Method and system using deep learning to identify unresponsive individuals in pulmonary vein isolation

Abstract

This invention provides a computer system and computer-aided method for the treatment of atrial fibrillation using pulmonary vein isolation. [Solution] The system trains a set of machine learning models to predict which atrial fibrillation patients will respond to which treatments using ablation lines. The models identify atrial fibrillation-related features associated with the patient and estimate the probability of reconnection of ablation sites associated with at least one ablation line treatment. The models identify patient-associated atrial fibrillation predictors based on patient-associated demographics and cardiac component dimension parameters, and use at least one ablation line treatment, reconnection probability, patient-associated atrial fibrillation predictors, and patient demographics to predict whether the patient will not respond to pulmonary vein isolation treatment alone.

Description

[Technical Field]

[0001] This disclosure relates to a computer system and computer-administered method for the treatment of atrial fibrillation using pulmonary vein isolation. [Background technology]

[0002] Atrial fibrillation (AF) is the most commonly diagnosed persistent arrhythmia, characterized by rapid and irregular activation of the atria. Atrial fibrillation can be paroxysmal, lasting less than 7 days with or without intervention, or it may persist for more than 7 days (persistent, PS-AF) or more than 12 months (long-standing, LS-PS-AF). Persistent atrial fibrillation is a term used for long-standing persistent atrial fibrillation when attempts to restore sinus rhythm have been abandoned or proven impossible. See, for example, Kaba, RA, Momin, A. & Camm, J. Persistent atrial fibrillation: The role of left atrial posterior wall isolation and ablation strategies. J. Clin. Med. 10, (2021).

[0003] Treatment strategies vary depending on the severity and duration of atrial fibrillation. Pulmonary vein isolation (PVI) is a medical procedure aimed at restoring normal rhythm and reducing or eliminating symptoms associated with atrial fibrillation. For example, paroxysmal atrial fibrillation is typically managed by wide antral circumferential ablation (WACA), i.e., by delivering radiofrequency energy in a site-by-site manner along a predefined line of the atrium. However, the majority of patients with atrial fibrillation have PS-AF or LS-PS-AF, which are more difficult to treat, likely due to a more widespread distribution of arrhythmic triggers and substrates outside the pulmonary veins. In patients with advanced cases of atrial fibrillation, pulmonary vein isolation alone may not be sufficient, and long-term failure of "pulmonary vein isolation alone" is observed. The posterior wall of the left atrium is a common site of these changes and is a target for ablation strategies to treat these more resistant forms of atrial fibrillation. Other common parts include the left and right sides of the carina and the roof. [Brief explanation of the drawing]

[0004] [Figure 1] Under one embodiment, a block diagram of an exemplary system for using deep learning to identify unresponsive individuals in pulmonary vein isolation is shown. [Figure 2] Under one embodiment, a block diagram of the inputs and outputs of an exemplary system for using deep learning to identify unresponsive individuals in pulmonary vein isolation is shown. [Figure 3A] This is a block diagram of a deep learning system for identifying unresponsive individuals in pulmonary vein isolation under one embodiment. [Figure 3B] This is a block diagram of a useful deep learning structure in the system shown in Figure 3A, according to one embodiment. [Figure 4A] In one embodiment, this is an exemplary block diagram of the left atrium for a deep learning system for identifying unresponsive individuals with pulmonary vein isolation. [Figure 4B]In one embodiment, this is an exemplary block diagram of the left atrium for a deep learning system for identifying unresponsive individuals with pulmonary vein isolation. [Figure 4C] In one embodiment, this is an exemplary block diagram of the left atrium for a deep learning system for identifying unresponsive individuals with pulmonary vein isolation. [Figure 4D] In one embodiment, this is an exemplary block diagram of the left atrium for a deep learning system for identifying unresponsive individuals with pulmonary vein isolation. [Figure 5] Under one embodiment, a block diagram of wall dimension calculation for using deep learning to identify unresponsive individuals in pulmonary vein isolation is shown. [Figure 6] Under one embodiment, a block diagram of a voltage map with ablation lines is shown for using deep learning to identify non-responsive individuals in pulmonary vein isolation. [Figure 7] Under one embodiment, a block diagram of a vector velocity field for using deep learning to identify unresponsive individuals in pulmonary vein isolation is shown. [Figure 8] Under one embodiment, a block diagram of bipolar window selection for using deep learning to identify unresponsive individuals in pulmonary vein isolation is shown. [Figure 9] Under one embodiment, a block diagram of a bipolar window with different waveforms for using deep learning to identify unresponsive individuals in pulmonary vein isolation is shown. [Figure 10] Under one embodiment, a pulsation-annotated block diagram is shown for using deep learning to identify unresponsive individuals in pulmonary vein isolation. [Figure 11] This flowchart shows a computer-aided method for using deep learning to identify unresponsive individuals in pulmonary vein isolation under one embodiment. [Figure 12] A block diagram showing exemplary hardware devices on which the subject may be implemented. [Modes for carrying out the invention]

[0005] The following detailed description includes many specific details to help you fully understand an example of the present invention. However, it will be understood by those skilled in the art that the present invention can be implemented in countless other examples or embodiments within the scope of the amended claims without following the specific details of the particular exemplary examples provided herein. In other cases, well-known methods, procedures, and components are not described in detail so as not to obscure the present invention.

