet another aspect, the second catheter is a coronary sinus catheter, and more particularly, includes a position sensor.

[0015] According to yet another aspect, the training data further includes electrocardiogram data generated by surface body electrodes.

[0016] According to yet another aspect, a neural network is trained to determine whether the electrocardiogram data is consistent with predetermined electrophysiological data of the cardiac structure of interest.

[0017] According to yet another aspect, the neural network is trained to determine that a first cardiac structure is a cardiac structure of interest when electrophysiological data for the first cardiac structure matches predetermined electrophysiological data for the cardiac structure of interest and the distance data is less than a predetermined value.

[0018] According to yet another aspect, the neural network is trained to determine that the first cardiac structure is not the cardiac structure of interest when the electrophysiological data for the first cardiac structure does not match the predetermined electrophysiological data for the cardiac structure of interest or when the distance data is greater than a predetermined value.

[0019] According to yet another aspect, the predetermined electrophysiological data of the cardiac structure of interest is stored in a database in communication with the neural network.

[0020] According to yet another aspect, the neural network is a convolutional neural network or a long-short-term memory neural network.

[0021] According to yet another aspect, the neural network training model determines that the first cardiac structure is a cardiac structure of interest, and this determination is compared to a database of locations of known cardiac structures to verify the accuracy of the determination. If the accuracy of the determination exceeds a predetermined accuracy threshold, the neural network training model is verified as the standard for cardiac mapping systems.

[0022] According to yet another aspect, the subject matter disclosed herein relates to a system for training a neural network to automatically detect a cardiac structure of interest, including a processor including a neural network training model that receives training data. The training data includes a first input including first electrophysiological data related to the first cardiac structure received by an electrode of a first catheter positioned within the heart and a second input including second data related to the first cardiac structure. The neural network training model generates as an output, based on the training data, a determination of whether the first cardiac structure is the cardiac structure of interest. [Brief explanation of the drawings]

[0023] A more detailed understanding may be had from the following description, given by way of example in conjunction with the accompanying drawings, in which like reference numerals indicate similar elements and in which: [Figure 1] 1 is a block diagram of an exemplary system for remotely monitoring and communicating patient biometric indicators in accordance with the subject matter of the present application. [Figure 2] 1 is a system diagram of an example computing environment in communication with a network in accordance with the subject matter of the present application. [Figure 3] 1 is a block diagram of an example device capable of implementing one or more features of the present disclosure in accordance with the subject matter of the present application. [Figure 4] 4 shows a graphical depiction of an artificial intelligence system incorporating the exemplary device of FIG. 3 in accordance with the subject matter of the present application. [Figure 5] 5 illustrates a method implemented in the artificial intelligence system of FIG. 4 according to the subject matter of the present application. [Figure 6] 1 illustrates an example of a Naive Bayes calculation probability according to the subject matter of the present application. [Figure 7] 1 illustrates an exemplary decision tree in accordance with the subject matter of the present application. [Figure 8] 1 illustrates an exemplary random forest classifier in accordance with the subject matter of the present application. [Figure 9] 1 illustrates an exemplary logistic regression according to the subject matter of the present application. [Figure 10] 1 illustrates an exemplary support vector machine according to the subject matter of the present application. [Figure 11] 1 illustrates an exemplary linear regression model in accordance with the subject matter of the present application. [Figure 12] 1 illustrates an exemplary K-means clustering according to the subject matter of the present application. [Figure 13] 1 illustrates an exemplary ensemble learning algorithm in accordance with the subject matter of the present application. [Figure 14] 1 illustrates an exemplary neural network in accordance with the subject matter of the present application. [Figure 15] 1 illustrates a hardware-based neural network in accordance with the subject matter of the present application. [Figure 16] 1 illustrates electrocardiogram (ECG) signals produced by the contraction (depolarization) and relaxation (repolarization) of the atrial and ventricular muscles of the heart, in accordance with the subject matter of the present application. [Figure 17] 1 illustrates an exemplary cardiac ablation system in accordance with the subject matter of the present application that may implement one or more features of the disclosed subject matter. [Figure 18A] A neural network, such as the neural network of FIG. 17, is shown receiving input data to train the neural network to automatically identify cardiac structures of interest, such as the His bundle, with greater efficiency and reliability than manual identification by a physician during a procedure, such as an ablation procedure. [Figure 18B] FIG. 1 is a flow diagram illustrating an embodiment of a module for training a neural network in accordance with the subject matter of the present application. [Figure 19A] 1A-1C illustrate cardiac images obtained by fluoroscopy showing various catheters positioned within the heart during a cardiac ablation procedure in accordance with the subject matter of the present application. [Figure 19B] 1 illustrates exemplary first and second catheters positioned within the heart in accordance with the subject matter of the present application. [Figure 20] 1 illustrates exemplary ECG and His bundle electrogram (HBE) signals that can be used in accordance with the disclosed systems and methods, in accordance with the subject matter of the present application. [Figure 21] 1 illustrates an exemplary convolutional neural network (CNN) in accordance with the subject matter of the present application. [Figure 22] 1 illustrates an exemplary recurrent neural network (RNN) in accordance with the subject matter of the present application. [Figure 23] 1 illustrates one implementation of the system as described. [Figure 24] 1 illustrates one implementation of the system as described. [Figure 25] 1 illustrates one implementation of the system as described. DETAILED DESCRIPTION OF THE INVENTION

[0024] Methods, systems, and programs are provided that provide for automatically detecting the location of specific structures within the heart, more preferably automatically detecting the location of the His bundle within the heart, and for training a neural network to automatically detect the location of such specific structures within the heart.

[0025] 1 is a block diagram of an exemplary system 100 for remotely monitoring and communicating patient biometrics (i.e., patient data). In the example shown in FIG. 1, the system 100 includes a patient biometric monitoring and processing unit 102 associated with a patient 104, a local computing device 106, a remote computing system 108, a first network 110, and a second network 120.

[0026] According to one exemplary embodiment, the monitoring and processing device 102 may be a device internal to the patient's body (e.g., subcutaneously implantable). The monitoring and processing device 102 may be inserted into the patient via any applicable method, including oral injection, surgical insertion via a vein or artery, endoscopic procedure, or laparoscopic procedure.

[0027] According to an exemplary embodiment, the monitoring and processing device 102 may be a device external to the patient. For example, as described in more detail below, the monitoring and processing device 102 may include an attachable patch (e.g., attached to the patient's skin). The monitoring and processing device 102 may also include a catheter with one or more electrodes, a probe, a blood pressure cuff, a weight scale, a bracelet or smartwatch biometric tracker, a glucose monitor, a continuous positive airway pressure (CPAP) machine, or virtually any device that can provide input regarding the patient's health or biometrics.

[0028] According to one exemplary embodiment, the monitoring and processing unit 102 may include both components internal to the patient and components external to the patient.

[0029] A single monitoring and processing device 102 is shown in Figure 1. However, an exemplary system may include multiple patient biometric monitoring and processing devices. A patient biometric monitoring and processing device may be in communication with one or more other patient biometric monitoring and processing devices. Additionally or alternatively, a patient biometric monitoring and processing device may be in communication with a network 110.

[0030] One or more monitoring and processing devices 102 may acquire patient biometric data (e.g., electrical signals, blood pressure, body temperature, blood glucose levels, or other biometric data) and may receive at least a portion of the patient biometric data representing the acquired patient biometric indicators, as well as additional information associated with the acquired patient biometric indicators from one or more other monitoring and processing devices 102. The additional information may be, for example, diagnostic information and / or additional information obtained from additional devices, such as wearable devices. Each monitoring and processing device 102 may process data including its own acquired patient biometric indicators as well as data received from one or more other monitoring and processing devices 102.

[0031] 1, network 110 is an example of a short-range network (e.g., a local area network (LAN) or a personal area network (PAN)). Information may be transmitted over short-range network 110 between monitoring and processing equipment 102 and local computing device 106 using any one of a variety of short-range wireless communication protocols, such as Bluetooth, Wi-Fi, Zigbee, Z-Wave, near field communication (NFC), Ultraband, Zigbee, or infrared (IR).

[0032] In an exemplary embodiment, network 120 may be a wired network, a wireless network, or may include one or more wired and wireless networks. For example, network 120 may be a long-range network (e.g., a wide area network (WAN), the Internet, or a cellular network). Information may be transmitted over network 120 using any one of a variety of long-range wireless communication protocols (e.g., TCP / IP, HTTP, 3G, 4G / LTE, or 5G / New Radio).

[0033] In one exemplary embodiment, the patient monitoring and processing device 102 may include patient biometric sensors 112, a processor 114, user input (UI) sensors 116, memory 118, and a transmitter-receiver (i.e., transceiver) 122. The patient monitoring and processing device 102 may continuously or periodically monitor, store, process, and communicate any number of various patient biometric indicators over the network 110. Examples of patient biometric indicators include electrical signals (e.g., ECG signals and brain biometrics), blood pressure data, blood glucose data, and temperature data. The patient biometric indicators may be monitored and communicated to treat any number of various diseases, such as cardiovascular diseases (e.g., arrhythmias, cardiomyopathies, and coronary artery disease) and autoimmune diseases (e.g., type 1 and type 2 diabetes).

[0034] In one embodiment, the patient biometric sensor 112 may include, for example, one or more sensors configured to sense a type of biometric patient biometric indicator. For example, the patient biometric sensor 112 may include electrodes configured to acquire electrical signals (e.g., cardiac signals, brain signals, or other bioelectrical signals), a temperature sensor, a blood pressure sensor, a blood glucose sensor, a blood oxygen sensor, a pH sensor, an accelerometer, and a microphone).

[0035] In one exemplary embodiment, as described in more detail below, the patient biometric monitoring and processing device 102 may be an ECG monitor for monitoring cardiac ECG signals. The patient biometric sensor 112 of the ECG monitor may include one or more electrodes for acquiring the ECG signals. The ECG signals may be used in the treatment of various cardiovascular disorders.

[0036] In one exemplary embodiment, the transceiver 122 may include a separate transmitter and receiver, or alternatively, the transceiver 122 may include a transmitter and receiver integrated into a single device.

[0037] In one exemplary embodiment, processor 114 may be configured to store patient data, such as patient biometric data acquired by patient biometric sensors 112, in memory 118 and to communicate the patient data over network 110 via a transmitter in transceiver 122. Data from one or more other monitoring and processing devices 102 may also be received by a receiver in transceiver 122, as described in more detail below.

[0038] According to one exemplary embodiment, the monitoring and processing device 102 includes a UI sensor 116, which may be a piezoelectric or capacitive sensor configured to receive user input, such as a tap or touch. For example, the UI sensor 116 may be controlled to implement capacitive coupling in response to the patient 104 tapping or touching the surface of the monitoring and processing device 102. Gesture recognition may be implemented via any one of a variety of capacitive types, such as resistive-capacitive, surface-capacitive, projected-capacitive, surface acoustic wave, piezoelectric, and infrared touch. The capacitive sensor may be positioned over a small area or length of the surface such that a tap or touch on the surface activates the monitoring device.

[0039] As described in more detail below, processor 114 may be configured to selectively respond to different tapping patterns (e.g., single tap or double tap) of a capacitive sensor, which may be UI sensor 116, so that different tasks of the patch (e.g., data acquisition, storage, or transmission) may be initiated based on the detected pattern. In some embodiments, audible feedback may be provided to the user from processing unit 102 when a gesture is detected.

[0040] In an exemplary embodiment, the local computing device 106 of the system 100 may be configured to communicate with the patient biometric monitoring and processing device 102 and act as a gateway to the remote computing system 108 via the second network 120. The local computing device 106 may be, for example, a smartphone, smartwatch, tablet, or other portable smart device configured to communicate with other devices via the network 120. Alternatively, the local computing device 106 may be a fixed or standalone device, such as, for example, a fixed base station including modem and / or router capabilities, a desktop or laptop computer using an executable program to communicate information between the processing device 102 and the remote computing system 108 via a wireless module in a PC, or a USB dongle. Patient biometric indicators may be communicated between the local computing device 106 and the patient biometric monitoring and processing device 102 via a short-range wireless network 110, such as a local area network (LAN) (e.g., a personal area network (PAN)), using short-range wireless technology standards (e.g., Bluetooth, Wi-Fi, ZigBee, Z-Wave, and other short-range wireless standards). In some embodiments, the local computing device 106 may also be configured to display the acquired patient electrical signals and information related to the acquired patient electrical signals, as described in more detail below.

[0041] In some demonstrative embodiments, remote computing system 108 may be configured to receive at least one of the monitored patient's biometric indicators and information associated with the monitored patient over network 120, which is a long-range network. For example, if local computing device 106 is a cellular phone, network 120 may be a wireless cellular network, and information may be communicated between local computing device 106 and remote computing system 108 via a wireless technology standard, such as any of the wireless technologies described above. As described in more detail below, remote computing system 108 may be configured to provide (e.g., visually display and / or audibly provide) at least one of the patient's biometric indicators and associated information to a medical professional (e.g., a physician).

[0042] 2 is a system diagram of an example computing environment 200 in communication with network 120. In some examples, computing environment 200 is incorporated into a public cloud computing platform (such as Amazon Web Services or Microsoft Azure), a hybrid cloud computing platform (such as HP Enterprise OneSphere), or a private cloud computing platform.

[0043] As shown in FIG. 2, computing environment 200 preferably includes a remote computing system 108 (hereinafter, computer system), which is an example of a computing system on which embodiments described herein may be implemented.

[0044] The remote computing system 108 can perform various functions via the processor 220, which may include one or more processors. For example, functions may include analyzing monitored patient biometrics and related information and providing alerts, additional information, or instructions (e.g., via the display 266) according to physician-determined or algorithm-driven thresholds and parameters. As described in more detail below, the remote computing system 108 can be used to provide a patient information dashboard (e.g., via the display 266) to a medical professional (e.g., a physician) so that the patient information may enable the medical professional to identify and prioritize patients with more significant needs than others.

[0045] 2, computer system 210 may include a communication mechanism, such as a bus 221, or other communication mechanism for communicating information within computer system 210. Computer system 210 further includes one or more processors 220 coupled with bus 221 for processing information. Processor 220 may include one or more CPUs, GPUs, or any other processors known in the art.

[0046] Computer system 210 may also include a system memory 230 coupled to bus 221 for storing information and instructions executed by processor 220. System memory 230 may include computer-readable storage media in the form of volatile and / or nonvolatile memory, such as read-only system memory (ROM) 231 and / or random access memory (RAM) 232. System memory RAM 232 may also include other dynamic storage device(s) (e.g., dynamic RAM, static RAM, and synchronous DRAM). System memory ROM 231 may also include other static storage device(s) (e.g., programmable ROM, erasable PROM, and electrically erasable PROM). Additionally, system memory 230 may be used to store temporary variables or other intermediate information during execution of instructions by processor 220. A basic input / output system (BIOS) 233 may include routines for transferring information, which may be stored in system memory ROM 231, between elements within computer system 210, such as during start-up. RAM 232 may contain data and / or program modules that are immediately accessible to and / or presently being operated on by processor 220. System memory 230 may also include, for example, an operating system 234, application programs 235, other program modules 236, and program data 237.

[0047] In one exemplary embodiment, computer system 210 also includes a disk controller 240 coupled to bus 221 for controlling one or more storage devices for storing information and instructions, such as a magnetic hard disk 241 and a removable media drive 242 (e.g., a floppy disk drive, a compact disk drive, a tape drive, and / or a solid state drive). Storage devices may be added to computer system 210 using an appropriate device interface (e.g., small computer system interface (SCSI), integrated device electronics (IDE), Universal Serial Bus (USB), or FireWire).

[0048] Computer system 210 may also include a display controller 265 coupled to bus 221 to control a monitor or display 266, such as a cathode ray tube (CRT) or liquid crystal display (LCD), for displaying information to a computer user. The illustrated computer system 210 includes a user input interface 260 and one or more input devices, such as a keyboard 262 and a pointing device 261, for interacting with a computer user and providing information to processor 220. Pointing device 261 may be, for example, a mouse, trackball, or pointing stick for communicating directional information and command selections to processor 220 and for controlling cursor movement on display 266. Display 266 may provide a touchscreen interface that may enable input that complements or replaces the communication of directional information and command selections by pointing device 261 and / or keyboard 262.

[0049] Computer system 210 may perform some or each of the functions and methods described herein in response to processor 220 executing one or more sequences of one or more instructions contained in a memory, such as system memory 230. Such instructions may be read into system memory 230 from another computer-readable medium, such as hard disk 241 or removable media drive 242. Hard disk 241 may include one or more data stores and data files used by the embodiments described herein. Data store contents and data files may be encrypted for improved security. Processor 220 may also be employed in a multi-processing configuration to execute one or more sequences of instructions contained in system memory 230. In alternative embodiments, hardwired circuitry may be used in place of or in combination with software instructions. Thus, embodiments are not limited to any specific combination of hardware circuitry and software.

[0050] As mentioned above, computer system 210 may include at least one computer-readable medium or memory for retaining programmed instructions according to the embodiments described herein and for containing the data structures, tables, records, or other data described herein. As used herein, the term “computer-readable medium” refers to any non-transitory, tangible medium that participates in providing instructions to processor 220 for execution. Computer-readable media may take many forms, including, but not limited to, non-volatile media, volatile media, and transmission media. Non-limiting examples of non-volatile media include optical disks, solid-state drives, magnetic disks, and magneto-optical disks, such as hard disk 241 or removable media drive 242. Non-limiting examples of volatile media include dynamic memory, such as system memory 230. Non-limiting examples of transmission media include coaxial cables, copper wire, and fiber optics, including the wires that comprise bus 221. Transmission media can also take the form of acoustic or light waves, such as those generated during radio wave and infrared data communications.

[0051] The computing environment 200 may further include a computer system 210 operating in a networked environment using logical connections to the local computing device 106 and to one or more other devices, such as a personal computer (laptop or desktop), a mobile device (e.g., a patient mobile device), a server, a router, a network PC, a peer device, or other common network node, and typically includes many or all of the elements described above with respect to the computer system 210. When used in a networked environment, the computer system 210 may include a modem 272 for establishing communications over a network 120, such as the Internet. The modem 272 may be connected to the system bus 221 via a network interface 270 or another appropriate mechanism.

[0052] Network 120 as shown in FIGS. 1 and 2 may be any network or system commonly known in the art, including the Internet, an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a direct connection or series of connections, a cellular telephone network, or any other network or medium capable of facilitating communication between computer system 610 and other computers (e.g., local computing device 106).

[0053] 3 is a block diagram of an example device 300 capable of implementing one or more features of the present disclosure. The device 300 may be, for example, the local computing device 106. The device 300 may include, for example, a computer, a gaming device, a handheld device, a set-top box, a television, a mobile phone, or a tablet computer. The device 300 includes a processor 302, a memory 304, a storage device 306, one or more input devices 308, and one or more output devices 310. The device 300 may also optionally include an input driver 312 and an output driver 314. It is understood that the device 300 may include additional components not shown in FIG. 3 , including an artificial intelligence accelerator.

[0054] In various alternatives, processor 302 includes a central processing unit (CPU), a graphics processing unit (GPU), a CPU and a GPU located on the same die, or one or more processor cores, each of which may be a CPU or a GPU. In various alternatives, memory 304 is located on the same die as processor 302 or is located separate from processor 302. Memory 304 includes volatile or non-volatile memory, such as random access memory (RAM), dynamic RAM, or a cache.

[0055] The storage device 306 includes fixed or removable storage means, such as a hard disk drive, solid state drive, optical disk, or flash drive. The input device 308 includes, but is not limited to, a keyboard, keypad, touch screen, touchpad, detector, microphone, accelerometer, gyroscope, biometric scanner, or network connection (e.g., a wireless local area network card for transmitting and / or receiving wireless IEEE 802 signals). The output device 310 includes, but is not limited to, a display, speaker, printer, haptic feedback device, one or more lights, antenna, or network connection (e.g., a wireless local area network card for transmitting and / or receiving wireless IEEE 802 signals).

[0056] The input driver 312 communicates with the processor 302 and the input device 308, allowing the processor 302 to receive input from the input device 308. The output driver 314 communicates with the processor 302 and the output device 310, allowing the processor 302 to send output to the output device 310. Note that the input driver 312 and the output driver 314 are optional components; the device 300 would operate in the same manner if the input driver 312 and the output driver 314 were not present. The output driver 314 may include an accelerated processing device ("APD") 316 coupled to a display device 318. The APD accepts computational and graphic rendering commands from the processor 302, processes the computational and graphic rendering commands, and provides pixel output to the display device 318 for display. As described in further detail below, APD 316 includes one or more parallel processing units that perform calculations according to the single-instruction-multiple-data ("SIMD") paradigm. Accordingly, although various functions are described herein as being performed by APD 316 or in conjunction with APD 316, in various alternatives, functions described as being performed by APD 316 are additionally or alternatively performed by other computing devices that are not driven by a host processor (e.g., processor 302) and have similar capabilities to provide graphical output to display device 318. For example, it is contemplated that any processing system that performs processing tasks according to the SIMD paradigm can perform the functions described herein. Alternatively, computing systems that do not perform processing tasks according to the SIMD paradigm are contemplated to perform the functions described herein.

[0057] FIG. 4 shows a functional graphical depiction of an artificial intelligence system 400 incorporating the exemplary device of FIG. 3 . The system 400 includes data 410, a machine 420, a model 430, multiple predicted outcomes 440, and underlying hardware 450. The system 400 operates by training the machine 420 using the data 410 while building a model 430 that enables the multiple outcomes 440 to be predicted. The system 400 may operate on the hardware 450. In such a configuration, the data 410 may be associated with the hardware 450, e.g., originating from the monitoring and processing unit 102. For example, the data 410 may be ongoing data or output data associated with the hardware 450. The machine 420 may operate as or be associated with a controller or data collection associated with the hardware 450. The model 430 may be configured to model the operation of the hardware 450 as well as model the data 410 collected from the hardware 450 to predict outcomes achieved by the hardware 450. The hardware 450 may be configured to use the predicted outcomes 440 to provide a particular desired outcome 440 from the hardware 450 .

[0058] Figure 5 illustrates a general method 500 implemented in the artificial intelligence system of Figure 4. Method 500 includes collecting data from hardware at step 510. This data may include currently collected historical data or other data from the hardware, or various combinations thereof. For example, this data may include measurements taken during a surgical procedure and may be correlated with the outcome of the procedure. For example, cardiac temperature may be collected and correlated with the outcome of a cardiac procedure.

[0059] At step 520, method 500 includes training the machine on the hardware. Training may include analyzing and correlating the data collected at step 510. For example, in the cardiac case, temperature and outcome data may be trained to determine if a correlation or association exists between cardiac temperature during treatment and outcome.

[0060] Method 500 includes building a model based on the hardware-related data at step 530. Building the model may include physical hardware or software modeling, algorithmic modeling, etc., as described below. The modeling may aim to represent the collected and trained data.

[0061] Method 500 includes predicting an outcome for a hardware-related model at step 540. This outcome prediction may be based on a trained model. For example, for the heart, if a procedure results in a positive outcome when the temperature during the procedure is between 97.7 and 100.2, then for a given procedure, the outcome may be predicted based on the temperature of the heart during the procedure. This model is rudimentary and is provided for illustrative purposes to facilitate understanding of the present disclosure.

[0062] The present systems and methods operate to train machines, build models, and predict outcomes using algorithms. These algorithms may be used to solve the trained models and predict outcomes related to the hardware. These algorithms can generally be categorized into classification algorithms, regression algorithms, and clustering algorithms.

[0063] For example, classification algorithms are used in situations where the dependent variable to be predicted is divided into multiple classes and one class, i.e., the dependent variable, is predicted for a given input. Thus, classification algorithms are used to predict outcomes from a number of fixed, predefined outcomes. Classification algorithms may include naive Bayes algorithms, decision trees, random forest classifiers, logistic regression, support vector machines, and k-nearest neighbors.