[0006] Embodiments of this specification enable the use of deep learning to identify non-responders of pulmonary vein isolation, as described. The system trains a set of machine learning models to predict which atrial fibrillation patients will respond to which treatments using ablation lines. The models identify a patient's ablation line treatments based on atrial fibrillation-related features associated with the patient. The models estimate the reconnection probability of the ablation site associated with at least one ablation line treatment. The models identify patient-associated patient atrial fibrillation predictors based on patient-associated demographics and cardiac component dimension parameters, and use at least one ablation line treatment, reconnection probability, patient atrial fibrillation predictors, and patient demographics to predict whether the patient will not respond to treatment with pulmonary vein isolation alone. In response to the prediction that the patient will not respond to treatment with pulmonary vein isolation alone, the system enables healthcare providers to offer the patient at least one ablation line treatment.

[0007] For example, the system trains two graph convolutional neural networks and three high-density layers within the neural networks to predict which atrial fibrillation patients will respond to which atrial fibrillation line treatment. The set of machine learning models extracts a map from specific intracardiac electrocardiogram data of a 61-year-old patient named "Pat," and uses the map data to identify treatments involving six ablation lines containing 33 ablation sites in the right atrium. The set of machine learning models then uses the ablation sites to estimate the probability of reconnection of the ablation sites. The set of machine learning models receives demographic data and right atrial fibrillation parameters of patient Pat, which are used collectively to predict that patient Pat's atrial fibrillation will progress. The machine learning model set uses six ablation lines containing 33 ablation sites within the right atrium, the highest estimated reconnection probability of previous ablation sites, and Pat's atrial fibrillation predictor. Pat's atrial fibrillation predictor is used collectively to predict that Pat's atrial fibrillation will not respond to treatment using pulmonary vein isolation alone, and that Pat's atrial fibrillation will respond to treatment with six ablation lines within Pat's right atrium. The system outputs these predictions to Pat's physician, which may allow for changes to the treatment Pat was about to receive.

[0008] The technology described herein relates to a system that uses a machine learning model to predict whether a patient with atrial fibrillation will be a responder to treatment using pulmonary vein isolation alone. This prediction is important for optimizing patient outcomes and tailoring treatment strategies. The system processes and trains on a multifaceted dataset that may include patient outcomes (follow-up) at X months, patient demographics, medical history, medical imaging data such as specific intracardiac electrocardiogram (ECG) data, voltage maps / cases, and ground truth regarding signal parameters and mapping analysis such as atrial cycle length + / - cycle length mapping of sites that meet catheter stability filters, ablation quality predictors, ablation reconnection, voltage, atrial dimensions and their relationships, atrial cycle length, signal fractions, and ablation reconnection probability. This approach involves several stages, including data preprocessing, deep learning model training, and post-processing of model proposal, recommendation, estimation, and prediction. As used herein, the term “long-term success” means a period of 3 to 24 months following treatment that is free from atrial fibrillation and any arrhythmias (“atrial fibrillation and arrhythmia-free period”).

[0009] Identification of ablation site System 100 receives cardiac mapping data 102, where the mapping data includes a three-dimensional object of the left atrium and feature embeddings, and the feature embeddings are generated using anatomical images produced by using an appropriate medical imaging system or technique such as electrogram (EGM), body surface ECG, intracardiac ECG, and / or baseline recording. The mapping data 102 includes both position information of a catheter moving around the atrium and electrocardiogram measurements such as voltage at each position. The system 100 determines locations for ablation by using electroanatomical maps 104 such as a segmentation map, a PVC-based delayed potential map extracted during premature ventricular contraction beats, a delayed potential map, a bipolar voltage map, a conduction velocity map, a map of activation types (fractionated, double potential, normal potential), a divergence map, and a rotation map. The extraction of each map is discussed in detail below in the sections describing the maps of vector velocity field divergence, conduction velocity, and activation type. The system 100 includes a matrix 106 that can represent all different maps, such as an N×M size feature matrix; in this example, M=9, since the system 100 typically has 6 different maps, the last 3 columns are defined by the (x, y, z) positions of the vertices, and the corresponding adjacency matrix has a size of N×N.

[0010] For implementations, a graph convolutional neural network may be employed, where the feature space is of size N×M and processed through a set of graph convolutional neurons. The system 100 can generate an output layer using softmax 108, such as APPNP (Approximate Personalized Propagation of Neural Predictions) softmax propagation = 1, 3, and 5, and ARMA softmax with 21 channels. This architecture has been successfully used for deep learning processing of three-dimensional objects. The output layer of the network is a 6000×1 vector 110, which represents the probability of ablating a given vertex of the patient's heart.