[0064] In general, the Naive Bayes algorithm follows Bayes' Theorem and follows a probabilistic approach, it being understood that other probability-based algorithms may be used and generally operate using similar principles of probability theory as those described below for the exemplary Naive Bayes algorithm.

[0065] Figure 6 shows an example of a naive Bayes calculation of probability. The probabilistic approach of Bayes' theorem essentially means that instead of jumping directly to the data, the algorithm has a set of prior probabilities for each target class. After the data is input, the naive Bayes algorithm may update the prior probabilities to form posterior probabilities. This is done using the following formula:

[0066]

number

[0067] The Naive Bayes algorithm and the Bayes algorithm in general can be useful when one needs to predict whether an input belongs to a given list of n classes or not. Because the probability of all n classes is very low, a probabilistic approach can be used.

[0068] For example, as shown in FIG. 6, a person's decision to play golf depends on factors including, but not limited to, the weather outside, as shown in a first data set 610. The first data set 610 lists the weather in a first column and the play outcomes associated with that weather in a second column. A frequency table 620 generates the frequency with which a particular event occurs. The frequency table 620 determines how often a person will or will not play golf in each weather condition. From this, a likelihood table is compiled and initial probabilities are generated. For example, the probability that the weather is cloudy is 0.29, but the general probability of playing is 0.64.

[0069] Posterior probabilities may be generated from likelihood table 630. These posterior probabilities may be configured to answer questions about weather conditions and whether golf will be played in those weather conditions. For example, the probability that it is sunny outside and golf will be played can be calculated using the Bayesian formula: P(Yes|Sunny) = P(Sunny|Yes) × P(Yes) / P(Sunny) It may be represented by: According to likelihood table 630, P(sunny|yes) = 3 / 9 = 0.33, P(sunny)=5 / 14=0.36, P(yes) = 9 / 14 = 0.64 is. Therefore, P(yes|sunny) = .33 × .64 / .36 or approximately 0.60 (60%).

[0070] Generally, a decision tree is a tree structure similar to a flowchart, where each outer node represents a test of an attribute and each branch represents the result of that test. The leaf nodes contain the actual predicted label. A decision tree starts at the root of the tree and attribute values are compared until a leaf node is reached. Decision trees can be used as classifiers when dealing with high-dimensional data and when little time is spent on data preparation. Decision trees can take the form of simple decision trees, linear decision trees, algebraic decision trees, deterministic decision trees, randomized decision trees, non-deterministic decision trees, and quantum decision trees. An exemplary decision tree is shown below in Figure 7.

[0071] 7 shows a decision tree that follows the same structure as the Bayesian example above when deciding whether to play golf. In the decision tree, a first node 710 examines whether the weather is sunny 712, cloudy 714, and rainy 716 as options for progressing down the decision tree. If the weather is sunny, the tree branch continues to a second node 720 that examines the temperature. In this example, the temperature at node 720 can be high 722 or normal 724. If the temperature at node 720 is high 722, a predicted outcome of "No Golf" 723 occurs. If the temperature at node 720 is normal 724, a predicted outcome of "Yes Golf" 725 occurs.

[0072] Furthermore, from the first node 710, the results Cloudy 714, Golf "Yes" 715 are generated.

[0073] From the first node, weather 710, a third node 730 (again) checks the temperature as a result of rain 716. If the temperature is normal 732 at the third node 730, the answer is "yes" 733 to play golf. If the temperature is low 734 at the third node 730, the answer is "no" 735 to not play golf.

[0074] From this decision tree, a golfer will play golf when it is cloudy 715, when it is sunny with normal temperatures 725, and when it is raining with normal temperatures 733, but the golfer will not play when it is sunny with high temperatures 723 or when it is raining with low temperatures 735.

[0075] A random forest classifier is a committee of decision trees, where each tree is given a subset of the data's attributes and makes a prediction based on that subset. The mode of the decision trees' actual predictions is taken into account to provide the final random forest answer. Random forest classifiers generally mitigate the overfitting that exists in standalone decision trees, making them more robust and accurate classifiers.

[0076] FIG. 8 illustrates an exemplary random forest classifier for classifying clothing color. As shown in FIG. 8, the random forest classifier includes five decision trees 8101, 8102, 8103, 8104, and 8105 (collectively or generally referred to as decision tree 810). Each of the trees is designed to classify clothing color. Because the individual trees generally operate as the decision trees of FIG. 7, a discussion of each of the trees and the decisions made is not provided. In this illustration, three of the five trees (8101, 8102, 8104) determine that the clothing is blue, one determines that the clothing is green (8103), and the remaining tree determines that the clothing is red (8105). The random forest receives these actual predictions of the five trees and calculates the mode of the actual predictions to provide the random forest's answer that the clothing is blue.

[0077] Logistic regression is another algorithm for binary classification tasks. It is based on the logistic function, also known as the sigmoid function. This S-shaped curve can take any real-valued value and map it between 0 and 1, asymptotically approaching these bounds. Logistic models can be used to model the probability of a particular class or event being present, such as pass / fail, win / lose, alive / dead, or healthy / sick. This can be extended to model several classes of events, such as determining whether an image contains a cat, dog, lion, etc. Each object detected in an image is assigned a probability between 0 and 1, and these probabilities sum to 1.

[0078] In a logistic model, the log odds (log of the odds) of a value labeled "1" are a linear combination of one or more independent variables ("predictors"), each of which can be binary (two classes coded by indicator variables) or continuous (any real-valued value). The corresponding probability of a value labeled "1" is so labeled because it can vary between 0 (definitely a value of "0") and 1 (definitely a value of "1"). The logistic function is so named because it converts log odds into probabilities. The unit of measurement for the log odds scale is called a logit, another name for the logistic unit. Similar models, such as probit models, can also be used, but with a sigmoid function instead of the logistic function. A distinctive feature of the logistic model is that increasing one of the independent variables multiplicatively scales the odds of a given outcome by a constant factor, and each independent variable has its own parameters. For binary dependent variables, this generalizes the odds ratio.

[0079] In a binary logistic regression model, the dependent variable has two levels (categories). Outputs with more than two values are modeled by multinomial logistic regression, and if the categories are ordered, they are modeled by ordinal logistic regression (e.g., proportional odds ordinal logistic model). Although a logistic regression model itself simply models the probability of the output with respect to the inputs and does not perform statistical classification (it is not a classifier), it can be used to create a classifier, for example, by selecting a cutoff value and classifying inputs with probabilities greater than the cutoff as one class and inputs with probabilities less than the cutoff as the other class. This is a common way to create a binary classifier.

[0080] FIG. 9 illustrates an exemplary logistic regression. This exemplary logistic regression allows for the prediction of an outcome based on a set of variables. For example, a school's acceptance outcome can be predicted based on an individual's grade point average. The relationship between past history of grade point average and acceptance allows for predictions. The logistic regression of FIG. 9 allows analysis of the grade point average variable 920 to predict an outcome 910 defined between 0 and 1. At the lower end of the S-curve 930, the grade point average 920 predicts an outcome 910 of not being accepted. At the upper end of the S-curve 940, the grade point average 920 predicts an outcome 910 of being accepted. Logistic regression can be used to predict home values, customer lifetime value in the insurance sector, etc.

[0081] A support vector machine (SVM) can be used to sort the data by making the margin between the two classes as far apart as possible. This is called margin-maximizing separation. Unlike linear regression, which uses the entire dataset for its purpose, SVM is able to take support vectors into account while plotting the hyperplane.

[0082] FIG. 10 illustrates an exemplary support vector machine. In the exemplary SVM 1000, data can be classified into two distinct classes, represented as squares 1010 and triangles 1020. The SVM 1000 operates by drawing a random hyperplane 1030. This hyperplane 1030 is monitored by comparing the distance (shown by line 1040) between the hyperplane 1030 and the nearest data points 1050 from each class. The data points 1050 closest to the hyperplane 1030 are known as support vectors. The hyperplane 1030 is drawn based on these support vectors 1050, with the optimal hyperplane having the greatest distance from each support vector 1050. The distance between the hyperplane 1030 and the support vectors 1050 is known as the margin.

[0083] The SVM 1000 may be used for data classification by using the hyperplane 1030 to maximize the distance between the hyperplane 1030 and the support vectors 1050. Such an SVM 1000 may be used, for example, to predict heart disease.

[0084] k Nearest Neighbors (KNN) generally refers to a set of algorithms that make no assumptions about the underlying data distribution and perform reasonably short training phases. Generally, KNN uses a large number of data points divided into classes to predict the classification of a new sample point. Operationally, KNN specifies an integer N with the new sample. The N entries in the model of the system that are closest to the new sample are selected. The most common classification of these entries is determined, and that classification is assigned to the new sample. KNN generally requires increasing storage space as the training set grows. This also means that estimation time increases linearly with the number of training points.

[0085] In regression algorithms, the output is continuous, so regression algorithms can be used when the target variable is a continuous variable. Linear regression is a common example of a regression algorithm. Linear regression can be used to measure true quality (such as housing costs, number of calls, or total transactions) by considering consistent variables. The connection between the variables and the outcome is found by fitting a line of best fit (hence the name linear regression). This line of best fit is known as the regression line and is expressed directly in terms Y=a×X+b. Linear regression is most often used in low-dimensional approaches.

[0086] 11 shows an exemplary linear regression model, in which a predictor variable 1110 is modeled against a measurement variable 1120. Clusters of instances of the predictor variable 1110 and the measurement variable 1120 are plotted as data points 1130. The data points 1130 are then fitted to a best-fit line 1140. Subsequent predictions then use the best-fit line 1140 to predict that instance of the predictor variable 1110, given the measurement variable 1120. Linear regression can be used to model and predict outcomes in surgical procedures, financial portfolio performance, salary predictions, real estate, and estimated arrival times in transportation.

[0087] Clustering algorithms may be used to model and train datasets. In clustering, inputs are assigned to two or more clusters based on feature similarity. Clustering algorithms typically learn patterns and useful insights from data without guidance. For example, unsupervised learning algorithms such as K-means clustering can be used to cluster audiences into similar groups based on interests, age, geography, etc.

[0088] K-means clustering is generally considered a simple unsupervised learning approach. In K-means clustering, similar data points can be clustered together and bound together in the form of clusters. One way to bind data points together is by calculating the centroid of the data points. In K-means clustering, the distance of each point from the cluster centroid is evaluated to determine effective clusters. Depending on the distance between the data point and the centroid, the data is assigned to the closest cluster. The goal of clustering is to determine the unique groupings of a set of unlabeled data. The "K" in K-means represents the number of clusters formed. The number of clusters (essentially the number of classes into which new instances of data can be classified) can be determined by the user. This determination can be made, for example, using feedback and observing the size of the clusters during training.

[0089] K-means is used when the dataset has distinct and well-separated points; otherwise, the modeling may render the clusters inaccurate if they are not separated. Additionally, K-means can be avoided if the dataset contains a large number of outliers or if the dataset is nonlinear.

[0090] FIG. 12 illustrates K-means clustering. In K-means clustering, data points are plotted and assigned a K value. For example, for K=2 in FIG. 12, the data points are plotted as shown in depiction 1210. Next, in step 1220, the points are assigned to similar centers. Cluster centroids are identified as shown in 1230. Once the centroids are identified, the points are reassigned to clusters such that the distance between the data points and the centroid of each cluster is minimized, as shown in 1240. New centroids of the clusters can then be determined as shown in depiction 1250. Once the data points have been reassigned to clusters and new centroids of the clusters have been formed, an iteration or series of iterations can occur to minimize the size of the clusters and determine the optimal centroid. Next, as new data points are measured, the new data points can be compared to the centroids and clusters and identified with those clusters.