[0011] An exemplary training set includes 60,000 cases (persistent cases and paroxysmal cases) from CartoNet, with ablation data for each case. The target vector for the algorithm is a 6000×1 size map with a value of 0 or 1 for each vertex, where 1 represents the vertex that was ablated in the case due to extra targets. To obtain this map, pulmonary vein isolation ablation sites + carina sites are filtered, the remaining ablation sites are projected onto the anatomy, any vertex (configurable) with a distance of less than 4mm is marked as 1, and all other vertices are marked as 0. 80% of the cases are used for training, 10% of the cases are used for validation, and 10% of the cases are used to test the performance of the algorithm. A binary cross-entropy loss function can be used to train this neural network.

[0012] The output of this neural network is further input into a set of dense layers 112 that output ablation line predictions 114, such as 13 types of ablation lines, and since more than one ablation line can be predicted by the neural network, mean squared error (MSE) loss can be used to predict the ablation lines. The same set of 60,000 cases can be used to train this neural network, but the target vector for each case is a 13×1 vector, where a 1 in this vector indicates that this ablation line was performed in the case. Using these techniques, the following ablation lines can be predicted: Left wide area circumferential ablation (WACA), Right wide area circumferential ablation, Left carina, Right carina, Roof line, Anterior line, Inferior line, Posterior line, Mitral valve line, Posterior wall debulking, Lateral wall debulking, Lower wall debulking, and Other lines are available as needed.

[0013] A machine learning system can be trained to identify a large number of many different types of patients who have atrial fibrillation and for whom the ablation line procedure has a higher probability of long-term success, provided that the trainer of the machine learning model has records of a sufficient number of different types of patients for whom the ablation line procedure has resulted in long-term success. If an insufficient number of records of various types of patients who have experienced long-term success with the ablation line procedure are available, it is proposed to use an alternative method to overcome the technical setback resulting from insufficient training data. This alternative approach is based on the assumption consistent with the "wisdom of crowds" (WoC) or "hive brain" principle, which states that in large numbers (e.g., including hundreds of thousands, millions, or more unrelated cases), the majority of individuals will make the correct decision. The use of physician recommendations when training machine learning models addresses the technical challenges of data scarcity related to actual outcomes in atrial fibrillation treatment, leveraging expert clinical judgment as a reliable substitute for outcome data, thereby facilitating the generation of robust and effective machine learning models despite limited availability of outcome data.

[0014] Continuity estimator Pulmonary vein isolation can be performed using radiofrequency energy in a site-specific WACA pattern (e.g., WACA procedure) with an ablation catheter, the endpoint of which is complete pulmonary vein isolation. After the final ablation of the WACA procedure, the physician can place a catheter at the ablation site to determine if spontaneous reconnection has occurred. If pulmonary vein reconnection does not occur, an intravenous adenosine bolus is administered to mask any site of dormant conduction. If necessary, ablation may be performed at any site of reconnection to achieve pulmonary vein isolation. In a study with 400 cases, ablation sites to which the physician added "modifications" were identified as continuity sites, and an appropriate network (e.g., system 100) was trained to estimate the ablation sites in the original WACA procedure near the modified sites as areas with potential reconnection sites. After the physician completes the ablation procedure (or performs ablation on a particular ablation line), system 100 analyzes real data from the ablation device to assess the probability of reconnection. By integrating data from demographics, electroanatomical mapping, and specific characteristics of the ablation sites performed, the model can update its predictions of whether a patient will ultimately respond to treatment. The dataset can be split into training, validation, and test sets in 80%, 10%, and 10% proportions. Binary cross-entropy can be used as the loss function for training the neural network.

[0015] The continuity input 116 may include the location (x, y, z) of the ablation site, as well as attributes of the ablation site such as impedance drop, high-frequency index, power, and force, and may be accessed by the sampler 118 used to generate the continuity matrix 120. The continuity matrix 120 may include a K×M continuity feature matrix, where K represents the number of ablation sites in the case, and may also include a K×K adjacency matrix representing the Euclidean distance between each of the ablation sites. The continuity matrix 120 is input to a graph convolutional neural (GCN) network layer, which may be implemented by a continuity softmax 122 such as APPNP and AMRA softmax, to generate a KX1 continuity probability 124, which stores 1 for each vertex of the heart predicted to require reconnection investigation and 0 for all other vertices of the heart. More detailed information regarding continuity is available in the following section titled "Continuity Estimation".

[0016] Continuous values ​​of demographic data such as age, body mass index (BMI), and left atrial (LA) diameter 126 serve as features of a high-density layer network 128. In addition to the left atrial predictor for atrial fibrillation, categorical variables such as sex, use of specific medical history, and presence of hypertension can be converted into one-hot coded vectors and serve as input to the high-density layer. The AF feature matrix 128 can include a predictor for atrial fibrillation, which is a set of 114 parameters calculated from three-dimensional representations of the heart, cycle length, and voltage maps. As the atrial fibrillation condition progresses, the morphology of the left atrial changes from tubular to spherical, and its volume expands. Evidence suggests a consistent association between left atrial volume expansion and worsening of both the severity and chronic nature of atrial fibrillation in affected patients. Therefore, factors such as atrial sphericity, low-voltage areas, and cycle length across the seven anatomical structures of the left atrium (posterior wall, anterior wall, left lateral wall, right lateral wall, roof, mitral valve, and inferior wall) may help predict atrial fibrillation.

[0017] System 100 includes a final high-density layer 130 that receives inputs from ablation line predictions 114, continuity probabilities 124, demographic data 126, and an AF feature matrix 128, and can also directly receive surface ECG and intracardiac ECG, signals, and parameters 132. Depending on how System 100 has been trained, the final high-density layer 130 can generate a variety of outputs from these various inputs. For example, if System 100 has been trained only on patients with pulmonary vein isolation only, System 100 can identify the atrial fibrillation features of the current patient that are similar to some of the atrial fibrillation features of these historical pulmonary vein isolation only patients, and can generate an output 134 that includes a prediction of whether the existing patient who is a candidate for pulmonary vein isolation only is likely to respond positively to pulmonary vein isolation only, or whether the patient will require pulmonary vein isolation plus treatment instead. In another example, after system 100 has been trained for all atrial fibrillation patient types with long-term outcomes, output 136 may include recommended ablation line treatments for current patients with specific atrial fibrillation characteristics, based on patients whose atrial fibrillation characteristics are similar and who have successfully undergone pulmonary vein isolation plus treatment. In yet another example, after system 100 has been trained for all atrial fibrillation patient types with long-term outcomes, output 138 may predict the probability of success regardless of the ablation approach.

[0018] Figure 2 shows the inputs and outputs of a deep learning system 200 for identifying unresponsive patients in pulmonary vein isolation, which includes an input patient database 202 that provides data to 204-216, generating outputs 216-220. The patient's demographics and medical history 204 may be similar to the demographic data 126 in Figure 1, and CARTO® mapping features 206, as well as the anatomical structures of the cardiac chambers and any of their subparts, which may include voltage LVZ, global voltage, and other calculations for voltage data, in addition to cycle length and its different calculations, anatomical surface, volume, length, height, and width, and the ratios between them, all of which may be similar to the mapping data 102 and map 104 in Figure 1. The electrical signals 208 (surface and intracardiac electrocardiograms) and their processing may be similar to the signals and parameters 132 in Figure 1, ablation parameters 210 which may include baseline impedance, impedance drop, CF, power, time, ablation index, catheter stability, inter-lesion distance, temperature, etc., and may be similar to the continuity input 116, ablation approach 212, and long-term treatment outcome 214 in Figure 1. The predicted output for a new patient 216 may be similar to the ablation line prediction 114 in Figure 1, based on the predicted output 218 of the physician-selected ablation approach for pulmonary vein isolation only or pulmonary vein isolation plus, and the predicted ablation approach output 220 may be for pulmonary vein isolation only or pulmonary vein isolation plus, according to the outcome, and may be similar to the ablation line prediction 114 in Figure 1. By adding the input of the long-term treatment outcome 212, the system 100 changes the predicted output 214 of the physician-selected ablation approach to the long-term treatment outcome 216.

[0019] Deep learning systems can be used to plan ablation lines, such as wide parasinus ablation (WACA) lines, and to check how closely the lines actually performed on the left atrium match the planned lines. In addition, the problem of defining ablation lines is a segmentation problem. Therefore, the input to such a system may be an anatomical map of the patient's heart, and the same type of deep learning structure may be used for both automatically segmenting the anatomical map and defining the locations of the ablation lines.

[0020] Herein, refer to Figure 3A showing the deep learning system 310 for the title. The deep learning system 310 includes three sections: an ablation line trainer 312 that constructs trained deep learning units for an ablation line proposer 313; an ablation line recommender 314 that generates ablation line recommendations using the ablation line proposer 313; and an ablation line guider 316 that provides ablation line guidance during the procedure. The ablation line trainer 312 may receive an anatomical map of the heart, which may be generated using anatomical images produced by using a preferred medical imaging system 320, or using fast anatomical mapping (FAM) technology available in the CARTO® system manufactured by Biosense Webster Inc. (Irvine, Calif.), or using any other preferred technology, or any preferred combination of the above. The map / model of the heart on which recommended ablation lines may be rendered may include any preferred type of three-dimensional (3D) anatomical map generated using any preferred technology. The map can be obtained from an intracardiac electrocardiogram (ECG) and may include both positional information of the catheter moving around the atria and ECG measurements such as voltage at each position.

[0021] The ablation line trainer 312 may include an automated map segmenter 322 that segments a three-dimensional map of the heart received from a medical imaging system 320 into various parts such as the posterior and anterior walls, superior wall, inferior wall, left lateral wall, septum, mitral valve, left atrial appendage (LAA), and four pulmonary veins (LSPV, LIPV, RSPV, RIPV), and an ablation line neural network trainer 324 that examines thousands of ablation cases on different maps received from the imaging system 320, along with their different ablation strategies, to generate trained deep learning units for the ablation line proposer 313. The ablation line recommender 314, including the ablation line proposer 313, provides an initial recommendation for a typical ablation line during the procedure, and the recommended line(s) are rendered on an anatomical structure, such as a model of the heart segmented by an automated map segmenter 322 and / or an anatomical map, thereby allowing the physician to ablate as close as possible to the proposed line, and thus resulting in an ablation line that is as close as possible to the ablation line preferred by the key opinion leader (KOL) in the field. However, it should be noted that the ablation line proposer 313 may display the recommended ablation lines described herein together with the ablation line indicator 325 on any digital depiction of an anatomical structure, whether it be an anatomical map or an ultrasound image, including a four-dimensional (4D) ultrasound image. Therefore, as used herein, “map” of an anatomical structure refers to any such digital depiction on which the ablation lines can be projected.