[0091] Ensemble learning algorithms may be used. These algorithms use multiple learning algorithms to achieve better predictive performance than would be possible from any of the constituent learning algorithms alone. Ensemble learning algorithms perform the task of searching a hypothesis space to find suitable hypotheses that make good predictions for a particular problem. Even if the hypothesis space contains hypotheses that are highly suitable for a particular problem, finding a suitable hypothesis can be very difficult. Ensemble algorithms combine multiple hypotheses to form a better one. The term ensemble is typically used to refer to methods that generate multiple hypotheses using the same base learner. The broader concept of a multi-classifier system also encompasses the hybridization of hypotheses that are not derived from the same base learner.

[0092] Because evaluating the predictions of an ensemble typically requires more computation than evaluating the predictions of a single model, ensembles can be thought of as a way to compensate for poor learning algorithms by performing a lot of extra computation. Fast algorithms such as decision trees are commonly used in ensemble methods such as random forests, but slower algorithms can also benefit from ensemble methods.

[0093] Ensembles are themselves supervised learning algorithms because they can be used to make predictions after training. A trained ensemble therefore represents a single hypothesis. However, this hypothesis is not necessarily contained within the hypothesis space of the model from which it was constructed. Thus, ensembles can be shown to have more flexibility in the functions they can represent. This flexibility theoretically allows them to overfit the training data more than a single model, but in practice, some ensemble methods (especially bagging) tend to reduce the problems associated with overfitting the training data.

[0094] Empirically, ensemble algorithms tend to perform better when there is a significant degree of diversity among the models. Therefore, many ensemble methods aim to promote diversity among the models they combine. While counterintuitive, more random algorithms (such as random decision trees) can be used to generate more powerful ensembles than highly cautious algorithms (such as entropy-reducing decision trees). However, the use of a variety of powerful learning algorithms has been shown to be more effective than using techniques that attempt to simplify models to promote diversity.

[0095] The number of component classifiers in an ensemble has a significant impact on the accuracy of prediction. This becomes even more important for online ensemble classifiers, in order to determine a priori the size of the ensemble and the volume and velocity of the big data stream. Theoretical frameworks suggest that there is an ideal number of component classifiers in an ensemble, and that having more or fewer classifiers than this number will result in a decrease in accuracy. Theoretical frameworks also suggest that using the same number of independent component classifiers as class labels will result in the highest accuracy.

[0096] Some common types of ensembles include Bayesian optimal classifiers, bootstrap aggregation (bagging), boosting, Bayesian model averaging, Bayesian model combination, and model bucketing and stacking. Figure 13 shows an exemplary ensemble learning algorithm in which bagging is performed in parallel (1310) and boosting is performed sequentially (1320).

[0097] A neural network is a network or circuit of neurons, or in the modern sense, an artificial neural network, composed of artificial neurons or nodes. The connections of biological neurons are modeled as weights. Positive weights reflect excitatory connections, while negative values imply inhibitory connections. The inputs are modified by the weights and summed using linear combinations. An activation function can control the amplitude of the output. For example, the allowed range of the output is usually 0 to 1, but can also be -1 to 1.

[0098] These artificial networks can be used in predictive modeling, adaptive control, and other applications, and can be trained through datasets. Self-learning resulting from experience can occur within the network, allowing it to draw conclusions from complex and seemingly unrelated sets of information.

[0099] For completeness, a biological neural network consists of a group of chemically connected or functionally associated neurons. One neuron may be connected to many other neurons, and the total number of neurons and connections in the network may vary widely. Connections, called synapses, are usually formed from axons to dendrites, although dendritic synapses and other connections are also possible. Apart from electrical signaling, other forms of signaling result from the diffusion of neurotransmitters.

[0100] Artificial intelligence, cognitive modeling, and neural networks are information processing paradigms inspired by the way biological nervous systems process data. Artificial intelligence and cognitive modeling attempt to simulate some of the properties of biological neural networks. In the field of artificial intelligence, artificial neural networks have been successfully applied to speech recognition, image analysis, and adaptive control, and to build software agents or autonomous robots (in computer and video games).

[0101] A neural network (NN) is an interconnected group of natural or artificial neurons that uses a mathematical or computational model based on a connectionist approach to computation for information processing, in the case of artificial neurons, called an artificial neural network (ANN) or simulated neural network (SNN). In most cases, ANNs are adaptive systems that change their structure based on external or internal information flowing through the network. More practically, neural networks are nonlinear statistical data modeling or decision-making tools. They can be used to model complex relationships between inputs and outputs and to find patterns in data.

[0102] Artificial neural networks comprise networks of simple processing elements (artificial neurons) that can exhibit complex global behavior determined by the connections between the processing elements and element parameters.

[0103] One classic type of artificial neural network is the recurrent Hopfield network. The utility of artificial neural network models lies in the fact that they can be used to estimate and use functions from observations. Unsupervised neural networks can also be used to learn representations of inputs that capture salient features of the input distribution, and more recently, in deep learning algorithms that can implicitly learn distribution functions for observed data. Training with neural networks is particularly useful in applications where the complexity of the data or the task makes it impossible to design such functions manually.

[0104] Neural networks can be used in a variety of fields. Tasks to which artificial neural networks are applied tend to span the following broad categories: function approximation or regression analysis, including time series prediction and modeling; pattern and sequence recognition; classification, including novelty detection and sequential decision making; data processing, including filtering, clustering, blind signal separation, and compression.

[0105] Areas of application of ANNs include identification and control of nonlinear systems (vehicle control, process control), game playing and decision making (backgammon, chess, racing), pattern recognition (radar systems, face identification, object recognition), sequence recognition (gesture, speech, handwritten text recognition), medical diagnosis, financial applications, data mining (or knowledge discovery in databases, or "KDD"), visualization, and email spam filtering. For example, it is possible to create semantic profiles of user interests resulting from photographs trained for object recognition.

[0106] 14 shows an exemplary neural network. The neural network has an input layer represented by multiple inputs, such as 14101 and 14102. The inputs 14101, 14102 are fed into a hidden layer, shown as including nodes 14201, 14202, 14203, and 14204. These nodes 14201, 14202, 14203, and 14204 are combined to produce output 1430 in the output layer. While neural networks perform simple processing through a hidden layer of simple processing elements, nodes 14201, 14202, 14203, and 14204, these nodes can exhibit complex global behavior determined by the connections between the processing elements and element parameters.

[0107] The neural network of Figure 14 may be implemented in hardware. Referring to Figure 15, a hardware-based neural network is shown.

[0108] Cardiac arrhythmias, and atrial fibrillation (AF) in particular, remain common and dangerous conditions, especially in the aging population. In patients with normal sinus rhythm, the heart, comprised of atria, ventricles, and excitable conduction tissue, is electrically excited to beat in a synchronous, patterned manner. In patients with cardiac arrhythmias, abnormal regions of cardiac tissue do not follow the synchronous beating cycle associated with normal conduction tissue, as in patients with normal sinus rhythm. In contrast, abnormal regions of cardiac tissue undergo abnormal conduction to adjacent tissue, disrupting the cardiac cycle and resulting in an asynchronous cardiac rhythm. Such abnormal conduction has long been known to occur in various regions of the heart, such as within the region of the sinoatrial (SA) node, along the conduction pathways of the atrioventricular (AV) node and the His bundle, or within the myocardial tissue forming the walls of the ventricles and atria.

[0109] Catheter ablation-based treatments may involve mapping the electrical properties of cardiac tissue, particularly the endocardium and cardiac volumes, and selectively ablating the cardiac tissue through the application of energy. Cardiac mapping, for example, generating an electrical potential map (voltage map) of wave propagation along cardiac tissue or a map of arrival times to where various tissues are located (local time activation (LAT) map), can be used to detect local dysfunction of cardiac tissue. Ablation, such as that based on cardiac mapping, can stop or modify the propagation of unwanted electrical signals from one part of the heart to another.

[0110] Cardiac ablation and other cardiac electrophysiology procedures are becoming increasingly complex as physicians treat challenging conditions such as atrial fibrillation and ventricular tachycardia. Treatment of refractory arrhythmias currently relies on the use of three-dimensional (3D) mapping systems to reconstruct the anatomical structure of the cardiac chamber of interest. For example, cardiologists rely on software such as the Complex Fractionated Atrial Electrogram (CFAE) module of the CARTO® 3 3D Mapping System manufactured by Biosense Webster, Inc. (Diamond Bar, California) to analyze intracardiac EGM signals and determine ablation points for treating various cardiac conditions, including unconventional atrial flutter and ventricular tachycardia. 3D maps can provide multiple pieces of information about the electrophysiological properties of tissues that represent the anatomical and functional substrates of these challenging arrhythmias.

[0111] Electrode catheters have been commonly used in medical practice for many years. They are used to stimulate and map electrical activity within the heart and to ablate sites of abnormal electrical activity. In use, an electrode catheter is inserted into a major vein or artery, such as the femoral artery, and then guided into a target heart chamber. A typical ablation procedure involves inserting a catheter with at least one electrode at its distal end into a heart chamber. A reference electrode is typically provided by a second catheter taped to the patient's skin or positioned within or near the heart. When RF (radio frequency) current is applied to the tip electrode of the ablation catheter, current flows through the medium (i.e., blood and tissue) surrounding the tip electrode toward the reference electrode. The distribution of the current depends on the amount of electrode surface in contact with the tissue compared to blood, which has a higher electrical conductivity than tissue. Tissue heating occurs due to the electrical resistance of the tissue. Sufficient tissue heating can cause cell destruction in the cardiac tissue, resulting in lesions within the non-conductive cardiac tissue. During this process, the electrode also heats due to conduction from the heated tissue to the electrode itself. If the electrode temperature becomes high enough, possibly exceeding 60°C, a thin, transparent film of dehydrated blood proteins can form on the electrode's surface. As the temperature continues to rise, this dehydrated layer can gradually thicken, causing blood to coagulate on the electrode surface. Because dehydrated biological material has a higher electrical resistance than endocardial tissue, the impedance to the flow of electrical energy into the tissue also increases. If the impedance becomes high enough, an impedance rise occurs, requiring the catheter to be removed from the body and the tip electrode to be cleaned.

[0112] A prerequisite for successful catheter ablation is that the source of the cardiac arrhythmia and the surrounding region of the heart must be accurately localized within the cardiac chamber. Such localization can be achieved by electrophysiological studies, in which electrical potentials are detected and spatially resolved by a mapping catheter introduced into the cardiac chamber. This electrophysiological study, also known as electroanatomical mapping, thus provides 3D mapping data that can be displayed on a monitor. In many cases, the mapping function and the therapeutic function (e.g., ablation) are provided by a single catheter or group of catheters, whereby the mapping catheter also simultaneously operates as a therapeutic (e.g., ablation) catheter.

[0113] Cardiac mapping can be performed using one or more techniques. As an example of a first technique, cardiac mapping may be performed by sensing electrical properties of cardiac tissue, such as regional activation time, as a function of precise location within the heart. Corresponding data may be acquired with one or more catheters advanced into the heart using catheters having electrical and position sensors at their distal tips. By way of example, position and electrical activity may be initially measured at approximately 10 to approximately 20 points on the inner surface of the heart. These data points may generally be sufficient to generate a preliminary reconstruction or map of the cardiac surface of satisfactory quality. The preliminary map may be combined with data taken at additional points to generate a more comprehensive map of the cardiac electrical activity. In clinical settings, it is not uncommon to accumulate data at 100 or more sites to generate a detailed, comprehensive map of the electrical activity of the cardiac chambers. The detailed map may then serve as a basis for making decisions regarding therapeutic action, such as tissue ablation, to alter the propagation of cardiac electrical activity and restore normal cardiac rhythm.

[0114] A catheter containing a position sensor may be used to determine the trajectory of each point on the heart's surface. These trajectories can be used to infer motion characteristics, such as the contractile force of the tissue. A map indicative of such motion characteristics may be constructed when trajectory information is sampled at a sufficient number of points within the heart.