[0022] The ablation line proposer 313 may allow a physician to specify the desired ablation line, which may be a known ablation line such as WACA, roof line, carina line (right, left), posterior line, anterior line, inferior line, mitral valve line, posterior wall debulking, anterior wall debulking, inferior wall debulking, left lateral debulking, or LAA isolation. The ablation line recommender 314 may also include an ablation line editor 327 to allow a physician to review and modify the selected ablation line according to the physician's requirements, preferences, and findings in the case. For example, a physician may want to edit the initial ablation line so that the initial ablation line includes a trigger observed in the posterior wall.

[0023] The ablation line proposer 313 may segment the WACA circle into sections and recommend an energy delivery per segment, i.e., a recommended ablation index (radio frequency (RF) or irreversible electroporation (IRE)) and / or a recommended distance between adjacent points, which may be predefined for a specific ablation machine such as a CARTO® machine and / or configured by the physician before performing the ablation and / or per ablation line type. The ablation line editor 327 may also allow the physician to update the recommended treatment delivery per segment. The ablation line recommender 314 may require the physician to approve the plan after reviewing all parameters.

[0024] The ablation line guider 316 provides ablation line guidance during the procedure. It can receive an anatomical map of the heart from a medical imaging system 320 during the procedure and may comprise an automated map segmenter 322 for showing various elements of the heart during the procedure, an ablation line guidance unit 330, and a performance reporter 332. During the procedure, based on the catheter position and the relevant ablation line from the approved plan, the ablation line guidance unit 330 may present a dynamic display of the next ablation site based on the previous site and the recommended ablation line (approved or edited by the physician). The next ablation site may then be displayed in the forward direction of the line, along with the minimum distance from the proposed line, based on a predefined distance, e.g., 4 mm from the previous site.

[0025] The objective of an ablation procedure is to form effective lesions that can block abnormal electrical pathways causing arrhythmias. For this purpose, the ablation line guidance unit 330 can calculate an ablation index indicating the quality of each radiofrequency (RF) ablation lesion during the procedure. The ablation index is a function of at least contact force, time, and power, and therefore increases as the ablation operation continues. The instantaneous ablation index provides the physician with real-time feedback, allowing them to monitor and adjust their technique during the procedure.

[0026] Since the physician can define a “required ablation index” for a given ablation line and a given segment within the line, the ablation line guidance unit 330 can present both the “required ablation index” and the instantaneous ablation index for the current site. This may allow the physician to achieve the desired ablation index. The performance reporter 332 can provide the physician with a report based on the physician's performance during the procedure with respect to the ablation plan generated by the ablation line recommender 314. The report may include data such as the maximum and median deviations (mm) from what was planned for each ablation line, deviations from the planned location including the number and percentage of ablation points with distances greater than (or less than) the planned distance, and deviations from the planned ablation for each ablation line. Note: This report may be particularly useful when training less experienced physicians.

[0027] The ablation line recommender 314 may be configured to always draw the recommended WACA (left-right) ablation lines. The same type of deep learning structure may be used, and the same type of input may be provided to both the automap segmenter 322 and the ablation line neural network trainer 324.

[0028] Figure 3B, referenced here, shows a deep learning structure, which will be described in more detail below, and indicates that the inputs are the anatomical map described above and the adjacency map described below. However, Figure 3B shows two types of outputs: segmentation codes for the Automated Map Segmenter 322 and ablation line codes for the Ablation Line Neural Network Trainer 324. The segmentation codes may be a set of codes indicating parts of the heart or parts of the atrium, such as 13 different parts of the atrium. The ablation line codes may be a set of codes indicating the type of ablation line, as well as codes for all regions lacking ablation lines. As described below, each training case may include an input map and an output map, and each location is labeled with one of the following: an adjacency code, a segmentation code for training the Automated Map Segmenter 322, or an ablation line code for training the Ablation Line Neural Network Trainer 324.

[0029] The initial dataset may consist of a sufficient number of paroxysmal and persistent atrial fibrillation cases (e.g., 5000 or more cases) each having an anatomical map of the case and at least one ablation line performed in that case. The ablation lines are typically one of the following: (1) left WACA, (2) right WACA, (3) ostial PVI, (4) roof line, (5) right carina, (6) left carina, (7) posterior line, (8) inferior line, (9) mitral isthmus line, (10) anterior line, (11) inferior vena cava tricuspid annulus (CTI) line, (12) superior vena cava isolation (SVCI), or (13) laryngeal atrial anesthetic (LAA) isolation. To prepare the initial dataset used as input for training and testing the automated map segmenter 322 and the ablation line neural network trainer 324, a specialist (such as a field key opinion leader (KOL)) may first manually annotate the ablation points onto an anatomical map of the case, such as one generated by the medical imaging system 320. Each point is associated with at least one of the ablation lines (1) to (13). Data preprocessing is performed so that each ablation site is rendered on the anatomical map using a specific radius, such as a 5 mm radius.

[0030] In another step, each map is examined, and maps with gaps / or blurry lines may be excluded from the training and test sets. Following this, the atria can undergo further data preprocessing to be transformed into a fixed dimension (all values ​​such as 1000, 2000, 4000, 10000 are configurable by generating a fixed mesh file with a fixed number of points, e.g., 2000 points and a 2000x2000 adjacency matrix). Using feature extraction for 3D general segmentation, a deep learning framework for 3D segmentation can be built, where the graph data is characterized by improving the local features of each node. This is done by vector (N x, Ny, Nz), using spatial features and structural features in the vicinity of one triangle of each mesh surface formed by points having positions (x, y, z) and normals N corresponding thereto. In a triangular mesh, each surface includes at most 3 adjacent surfaces. The 3D mesh structure data is characterized by M={V, E, F} of vertices, edges and faces.

[0031] Therefore, one triangular mesh is converted into a graph G={feature, adjacency}. The faces {F} on the mesh M have 3 corresponding nodes (vertices) on the graph {G}. The adjacency matrix {adjacency} is a 6000×6000 matrix in the case of the present invention. The adjacency matrix represents connections between vertices in M based on {F}. To improve the perception of the local region of graph nodes, spatial and structural geometric features are used based on a single vertex of the simplified mesh (which in this example consists of 6000 vertices). The feature space is a 6000×9 matrix, and for each vertex (a triangle composed of 3 vertices with 3 adjacent faces), the preprocessing utilizes the following features. p i -Normalized (x of vertex Vi in Euclidean space i , y i , z i ) position, where i represents the three indices of the vertex. sp i -Spherical coordinates (φ of vertex i i , θ i , r i ), where φ i is normalized by π, and θ is normalized by 2π. ni-The normal of vertex i defined by (xni, yni, zni).

[0032] The network architecture may include an automated map segmenter 322 that can utilize a machine learning model consisting of two layers, the first layer consisting of six graph convolutional neural network units, and the second layer being a decision layer consisting of another graph convolutional neural (GCN) network having softmax units. Each graph convolutional neural (GCN) network unit gives a prediction of a given vertex to be associated with each of the target categories (anatomical structures or ablation lines).

[0033] As shown in Figure 3B, the anatomical map and the adjacency map of the adjacency matrix serve as inputs to all units in the network. Two types of graph convolutional neural (GCN) networks are used: 1) approximation personalized propagation of neural prediction (APPNP) and 2) auto-regressive moving average layer (ARMA) units, which are highly efficient as softmax layers to increase the overall performance of the network during the training and cross-validation phases, provide a more flexible frequency response, are more robust to noise, and better capture the global graph structure.

[0034] Personalized graph neural networks are described in Predict then propagate: Graph neural networks meet personalized PageRank.Int.Conf.Learn.Represent.(2018) by Gasteiger, J., Bojchevski, A., and Unnemann, SG. Graph CNN neural networks with ARMA filters are described in Graph Neural Networks with Convolutional ARMA Filters (2018) by Maria Bianchi, F., Grattarola, D., Livi, L., and Alippi, C.

[0035] The Approximate Personalized Propagation (APPNP) unit for neural prediction models the relationship between graph convolutional neural (GCN) networks and the PageRank algorithm (originally a link analysis algorithm designed by Google that assigns numerical weights to each node based on the quantity and quality of links pointing to each node) and extends it to personalized PageRank. It uses information from a large, adjustable neighborhood to classify each node. The model is computationally efficient and outperforms several state-of-the-art methods for semi-supervised classification across multiple graphs. Approximate Personalized Propagation (APPNP) for neural prediction has 13 channels with a multilayer perceptron (MLP) having two hidden layers with 128 and 64 neurons in each layer, respectively.

[0036] Network weights are trained using a weighted categorical cross-entropy loss function. The automatic map segmenter 322 can provide segmentation codes to its segmentation loss function, and the ablation line neural network trainer 324 can provide ablation line codes to its ablation line loss function. In an exemplary embodiment, the training set includes 5000 maps. Each map is represented using three matrices, namely a 6000×9 feature matrix, a 6000×6000 adjacency matrix, and a 6000×1 target matrix, one for segmentation codes and one for ablation codes. For ablation codes, 0 represents no lines, and a non-zero represents an ablation line ID.

[0037] Once the neural network of the automated map segmenter 322 is trained, the automated map segmenter 322 may segment the anatomical map of each atrium and provide the segmented map to the ablation line indicator 325. From segmentation to lines, the ablation line proposer 313 can create a concentrated line on the anatomical map with potential locations for point-by-point ablations. The point-by-point ablation "points" may be updated as needed based on the actual ablations performed by the physician during each case. The mechanism for updating the points may involve minimizing the distance between the planned line and the next dot in the forward direction of the ablation line.