[0115] Electrical activity at a point within the heart can typically be measured by advancing a catheter containing an electrical sensor at or near its distal tip to the point within the heart, contacting tissue with the sensor, and acquiring data at the point. Multi-electrode catheters may be implemented using any applicable shape, such as a linear catheter with multiple electrodes, a balloon catheter including electrodes distributed on multiple spines forming a balloon, a lasso or loop catheter with multiple electrodes, or any other applicable shape.

[0116] According to one example, a multi-electrode catheter can be advanced into a cardiac chamber. Anterior-posterior (AP) and lateral fluorograms can be acquired to determine the location and orientation of each electrode. Electrograms can be recorded from each electrode in contact with the cardiac surface relative to a temporal reference, such as the onset of the P wave in sinus rhythm obtained from a surface ECG. The system further disclosed herein can distinguish between those electrodes that record electrical activity and those that do not due to lack of proximity to the endocardial wall. After an initial electrogram is recorded, the catheter can be repositioned, and fluorograms and electrograms can be recorded again. An electrical map can then be constructed from a repetition of the above process.

[0117] According to another embodiment, a technique and apparatus for mapping the electrical potential distribution of a cardiac chamber can be implemented. An intracardiac multi-electrode mapping catheter assembly can be inserted into a patient's heart. The mapping catheter assembly can include a multi-electrode array with an integral reference electrode, or preferably, a companion reference catheter. The electrodes can be deployed in the form of a substantially spherical array. The electrode array can be spatially referenced to a point on the endocardial surface by the reference electrode or by a reference catheter in contact with the endocardial surface. A preferred electrode array catheter can carry a large number of individual electrode sites (e.g., at least 24). Additionally, the technique of this embodiment can be implemented by knowing the location of each electrode site on the array as well as the cardiac geometry. These locations are preferably determined by the technique of impedance plethysmography.

[0118] According to other embodiments, body patches and / or body surface electrodes may be positioned on or adjacent to a patient's body. A catheter having one or more electrodes may be positioned within the patient's body (e.g., within the patient's heart), and the position of the catheter may be determined by the system based on signals transmitted and received between one or more electrodes of the catheter and the body patch and / or body surface electrodes. In addition, the catheter electrodes may sense biometric data (e.g., LAT values) from within the patient's body (e.g., within the heart). The biometric data may be associated with the determined position of the catheter such that a rendering of the patient's body part (e.g., heart) can be displayed and the biometric data can be shown superimposed on the shape of the body part, as determined by the position of the catheter.

[0119] Electrical signals, such as electrocardiogram (ECG) signals, are often detected before and / or during cardiac procedures. For example, ECG signals can be used to identify potential locations in the heart where arrhythmias may occur. In general, ECG signals are signals that describe the electrical activity of the heart. ECG signals may also be used to map portions of the heart.

[0120] An ECG signal is generated by the contraction (depolarization) and relaxation (repolarization) of the atrial and ventricular muscles of the heart. As shown by signal 1602 in FIG. 16, the ECG signal includes a P wave (due to atrial depolarization), a QRS complex (due to atrial and ventricular depolarization), and a T wave (due to ventricular repolarization). To record the ECG signal, electrodes can be placed at specific locations on the body or can be positioned inside the body via a catheter. Artifacts (e.g., noise) are unwanted signals that merge with electronic signals such as ECG signals and sometimes impair the diagnosis and / or treatment of cardiac disorders. Electrical signal artifacts can be baseline wander, power line interference, electromyogram (EMG) noise, power line noise, etc.

[0121] Additionally, biometric (e.g., biopotential) patient monitors may use surface electrodes to make measurements of biopotentials, such as ECG or electroencephalogram (EEG). The fidelity of these measurements is limited by the effectiveness of the electrode connection to the patient. The resistance of the electrode system to the flow of electrical current, known as electrical impedance, characterizes the effectiveness of the connection. Typically, the higher the impedance, the lower the fidelity of the measurement. Several mechanisms can result in lower fidelity.

[0122] FIG. 17 is a diagram of an example system 1720 capable of implementing one or more features of the presently disclosed subject matter. All or a portion of the system 1720 can be used to collect information for a training dataset and / or to implement a trained model. The system 1720 can include a component, such as a catheter 1740, configured to injure a tissue region of an internal organ. The catheter 1740 may also be further configured to acquire biometric data, including electronic signals. While the catheter 1740 is shown as a point catheter, it will be understood that any shape of catheter including one or more elements (e.g., electrodes) can be used to implement the embodiments disclosed herein. The system 1720 includes a probe 1721 having a shaft that can be navigated by a physician 1730 into a body part, such as a heart 1726, of a patient 1728 reclining on a table 1729. According to an embodiment, multiple probes may be provided; for simplicity, a single probe 1721 is described in this example, although it will be understood that probe 1721 may represent multiple probes. As shown in FIG. 17 , a physician 1730 can insert a shaft 1722 through a sheath 1723 while manipulating the distal end of the shaft 1722 using a manipulator near the proximal end of the catheter 1740 and / or deflection from a sheath 1723. As shown in inset 1725, a catheter 1740 can be attached to the distal end of the shaft 1722. The catheter 1740 can be inserted through the sheath 1723 in a collapsed state and then expanded within the heart 1726. As described further herein, the catheter 1740 can include at least one ablation electrode 1747 and a catheter needle.

[0123] According to embodiments, catheter 1740 may be configured to ablate a tissue region in a chamber of heart 1726. Inset 1745 shows a close-up of catheter 1740 inside a chamber of heart 1726. As shown, catheter 1740 may include at least one ablation electrode 1747 coupled to the body of the catheter. According to other embodiments, multiple elements may be connected via splines that form the shape of catheter 1740. One or more other elements (not shown) may be provided and may be any element configured to perform ablation or acquire biometric data, such as an electrode, a transducer, or one or more other elements.

[0124] According to embodiments disclosed herein, an ablation electrode, such as electrode 1747, may be configured to deliver energy to a tissue region of a body organ, such as heart 1726. The energy may be thermal energy and may cause damage to the tissue region starting at the surface of the tissue region and extending through the thickness of the tissue region.

[0125] According to embodiments disclosed herein, the biological data may include one or more of LAT, electrical activity, topology, bipolar mapping, dominant frequency, impedance, etc. The local activation time may be the time point of a threshold activation corresponding to local activation calculated based on a normalized initial onset. The electrical activity may be any applicable electrical signal that can be measured based on one or more thresholds and detected and / or enhanced based on signal-to-noise ratio and / or other filters. The topology may correspond to the physical structure of a body part or a portion of a body part, or may correspond to changes in the physical structure for different portions of the body part or for different body parts. The dominant frequency may be a frequency or range of frequencies prevalent in a portion of a body part and may differ in different portions of the same body part. For example, the dominant frequency of the pulmonary veins of a heart may be different from the dominant frequency of the right atrium of the same heart. The impedance may be a resistance measurement in a given region of a body part.

[0126] 17 , the probe 1721 and catheter 1740 can be connected to a console 1724. The console 1724 can include a processor 1741, such as a general-purpose computer with suitable front-end and interface circuitry 1738, for transmitting and receiving signals to and from the catheter, as well as for controlling other components of the system 1720. In some embodiments, the processor 1741 can be further configured to receive biometric data, such as electrical activity, and determine whether a given tissue region conducts electricity. According to one embodiment, the processor can be external to the console 1724, for example, located in the catheter, an external device, a mobile device, a cloud-based device, or can be a stand-alone processor.

[0127] As mentioned above, processor 1741 may include a general-purpose computer, which may be programmed with software to perform the functions described herein. The software may be downloaded in electronic form to the general-purpose computer, for example, over a network, or alternatively or additionally, may be provided and / or stored on non-transitory tangible media, such as magnetic, optical, or electronic memory. The exemplary configuration shown in FIG. 17 may be modified to implement embodiments disclosed herein. Embodiments of the present disclosure may be similarly applied using other system components and configurations. Additionally, system 1720 may include additional components, such as elements for sensing electrical activity, wired or wireless connectors, processing and display devices, etc.

[0128] According to one embodiment, the display 1727 connected to the processor (e.g., processor 1741) may be located at a remote location, such as a separate hospital, or within a separate healthcare provider network. Additionally, the system 1720 may be part of a surgical system configured to obtain anatomical and electrical measurements of a patient's organs, such as the heart, and to perform cardiac ablation procedures. One example of such a surgical system is the Carto® system sold by Biosense Webster.

[0129] System 1720 can also, and optionally, acquire biometric data, such as anatomical measurements of the patient's heart, using ultrasound, computed tomography (CT), magnetic resonance imaging (MRI), or other medical imaging techniques known in the art. System 1720 can acquire electrical measurements using a catheter, an electrocardiogram (EKG), or other sensors that measure the electrical properties of the heart. The biometric data, including the anatomical and electrical measurements, may then be stored in memory 1742 of mapping system 1720, as shown in FIG. 17 . The biometric data may be transmitted from memory 1742 to processor 1741. Alternatively, or in addition, the biometric data may be transmitted to server 1760, which may be local or remote, using network 1762.

[0130] The network 1762 may be any network or system commonly known in the art, such as an intranet, a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a direct or series of connections, a cellular telephone network, or any other network or medium capable of facilitating communication between the mapping system 1720 and the server 1760. The network 1762 may be wired, wireless, or a combination thereof. Wired connections may be implemented using Ethernet, Universal Serial Bus (USB), RJ-11, or any other wired connection commonly known in the art. Wireless connections may be implemented using Wi-Fi, WiMAX, Bluetooth, infrared, cellular networks, satellite, or any other wireless connection method commonly known in the art. Additionally, several networks may operate alone or in communication with each other to facilitate communication within the network 1762.

[0131] In some cases, server 1760 may be implemented as a physical server. In other cases, server 1762 may be implemented as a virtual server, a public cloud computing provider (e.g., Amazon Web Services (AWS)).

[0132] According to one exemplary embodiment, server 1760 may be implemented as or in communication with a processor that stores a machine learning algorithm, such as neural network 1790. In another embodiment, neural network 1790 may be implemented within console 1724. For example, without limitation, neural network 1790 may be implemented on one or more CPU processors, one or more GPU processors, one or more FPGA chips, or a specialized ASIC that performs deep learning computations, such as the Intel® Nervana™ Neural Network Processor. According to one exemplary embodiment, neural network 1790 may be located, without limitation, in a medical procedure room, on a server or processor at a hospital or medical facility, on a remote server or processor, or in the cloud.

[0133] The control console 1724 may be connected by a cable 1739 to body surface electrodes 1743, which may include adhesive skin patches that are affixed to the patient 1730. A processor in conjunction with the current tracking module may determine position coordinates of the catheter 1740 within the patient's body part (e.g., the heart 1726). The position coordinates may be based on impedance or electromagnetic fields measured between the body surface electrodes 1743 and electrodes 1747 or other electromagnetic components of the catheter 1740. Additionally or alternatively, the location pad may be located on the surface of the bed 1729 or may be separate from the bed 1729.

[0134] Processor 1741 may include real-time noise reduction circuitry, typically configured as a field programmable gate array (FPGA), followed by an analog-to-digital (A / D) ECG (electrocardiograph) or EMG (electromyogram) signal conversion integrated circuit. Processor 1741 may communicate signals from the A / D ECG or EMG circuitry to another processor and / or may be programmed to perform one or more functions disclosed herein.

[0135] The control console 1724 may also include an input / output (I / O) communication interface that allows the control console to communicate signals to and / or from the electrodes 1747 .

[0136] During a procedure, processor 1741 facilitates presentation of body part rendering 1735 to physician 1730 on display 1727 and may store data representing body part rendering 1735 in memory 1742. Memory 1742 may comprise any suitable volatile and / or non-volatile memory, such as random access memory or a hard disk drive. In some embodiments, medical professional 1730 may be able to manipulate body part rendering 1735 using one or more input devices, such as a touchpad, mouse, keyboard, gesture recognizer, etc. For example, the input device may be used to change the position of catheter 1740 so that rendering 1735 is updated. In an alternative embodiment, display 1727 may include a touch screen, which may be configured to receive input from medical professional 1730 in addition to presenting body part rendering 1735.