[0038] As described, system 310 can use mapping data and automated segmentation data to recommend which of the ablation lines is more suitable for a given case. The recommendation may be based on instructions from a specific research hospital or a specific physician (e.g., a key opinion leader (KOL) in the field of cardiac ablation), or on data from a large number of physicians and / or research hospitals. In an alternative embodiment, the structure used in both the automated map segmenter 322 and the ablation line trainer 324 may include a classifier that operates on intracardiac electrocardiogram (ECG) data. During intracardiac ECG recording, a catheter is inserted into the heart to measure the electrical activity of the heart. The electrical signal for each anatomical location is expressed as a voltage amplitude.

[0039] In this alternative embodiment, the classifier may be any suitable classifier, such as a random forest classifier, a support vector machine (SVM) classifier, or a deep learning classifier. Each case in the training dataset may be preprocessed with an indication of whether one of a given number of ablation lines (e.g., 14) has been performed by a key opinion reader (KOL), so that the target is a matrix of size N × 14, where N represents the number of cases in the training dataset.

[0040] Input data: Electrocardiogram (ECG) data from intracardiac ECG is stored as an additional 6000×22 feature matrix. The following triggers and substrate maps from intracardiac ECG data are used as the feature space. 1) The focal source per vertex of the "finder" (6000 x 1) represents the number of S-waves per second. 2) "Finder" RAP - Each vertex is represented by the number of rotation patterns per second. 3) CL - Cycle Length Mapping - Each vertex is represented by its cycle length in milliseconds. 4) STD CL - Cycle Length Standard Deviation - Each peak is represented by the cycle length standard deviation per second. 5) Spatial and temporal variance - Each vertex represents the proportion of spatial and temporal variance (within 10-30 seconds of the recording). 6) Periodic spatial and temporal variance - Each peak represents the proportion of periodic spatial and temporal variance (within 10-30 seconds of the recording). 7) Voltage map - Each peak represents the bipolar amplitude in mV. 8) CLM ROI - Each vertex represents the existence of a region of interest for cycle length, based on the cutoffs of cycle length < minimum CL + 15 milliseconds and standard CL < 15 milliseconds. 9) The segmentation hot one coding is the output of the automap segmenter 322, a matrix of size (6000 × K), where K represents the number of anatomical structures (13 for LA) (output from the segmentation algorithm). The ablation line coding is the output of the ablation line trainer 324. 10) Global voltage per segmentation region (6000 × 1) mV. 11) Low-voltage zones are areas within the left atrium where the recorded voltage amplitude is relatively low compared to the surrounding area. For example, these may be areas where the voltage is 0.2–0.5 mV. Typically, scarring can be defined as areas where the voltage is less than 0.2 mV. It will be understood that these thresholds may vary based on the operator's decisions in the procedure.

[0041] The classifier may receive a feature matrix as input, which may have dimensions of 6000 × M, where 6000 represents the number of vertices on the map and M represents the number of features (e.g., focus, substrate association, ripple, frequency, anatomical location). The target vector has dimensions of K × 1, where K represents the number of ablation approaches or IDs. In this embodiment, the data may come from successful treatment cases, specifically cases where there have been no arrhythmias for 12 months. For each of these cases, the K × 1 vector is set to 1 if a particular ablation approach was performed during the treatment.

[0042] The classifier's neural network may utilize a set of dense nets, where the final layer consists of a softmax layer with K neurons in the output layer. This modification enables the prediction of K ablation approaches. The systems and methods described herein include, but are not limited to, those described in U.S. Patent No. 11,198,004 or U.S. Patent Application Publication No. 2020 / 0352652(A1), which are owned by the applicant and incorporated herein by reference. It should be understood that actual ablation line data may be noisy, and the ablation lines they represent may not be ideal lines. In another embodiment, the system 310 may further include a keypoint filter 323 for reviewing noisy ablation line data, which may find those ablation points in the ablation line data that can be defined as keypoints (i.e., important points along the ablation line that should not be missed) defined by experts such as associated key opinion leaders (KOLs) based on a segmentation map received from an automated map segmenter 322.

[0043] Here, we refer to Figure 4A, which shows an exemplary left atrium 402 in a posterior-anterior (PA) view, illustrating several exemplary regions such as the superior wall 404 and the posterior wall 406. To its right are the superior right pulmonary vein (RSPV) 408 and the inferior right pulmonary vein (RIPV) 410, and to its left are the inferior left pulmonary vein (LIPV) 412 and the superior left pulmonary vein (LSPV) 414. To construct a deep learning framework for 3D segmentation, the graph data is, for example, a vector (N XAs shown in Figure 4B, which shows a mesh surface formed from point 416 having position (x, y, z) and normal N to it, indicated by Ny, Nz, the local features of each node are improved by utilizing the spatial and structural features of one triangular neighborhood of each mesh surface. Figure 4C shows examples of ablation lines 418 and 420 having potential locations of point-by-point ablations on the anatomical map, such as the point-by-point ablation indicated by point 422.