[0137] According to one embodiment, a neural network 1790 may be provided to automatically detect and identify the location of particular structures within the heart, such as the His bundle. The neural network 1790 may be of the form described above in connection with Figures 14 and 15.

[0138] The His bundle is a portion of the myocardium that originates near the orifice of the CS. The His bundle is an important part of the cardiac electrical conduction system, as it transmits electrical impulses from the atrioventricular (AV) node, located between the atria and ventricles, to the ventricles of the heart. The His bundle is located in a vulnerable position within the heart, and if accidentally ablated during a catheter ablation procedure, it can cause harmful and undesirable effects on the cardiac electrical conduction system. During conventional ablation procedures, physicians can manually tag the His bundle to identify its location within the heart so that it can be avoided during the ablation procedure. Such manual tagging of the His bundle is tedious and time-consuming. Manual tagging can also result in false-positive readings, in which an electrocardiogram (ECG) signal appears to resemble a His bundle impulse, but the location of the impulse does not exactly correspond to the location of the His bundle. In other conventional ablation procedures, physicians sometimes fail to tag the His bundle, which increases patient risks during the ablation procedure.

[0139] According to the exemplary embodiment of Figure 18A, a neural network 1800, providing an example of the neural network 1790 of Figure 17, receives input data 1810 to train the neural network 1800 to automatically identify cardiac structures of interest, such as the His bundle 1820, with greater efficiency and reliability than manual identification by a physician during a procedure, such as an ablation procedure. Non-limiting examples of the input data 1810 may include an intracardiac electrogram (EGM) or ECG signal 1830 received by an electrode or bipolar electrode pair of a catheter, such as a mapping catheter, a distance 1840 (typically measured in millimeters) between an electrode of a first mapping catheter and a point on a second reference catheter, or other input 1850. Other inputs 1850 may include a discrete Boolean value (i.e., 0 or 1) indicating whether the distance between the electrode of the first mapping catheter and the point on the second reference catheter is less than a predetermined threshold, the force (typically measured in grams) applied by the catheter to the cardiac structure of interest as measured by a force sensor within the catheter, an indicator of the proximity between the catheter electrode and the cardiac structure of interest, an impedance value (typically measured in ohms) of the cardiac structure of interest measured by the catheter electrode, an electrocardiogram (ECG) signal 1830 received by the body surface electrode(s), manual mapping data of the cardiac structure of interest, and any other electrophysiological data measured by the catheter electrode.

[0140] In one embodiment, one or more of the input data 1810 are provided to the neural network 1800. The input data 1810 may be stored in various locations, including but not limited to, at a hospital or medical facility, at a remote server location, or in the cloud. The training data 1810 may be transferred to the neural network 1800 in real time, at predetermined intervals, or on demand. Once trained, the neural network 1800 can identify the His bundle 1820 in real time during a catheter ablation procedure.

[0141] In one embodiment, the output 1820 of the neural network 1800 may include, but is not limited to, a discrete Boolean value indicating whether the electrodes of the mapping catheter are sufficiently close to the cardiac structure of interest, such as the His bundle, and a continuous value, such as a match index, indicating the match between the cardiac characteristics of interest measured by the electrodes of the mapping catheter and the characteristics of the cardiac structure of interest, such as those obtained by manual mapping.

[0142] According to an exemplary embodiment, the neural network 1800 may include a recurrent neural network (RNN), such as a convolutional neural network (CNN) or a long short-term memory (LSTM) neural network. A convolutional neural network (CNN) is a deep learning algorithm preferably used in the fields of computer vision and / or image recognition. A CNN assigns importance (learnable weights) to various aspects or features in an input image in order to distinguish one from another. An LSTM neural network is a recurrent neural network with feedback connections used in deep learning.

[0143] In an exemplary embodiment, after each training, the trained model, including its output, can be run against a standard database, such as a gold standard database, to verify its accuracy. In one non-limiting example, the gold standard database consists of points known to be the cardiac structure of interest (e.g., the His bundle), points known not to be this structure, associated catheter locations, ECG signals, and other relevant parameters. In an exemplary embodiment, if the accuracy of the newly trained model falls below a threshold, or alternatively, if the accuracy of the newly trained model is less than the accuracy of the previous model, the model can be discarded. Similarly, if the accuracy of the newly trained model is equal to or greater than a threshold, or alternatively, if the accuracy of the newly trained model is greater than the accuracy of the previous model, the model can be published to the on-site mapping system. In an exemplary embodiment, publishing a new model can be performed manually, for example, by an operator downloading a file from a web address and uploading it to the mapping system 1720. Alternatively, the new model can be pushed to the on-site mapping system 1720 via the internet.

[0144] 18B is a flow diagram illustrating an exemplary embodiment of a module 1860 for training a neural network 1800 to automatically detect and identify the location of specific structures within the heart. While FIG. 18B illustrates a module for automatically identifying the His bundle within the heart in conjunction with FIG. 18A, one skilled in the art will recognize that other cardiac structures or signals may be identified according to module 1860.

[0145] For example, during a cardiac ablation procedure, multiple catheters may be utilized to acquire various data recordings of the heart. FIG. 19A illustrates multiple intracardiac catheters and body surface electrodes used to acquire electrophysiological data of a heart 1910. According to one exemplary embodiment, for example, the catheters may include, but are not limited to, an ablation catheter 1920, a CS reference catheter 1930, and a His bundle mapping catheter 1940. Other probes, such as a right ventricular apex (RVA) catheter 1950, may also be used if desired. The ablation catheter 1920 is utilized to perform the ablation procedure as discussed above. The CS reference catheter 1930 is positioned inside the coronary sinus 1932 of the heart 1910. The His bundle mapping catheter 1940 is configured to contact the His bundle and record a His bundle electrogram (HBE), as shown in FIG. 20, discussed herein. Body surface electrodes 1960 are positioned on the body surface to record cardiac ECG signals, as described above and shown in FIGS.

[0146] According to one exemplary embodiment, electrophysiological data obtained from catheters 1920, 1930, 1940, 1950, and electrodes 1960 may be provided to a processor such as processor 1741 for analysis and output to display 1727, and preferably transmitted to neural network 1790 as shown in FIG. 17.

[0147] At step 1865, the neural network 1800 receives first input data from a first catheter. In one embodiment, the first input data 1810 is preferably electrophysiological data, more preferably an intracardiac electrogram (EGM) signal 1840 received by an electrode or bipolar electrode pair of the first catheter. In one embodiment, the first catheter is a mapping catheter that receives EGM signals from a cardiac structure of interest. In one embodiment, the first catheter is a His bundle mapping catheter 1940 that receives His bundle electrogram (HBE) signals, as shown in FIGS. 19A-19B. The first catheter can include multiple electrodes. FIG. 19B shows a His bundle mapping catheter 1940 having four electrodes 1942a, 1942b, 1942c, and 1942d. However, one skilled in the art will recognize that the His bundle mapping catheter 1940 can include any number of electrodes. The His bundle mapping catheter 1940 can receive EGM signals at any of the electrodes 1942a, 1942b, 1942c, 1942d.

[0148] At step 1875, neural network 1800 receives additional input data. In one embodiment, the additional input data may be a second EGM signal received from a second electrode, such as electrode 1942b of His bundle mapping catheter 1940. Those skilled in the art will recognize that the additional input data may include multiple EGM signals received from different electrodes of His bundle mapping catheter 1940.

[0149] To help illustrate aspects of the present disclosure, FIG. 20 shows exemplary surface ECG and intracardiac HBE tracings 2010 and 2020, respectively, that can be used in accordance with the methods of the present disclosure. As shown in HBE tracing 2020, the A-wave indicates low right atrial activation, His bundle activity is indicated by an H, and V-deflection indicates ventricular activation. Conventionally, the period between the onset of the A-wave and the onset of the subsequent V-deflection is known as the AV interval 2022. When the His bundle mapping catheter 1940 contacts the His bundle, the HBE signal preferably has a pattern as shown in HBE tracing 2020.

[0150] As known in the art and discussed above, the body surface electrode 1960 functions as a reference electrode and produces a body surface ECG recording 2010 of the cardiac cycle, including a P wave, a QRS complex, and a T wave, as shown. The P wave represents the atrioventricular polarization phase, the QRS complex represents ventricular repolarization, and the T wave represents ventricular depolarization. Correspondingly, the period between the onset of the P wave and the onset of the QRS complex is known as the PR interval 2012. Line 2030 indicates the point in the body surface ECG recording 2010 where the electrical impulse passes through the His bundle.

[0151] In one embodiment, neural network 1800 identifies electrophysiological data, such as an HBE signal, corresponding to the location of the His bundle based on the input data. For example, neural network 1800 identifies whether the input data includes electrophysiological data corresponding to the His bundle.

[0152] However, even if the HBE signal has a pattern as shown in HBE recording 2020, it may be that the His bundle mapping catheter 1940 gives a false positive reading, such as when the His bundle mapping catheter 1940 is close to but not touching the His bundle. Therefore, the neural network may also rely on the distance between the electrodes of the His bundle mapping catheter 1940 and the CS reference electrode 1930, as discussed herein.

[0153] In one embodiment, the additional input data may also include the distance between the electrodes on the His bundle mapping catheter 1940 and the electrodes on the reference catheter. For example, the reference catheter may be the CS reference catheter 1930, which is inserted into the coronary sinus of the heart, as shown in FIGS. 19A-19B. For example, it is well understood that the His bundle is anatomically located close to the CS. According to one exemplary embodiment, the distance between the electrodes on the His bundle mapping catheter 1940 and the electrodes on the CS reference catheter 1930 is used as proximity data for determining the reliability of the HBE recording 2020.

[0154] In one embodiment, the CS reference catheter 1930 can include multiple electrodes. While FIG. 19B shows the CS reference catheter 1940 with ten electrodes 1932a, 1932b, 1932c, 1932d, 1932e, 1932f, 1932g, 1932h, 1932i, and 1932j, one skilled in the art would recognize that the CS reference catheter 1940 can include any number of electrodes. In one embodiment, a distance can be measured between each electrode 1942a-d of the His bundle mapping catheter 1940 and the closest point on the CS reference catheter 1930. For example, as shown in FIG. 19B, distance D3 is the distance between electrode 1942c of the His bundle mapping catheter 1940 and the closest point on the CS reference catheter 1930. Alternatively, or in addition, distance can be measured between each electrode 1942a-d of the His bundle mapping catheter 1940 and the nearest electrode 1932a-j of the CS reference catheter 1930. Alternatively, or in addition, distance can be measured between each electrode 1942a-d of the His bundle mapping catheter 1940 and a selected electrode of the CS reference catheter 1930, which can be any one of the electrodes 1932a-j.

[0155] Those skilled in the art will recognize that the first input data and additional input data referenced in steps 1865 and 1885 may be any of the input data discussed herein and any other electrophysiological data measured by the electrodes of the catheter.

[0156] In step 1885, the neural network 1800 applies a machine learning algorithm to each received input data to identify the location of a cardiac structure of interest, such as the His bundle. For example, the EGMs received by each electrode 1942a-d of the His bundle mapping catheter 1940 and the distance between each electrode 1942a-d and the nearest point on the CS reference catheter 1930 can be used to determine whether any of the electrodes 1942a-d are positioned on the His bundle, and if so, which electrode is positioned on the His bundle. For example, as shown in FIG. 19B , the EGMs received by electrode 1942c of the His bundle mapping catheter 1940 and the distance D3 between electrode 1942c and the CS reference catheter 1930 can be used as inputs to the neural network 1800 to determine whether electrode 1942c of the His bundle mapping catheter 1940 is positioned on the His bundle. As described in more detail below, determining whether an electrode is positioned on the His bundle may include review of other inputs including, for example, the ECG signal, distance from the CS catheter, and force, contact state.

[0157] In another example, if the HBE recording 2020 at a first selected electrode of the His bundle mapping catheter 1940 has characteristics of the His bundle, but the spatial location of the first selected electrode has a distance greater than a predetermined threshold or range from the CS reference catheter 1930, as shown with reference to distance D2 in FIG. 19A , the neural network 1800 determines that the first selected electrode is not the correct location of the His bundle. On the other hand, if the HBE recording 2020 has characteristics of the His bundle, and the spatial location of a second selected electrode along the His bundle mapping catheter 1940 has a distance less than a predetermined threshold or range from the CS reference catheter 1930, as shown with reference to distance D1 in FIG. 19A , the neural network 1800 determines that the second selected electrode is the correct location of the His bundle.

[0158] In one embodiment, the neural network learns a predetermined threshold based on the location of the His bundle manually marked by the physician, the distance of the His bundle manually marked to the CS reference catheter 1930, and electrophysiological data such as HBE recording 2020 or ECG recording 2010.

[0159] As a result, the neural network 1800 learns to automatically detect the location of particular structures within the heart, such as the His bundle, based on input data such as electrophysiological data received from a first catheter, such as EGM data received by the electrodes of the His bundle mapping catheter 2040, and optionally additional data such as EGM data received from other electrodes of the first catheter, electrophysiological data received from a second catheter, such as a CS reference catheter, manual mapping data, ECG data, EGM data, distance data, force data, proximity indicator data, impedance data, and any other electrophysiological data measured by the electrodes of the catheters. This additional data may be relevant to detecting the His bundle, and although the additional data may seem unimportant, the AI algorithms described herein can find correlations and significance with the data in these input data.

[0160] In step 1890, the neural network 1800 generates an output of whether the first catheter electrode is positioned on or sufficiently close to the cardiac structure of interest, such as the His bundle. As discussed above, the output of the neural network may include, but is not limited to, a discrete Boolean value indicating whether the catheter electrode is sufficiently close to the cardiac structure of interest, such as the His bundle, or a continuous value, such as a match index, indicating a match between a characteristic of the cardiac structure of interest measured by the catheter electrode and a characteristic of the cardiac structure of interest, such as obtained by manual mapping.

[0161] In another embodiment, the neural network 1800 can be used to detect Local Abnormal Ventricular Activations (LAVA) signals within the heart. In such an embodiment, LAVA and non-LAVA signals are used as training data in the model, and the neural network learns to distinguish between LAVA and non-LAVA signals in order to automatically detect LAVA signals in a clinical setting.

[0162] An exemplary embodiment of a convolutional neural network (CNN) 2100 for automatically identifying cardiac structures of interest is shown in FIG. 21. As shown in FIG. 21, the CNN 2100 preferably receives input data 2110. The input data may include a single input or multiple inputs 2110-1, 2110-2, 2110-3, ..., 2110-n, where "n" is the last of the multiple inputs. By way of example, and not limitation, the first input 2110-1 may include a first EGM signal received by a first electrode of the mapping catheter, the second input 2110-2 may include a second EGM signal received by a second electrode of the mapping catheter, the third input 2110-3 may include the distance between the first electrode of the mapping catheter and the nearest electrode of the reference catheter, and the final input (2110-n) may include the distance between the second electrode of the mapping catheter and the nearest electrode of the reference catheter. Inputs 2110 are provided to a first hidden layer 2120 including nodes 2120-1, 2120-2, 2120-3, ... 2120-n, and optionally to a second or more hidden layers 2130 including nodes 2130-1, 2130-2, 2130-3, ... 2130-n, which are combined to generate an output 2140, such as a Boolean value or a match indicator. For example, in CNN 2100, all EGM inputs 2110 are fed into the neural network at once to compute output 2140. The neural network may include, for example, a series of convolutional layers feeding one or more pooling and flattening layers to provide outputs to the hidden layers, as described more particularly below.

[0163] An exemplary embodiment of a recurrent neural network (RNN) 2200, such as a long-short-term memory (LSTM) neural network, for automatically identifying cardiac structures of interest is shown in FIG. 22. As shown in FIG. 22, the RNN 2200 preferably receives input data 2210. The input data 2210 may include a single input or multiple inputs 2210-1, 2210-2, 2210-3, ..., 2210-n, where "n" is the last of the multiple inputs. By way of example and not limitation, the first input 2210-1 may include a first EGM signal received by a first electrode of the mapping catheter, the second input 2210-2 may include a distance between the first electrode of the mapping catheter and the nearest electrode of the reference catheter, the third input 2210-3 may include force data received from the electrodes of the mapping catheter, and the final input 2210-n may include impedance data received from the electrodes of the mapping catheter. Inputs 2210 are provided to RNN 2200 and combined to generate output 2240, such as a Boolean value or a match indicator. For example, in RNN 2200, EGM inputs 2210-1 are fed into the neural network one at a time. The more EGM samples are fed into RNN 2200, the more accurate the output 2240 becomes.

[0164] In step 1895 of FIG. 18B , the output of neural network 1800, such as output 2140 or 2240, is used to train neural network 1800. Specifically, the output of neural network 1800 provides a system output, which is further provided to recursively train neural network 1800 to achieve an improved output. For example, as discussed above, after each training, the training model including its output can be run against a standard database, such as a gold standard database, to verify its accuracy. For example, the output may be a legitimate output indicating where the His bundle is or is not present, and the output may be used to further train the algorithm. Additionally, as shown with reference to arrow 2230 in FIG. 22 , if the accuracy of the newly trained model is above a threshold, or alternatively, if the accuracy of the newly trained model is higher than the accuracy of the previous model, the model can be used as input for the neural network.

[0165] In one embodiment, the training of neural network 1800 may be supervised at the facility where the cardiac procedure is performed, such as a hospital or medical facility, or at a remote location, such as a training center.

[0166] Once the neural network 1800 is trained, it can be utilized in real time to automatically locate specific structures within the heart, such as the His bundle.

[0167] 23 shows an implementation 2300 of the system as described. Implementation 2300 includes a series of inputs to network 2100, including distance ECG input 1830, distance from catheter 1840, and other inputs 1850, to generate output 1820 including probability (or location) of His bundle detection and probability (or location) of non-His bundle detection. As mentioned above, ECG input 1830 includes a first ECG 18301, a second ECG 18302, ..., a final ECG 1830. N The ECG data may include any number of ECG data, including:

[0168] The network 2100 may be, for example, a CNN network as described herein. For simplicity, the network 2100 may include multiple convolutional layers 2310 interconnected with multiple pooling layers 2320, and further interconnected with multiple flattening layers 2330, which may include one or more retransmission and / or fully connected layers. As will be appreciated, the softmax layer 2340 may be the last layer in the network 2100.

[0169] The convolutional layer 2310 is described herein, including with respect to at least FIG. 21 . The pooling layer 2320 may be used to reduce the dimensionality of the feature map. The pooling layer 2320 may be used to reduce the number of parameters to learn and the amount of computation performed within the network 2100. The pooling layer 2310 summarizes features present within a region of the feature map generated by the convolutional layer 2310. The flattening layer 2330 converts the pooled feature map into a single column that is passed to a fully connected layer, which is added to the neural network 2100.

[0170] The Softmax layer 2340 provides a function that transforms a vector of K real values into a vector of K real values that sum to 1. The input values may be positive, negative, zero, or greater than 1. The Softmax layer 2340 may transform the inputs to the Softmax layer 2340 into values between 0 and 1 to allow interpretation as probabilities. If one of the inputs is small or negative, the Softmax layer 2340 may transform the input into a small probability, and if the input is large, the Softmax layer 2340 transforms the input into a large probability.

[0171] The softmax layer 2340 is sometimes referred to as a softargmax function or multi-class logistic regression. The softmax layer 2340 may be a generalization of logistic regression that can be used for multi-class classification, and its formula is very similar to the sigmoid function used in logistic regression. The softmax layer 2340 function can be used in a classifier only when the classes are mutually exclusive.

[0172] The Softmax layer 2340 converts the scores into a normalized probability distribution that can be displayed to the user or used as input to other systems. The Softmax layer 2340 is the final layer of the neural network 2100 and can produce an output 1820 that includes His bundle and non-His bundle probabilities.

[0173] By directly feeding the distance 1840 as an input to the network 2100, training may require additional time.

[0174] 24 shows an implementation 2400 of the system as described. The implementation 2400 includes a series of inputs to the network 2100, including an ECG input 1830, to generate an output 1820 including a His bundle probability and a not-His bundle probability. As mentioned above, the ECG input 1830 includes a first ECG 18301, a second ECG 18302, ..., a final ECG 1830. N The ECG data may include any number of ECG data, including:

[0175] The network 2100 may be, for example, a CNN network as described herein. For simplicity, the network 2100 may include multiple convolutional layers 2310 interconnected with multiple pooling layers 2320, and further interconnected with multiple flattening layers 2330, which may include one or more resnets and / or fully connected layers. As will be appreciated, the softmax layer 2340 may be the last layer in the network 2100.

[0176] The ECG input 1830 may be provided within the network 2100 terminating before the Softmax layer 2340 .

[0177] A distance from catheter input 1840 may be provided as an input separate from the ECG input 1830. After multiplying by a weight and adding a bias 2420, the distance input 1840 may be provided to an activation function 2430. If d is the distance, b is the bias, w is the weight, and f is the activation function, the output of the activation function is f(wd+b). As understood in the art, the activation function 2430 defines the output of that node given an input or set of inputs. Examples of activation functions 2430 include sigmoid, hyperbolic tangent, ELU, and LeakyReLU. The input of the softmax layer 2340 may be multiplied by the output of the activation function 2430. The softmax layer 2340 converts the scores into a normalized probability distribution that can be displayed to a user or used as input to other systems. Softmax layer 2340 is the final layer of neural network 2100 and can generate output 1820, which includes His bundle and non-His bundle probabilities. Alternatively, instead of softmax layer 2340 taking two inputs and providing two outputs, a single output between -∞ and +∞ can be generated by neural network 2330, which can be converted to a probability between 0 and 1 by an activation function such as a sigmoid. In this configuration, the input of the final activation function can be multiplied by the output of activation function 2430.

[0178] 25 shows an implementation 2500 of the system as described. Implementation 2500 includes a series of inputs to network 2100, including ECG input 1830, to generate output 1820 including His bundle probability and not-His bundle probability. As mentioned above, ECG input 1830 includes a first ECG 18301, a second ECG 18302, ..., a final ECG 1830. NThe network 2100 can output a median probability 2520 of the His bundle.

[0179] Network 2100 may be, for example, a CNN network as described herein. For simplicity, network 2100 may include multiple convolutional layers 2310 interconnected with multiple pooling layers 2320, and further interconnected with multiple flattening layers 2330, which may include one or more resnets and / or fully connected layers.

[0180] The distance from catheter input 1840 and other inputs 1850 may be provided after the convolutional layer 2310, the pooling layer 2320, and the flattening layer 2330 of the convolutional network. As shown, all three layers 2310, 2320, and 2330 may be included, as will be appreciated, skipping the distance 1840 and other inputs 1850, or only portions of layers 2310, 2320, and 2330. The distance input 1840 and any other inputs 1850 may be provided to the output of the ECG network 2100 and combined using a non-convolutional neural network with one or more hidden layers 2500 to generate an output 1820 including the His bundle and non-His bundle probabilities. In this network architecture, because the inputs are scalars, e.g., distance and applied force, the inputs are treated as "just another input," which may cause the network to take too long to train and make it difficult to converge. The configurations described in Figures 24 and 25 solve this problem.

[0181] Although two networks 2100, 2510 are shown, these networks are trained monolithically as a single network. Assume two locations within the heart are to be located, one being the desired location (the His bundle) and the other simply an arbitrary point at some distant location within the heart. For the sake of this example, assume that the ECG signals received from these two locations are very similar, and the only way to understand which is the His bundle is by looking at the distance. The first neural network is not provided with information about distance because distance is not an input. If the first network were trained alone, the two very similar signals would need to be fed to the neural network once as the His bundle and once as "not the His bundle," causing confusion within the neural network. The neural network would not converge.

[0182] In this example, two very similar signals were included to provide a better understanding. The neural network can learn some "attributes" of the signals to determine the His bundle, and these attributes may exist in different locations, so even with different signals, the same problem would still exist. A human would understand the difference, but a neural network that receives an ECG as its only input would be confused and would have difficulty converging. In the best case, training would take too long, and in the worst case, the network would not converge.

[0183] When training the two networks as a monolithic whole, the first network has the freedom to give a "garbage" output if the distance is too large and still converge. In Figure 24, multiplication can reduce the influence of "garbage" if the distance is too large. In Figure 25, the second neural network can reduce the influence of "garbage" if the distance and other parameters indicate a low probability of His bundles. Also, unlike handcrafted algorithms, this network learns what the relevant distances are. This network can learn that if the distance is large enough, the activation function 2430 should output a coefficient "near zero." A "near zero" coefficient can cause the second network to output a low probability of His bundles, regardless of the output of the first network.

[0184] In general, cutoffs do not help the results. For example, if ECG signals whose distance is less than 1 cm are considered, all other signals are filtered out during the training phase. In theory, such a configuration and cutoff would provide the same "fast convergence" advantage, which would also solve the problem. The effect of distance should be a continuous function, not a discrete function. The effect of the architectures in Figures 24 and 25 is as if the neural network learns to construct a function that represents the effect of distance. During training, the network gradually de-emphasizes inputs that are farther away. This solution can be extended to any additional scalar input, such as applied force, tissue proximity indicators, etc.

[0185] In Figure 25, the first network can output more outputs to indicate different attributes of the signal, and the second network can be trained to combine the attributes with the distance. For example, if the distance is small, the second network can be trained to give more importance to attributes such as attributes A, B, and C. If the distance is large, the second network can be trained to give more importance to attributes such as attributes D and E.

[0186] The first portion of network 2100 may or may not terminate in a softmax or sigmoid layer. If it does terminate in a softmax or sigmoid layer, the input to the second portion of the network is between 0 and 1. Otherwise, the input to the second portion of the network is between -∞ and +∞.

[0187] Although automatic detection of the His bundle is described herein as a result of utilizing the neural networks described herein, the presently disclosed subject matter is not limited to automatic detection of the His bundle. Automatic identification of other cardiac structures and / or signals is within the scope of the disclosed subject matter. For example, the cardiac cycle begins with the sinoatrial (SA) node, which transmits electrical impulses through the atria and via the atrioventricular (AV) node to the His bundle. The His bundle transmits electrical impulses from the AV node to the left and right bundle branches, which then transmit the electrical signal to the Purkinje fibers, which provide the ventricles. In one embodiment, the presently disclosed subject matter can be used to automatically detect other cardiac structures in the cardiac cycle, including, but not limited to, the SA node, left and right bundle branches, Purkinje fibers, etc. Additionally, as another example, the presently disclosed subject matter can be used to detect LAVA signals, as discussed above. In another embodiment, the presently disclosed subject matter can be used to detect the position of a catheter and can serve as a warning system for a physician to detect if a catheter has unintentionally moved from one atrium to another.

[0188] Although features and elements are described above in particular combinations, those skilled in the art will understand that each feature or element can be used alone or in combination with other features and elements. Additionally, although the process steps are described above in a particular order, the steps can be performed in any other desired order.

[0189] The methods, processes, modules, and systems described herein can be implemented in a computer program, software, or firmware embodied in a computer-readable medium for execution on a computer or processor. Examples of computer-readable media include electronic signals (transmitted via wired or wireless connections) and computer-readable storage media. Examples of computer-readable storage media include, but are not limited to, read-only memory (ROM), random-access memory (RAM), registers, cache memory, semiconductor memory devices, magnetic media such as internal hard disks and removable disks, magneto-optical media, and optical media such as CD-ROM disks and digital versatile disks (DVDs). A processor in association with software can be used to implement a radio frequency transceiver for use in a WTRU, UE, terminal, base station, RNC, or any host computer.

[0190] Further embodiments of the present specification may be formed by adding to an embodiment one or more elements from any one or more other embodiments of the present specification and / or by substituting one or more elements from an embodiment with one or more elements from one or more other embodiments of the present specification.

[0191] It is understood, therefore, that the disclosed subject matter is not limited to the particular embodiments disclosed, but is intended to encompass all modifications that are within the spirit and scope of the present disclosure as defined by the appended claims, the above description, and / or as illustrated in the accompanying drawings.

[0192] [Embodiment] (1) A system for automatically detecting cardiac structures, comprising: a plurality of sensing devices positioned within the heart to receive electrophysiological data related to a first cardiac structure, each sensing device providing a one-dimensional signal; A processor including a neural network, receiving the one-dimensional signal from at least one of the plurality of sensing devices; receiving distance data relating to a distance between a pair of the plurality of sensing devices; applying the neural network to the received one-dimensional signal to determine an output; applying weights and biases to the distances; applying an activation function to the weighted and biased distances; multiplying the determined output by the output of the activation function to determine whether the first cardiac structure is the cardiac structure of interest based on the electrophysiological data and the distance data. a processor; A system comprising: (2) The system of embodiment 1, wherein the cardiac structure of interest includes the His bundle. (3) The system of embodiment 1, wherein the electrophysiological data regarding the first cardiac structure received by the plurality of sensing devices includes an ECG signal. (4) The system of embodiment 1, wherein the plurality of sensing devices comprises a plurality of different electrodes. (5) The system of embodiment 1, wherein the distance data is the distance of the mapping electrodes.

[0193] (6) The system of embodiment 1, wherein the neural network is a convolutional neural network or a recurrent neural network. (7) The system of claim 1, further comprising monolithically training the neural network on the one-dimensional signal and the range data. (8) A method for automatically detecting cardiac structures, comprising: receiving electrophysiological data comprising a plurality of one-dimensional signals via a plurality of sensing devices positioned within the heart; receiving, via a neural network, the one-dimensional signal from at least one of the plurality of sensing devices and distance data relating to a distance between a pair of the plurality of sensing devices; applying the neural network to the received one-dimensional signal to determine an output; applying weights and biases to the distances and applying an activation function to the weighted and biased distances; multiplying the determined output by the output of the activation function to determine whether the first cardiac structure is the cardiac structure of interest based on the electrophysiological data and the distance data; A method comprising: (9) The method of embodiment 8, wherein the cardiac structure of interest includes the His bundle. (10) The method of embodiment 8, wherein the electrophysiological data regarding the first cardiac structure received by the plurality of sensing devices includes an ECG signal.

[0194] (11) The method of embodiment 8, wherein the plurality of sensing devices comprises a plurality of different electrodes. (12) The system of embodiment 1, wherein the distance data is the distance of a mapping electrode. (13) The system of embodiment 1, wherein the neural network is a convolutional neural network or a recurrent neural network. (14) The system of claim 1, further comprising monolithically training the neural network on the one-dimensional signal and the range data. (15) A system for automatically detecting cardiac structures, comprising: a plurality of sensing devices positioned within the heart to receive electrophysiological data related to a first cardiac structure, each sensing device providing a one-dimensional signal; a processor including a first neural network that receives the one-dimensional signal from at least one of the plurality of sensing devices and applies the neural network to the received one-dimensional signal to determine an output; a second neural network, the processor configured to receive a plurality of scalar values relating to distances between at least one pair of the plurality of sensing devices, apply weights and biases to the plurality of scalar values, and apply an activation function to the weighted and biased distances; The system combines the output of the first neural network and the second neural network to determine whether the first cardiac structure is the cardiac structure of interest based on the electrophysiological data and the distance data.

[0195] (16) The system of embodiment 15, wherein the cardiac structure of interest includes the His bundle. (17) The system of embodiment 15, wherein the electrophysiological data regarding the first cardiac structure received by the plurality of sensing devices includes an ECG signal. (18) The system of embodiment 15, wherein the plurality of sensing devices comprises a plurality of different electrodes. (19) The system of embodiment 15, wherein the distance data is the distance of a mapping electrode. (20) The system of embodiment 1, further comprising monolithically training the first neural network and the second neural network on the one-dimensional signal and the range data.

Claims

1. 1. A system for automatically detecting cardiac structures, comprising: a plurality of sensing devices including a plurality of electrodes positioned within the heart to receive electrophysiological data relating to cardiac structures including the region of the sinoatrial node, the atrioventricular (AV) node and bundle of His, myocardial tissue forming the walls of the ventricles and atria, and / or the endocardium, each of which provides EGM or ECG signals constituting the electrophysiological data; A processor including a neural network and a CPU, the CPU receives the EGM or ECG signal from at least one of the plurality of sensing devices; the CPU receiving distance data relating to a distance between a pair of the plurality of sensing devices; the neural network outputs a score indicative of whether the cardiac structure is of interest in response to the received EGM or ECG signal; the CPU applies an activation function to the weighted and biased distance inputs of the distance data to generate an output of the activation function; the CPU multiplies the score by the output of the activation function to determine whether the cardiac structure is a cardiac structure of interest for being a site of abnormal electrical activity based on the electrophysiological data and the distance data. a processor; A system comprising:

2. The system of claim 1 , wherein the cardiac structure of interest includes the bundle of His.

3. The system of claim 1 , wherein the distance data is data indicating a distance between a pair of the electrodes of the plurality of sensing devices.

4. The system of claim 1 , wherein the neural network is a convolutional neural network or a recurrent neural network.

5. The system of claim 1 , further comprising monolithically training the neural network on the EGM or ECG signals and the distance data.

6. 1. A method of operating a system for automatically detecting cardiac structures, comprising: a CPU of a processor of the system receiving electrophysiological data including a plurality of EGM or ECG signals; receiving, by the CPU, distance data relating to a distance between a pair of a plurality of sensing devices, the plurality of sensing devices including a plurality of electrodes; a neural network of the processor outputting a score indicative of whether the cardiac structure is of interest in response to the received EGM or ECG signal; the CPU applying an activation function to the weighted and biased distance inputs to generate an output of the activation function; the CPU multiplying the score by the output of the activation function to determine, based on the electrophysiological data and the distance data, whether the cardiac structures, including the region of the sinoatrial node, the atrioventricular (AV) node and bundle of His, myocardial tissue forming the walls of the ventricles and atria, and / or the endocardium, are of interest as sites of abnormal electrical activity; A method of operation comprising:

7. The method of claim 6 , wherein the cardiac structure of interest includes the bundle of His.

8. The operating method according to claim 6 , wherein the distance data is data indicating a distance between a pair of the electrodes of the plurality of sensing devices.

9. 7. The method of claim 6, wherein the neural network is a convolutional neural network or a recurrent neural network.

10. The method of claim 6 further comprising monolithically training the neural network on the EGM or ECG signals and the distance data.

11. 1. A system for automatically detecting cardiac structures, comprising: a plurality of sensing devices including a plurality of electrodes positioned within the heart to receive electrophysiological data relating to cardiac structures including the region of the sinoatrial node, the atrioventricular (AV) node and bundle of His, myocardial tissue forming the walls of the ventricles and atria, and / or the endocardium, each of which provides EGM or ECG signals constituting the electrophysiological data; a first neural network, a second neural network, and a processor including a CPU; the CPU receives the EGM or ECG signal from at least one of the plurality of sensing devices; and outputting a score from the first neural network in response to the received EGM or ECG signal indicating whether the cardiac structure is of interest; the CPU receiving a plurality of scalar values relating to distances between at least one pair of the plurality of sensing devices; the second neural network applies an activation function to the weighted and biased distance inputs of the plurality of scalar values to generate an activation function output; The second neural network multiplies the score by the output of the activation function to output a probability of whether the cardiac structure is a cardiac structure of interest, based on the electrophysiological data and the plurality of scalar values, as to whether the cardiac structure is a site of abnormal electrical activity.

12. The system of claim 11 , wherein the cardiac structure of interest includes the bundle of His.

13. The system of claim 11 , wherein the plurality of scalar distance values are values indicating a distance between a pair of electrodes of the plurality of sensing devices.

14. 12. The system of claim 11, further comprising monolithically training the first neural network and the second neural network on the EGM or ECG signal and the plurality of scalar values.

Images