[0044] Actual ablation line data can be noisy, and the ablation lines they represent may not be ideal. In another embodiment shown in Figure 3A, referenced herein, the system 310 may further include a keypoint filter 323 for reviewing noisy ablation line data, which may find those ablation points in the ablation line data that can be defined as keypoints (i.e., important points along the ablation line that should not be missed) defined by experts such as the associated Key Opinion Reader (KOL), based on a segmentation map received from an automated map segmenter 322. Typically, these keypoints may be points at the intersection of two parts of the atrium.

[0045] This is shown in Figure 4D, which is referenced here. The input 416 to the keypoint filter 323 may be an image or mesh of the left atrium having actual ablation lines, which may have been segmented 418 by the automap segmenter 322. The keypoint filter 323 can detect keypoints 420 by comparing the ablation points of each case to the intersections on the segmented map of that case and selecting the ablation points closest to the intersections. The keypoint filter 323 can combine keypoints from multiple cases to generate the resulting ablation lines 422 for training, which may be thinner and better defined than otherwise. The keypoint filter 323 may provide the resulting ablation lines 422 to the ablation line neural network trainer 324 for training.

[0046] Continuity estimation Pulmonary vein isolation can be performed using radiofrequency energy in a site-specific WACA pattern (e.g., WACA procedure) with an ablation catheter, the endpoint of which is complete pulmonary vein isolation. After the final ablation of the WACA procedure, the physician places a catheter at the ablation site to determine if spontaneous reconnection has occurred. If pulmonary vein reconnection does not occur, an intravenous adenosine bolus is administered to mask any site of dormant conduction, and if necessary, ablation can be performed at any site of reconnection to achieve pulmonary vein isolation again. In a study with 400 cases, sites where physicians added “modifications” as continuity sites were identified, and an appropriate network (e.g., Figure 1) was trained to estimate the ablation sites in the original WACA procedure near the modified sites as regions with potential reconnection sites. The dataset can be divided into training, validation, and test sets in 80%, 10%, and 10% ratios. Binary cross-entropy can be used as a loss function for training neural networks.

[0047] The locations and attributes of the ablation sites may be used to generate a K×M feature matrix, where K represents the number of ablation sites in the case. A K×K adjacency matrix representing the Euclidean distances between each of the ablation sites may also be computed as part of the input to the graph convolutional neural (GCN) network layer.

[0048] The following are examples of features that can be computed for this task: Normalization x:(x-min(x)) / (max(x)-min(x)) Normalized y: (y - minimum(y)) / (maximum(y) - minimum(y)) Normalization z:(z-min(z)) / (max(z)-min(z)) Normalization by maximum range x: (x - minimum(x)) / (maximum range) Normalized by maximum range y: (y - minimum(y)) / (maximum range) Normalization by maximum range z: (z - minimum(z)) / (maximum range)

[0049] The following features are "circular base" features and are used to establish orientation positioning, lower roof, and front-to-back: arctan(x / y) arctan(y / z) arctan(x / z)

[0050] The following set of features is specifically designed to add particular information to both the left-right WACA, front-back, and top-bottom directions. Each ablation site can be normalized based on its relationship to the WACA ring: Average x-ring - Average of the x-values ​​of the rings Average y-ring - Average of the y-values ​​of the rings Average z-ring - Average of the z-values ​​of the rings Normalized x-ring: (x-ring - minimum(x-ring)) / (maximum(x-ring) - minimum(x-ring)) Normalized y-ring: (y-ring - minimum(y-ring)) / (maximum(y-ring) - minimum(y-ring)) Normalized z-ring: (z-ring - minimum(z-ring)) / (maximum(z-ring) minimum(z-ring)) Normalized x-ring by maximum range: (x-ring - minimum(x-ring)) / (maximum range) Normalized y-ring by maximum range: (y-ring - minimum(y-ring)) / (maximum range) Normalized z-ring by maximum range: (z-ring - minimum(z-ring)) / (maximum range) arctan(mean x-ring / mean y-ring) arctan(mean y-ring / mean z-ring) arctan(mean x-ring / mean z-ring) Ring XY - (Range x) * (Range ring y) Ring XZ - (Range x) * (Range ring z) Ring YZ - (Range Ring Y) * (Range ring z) Ring size: [Number of ablation sites within the ring] / [Number of ablation sites]

[0051] The WACA random forest algorithm may be trained on a set of 25 location features and an additional 8 ablation site features obtained from the CARTO® VisiTag export: Duration (total duration of ablation at the ablation site, measured in seconds) average force Maximum temperature [℃] Maximum power [W] Base impedance [Ω] Impedance drop [Ω] Force-Time Integral (FTI) Ablation Index

[0052] The maximum spherical stability can be calculated as the maximum Euclidean distance between the center of the ablation site and all ablation sites associated with that site. The unit of this feature is [mm].

[0053] The mean spherical stabilization velocity can be calculated by dividing the mean Euclidean distance between the center of the ablation site and all associated ablation sites by the duration of the ablation in seconds. The unit of this feature is [mm / second].

[0054] Line stability maxima minimizes the least-squares minimization of the distance from a set of sites to a line in 3D space. Line stability maxima can be calculated as the maximum distance between the ablation site and the estimated least-squares line. The unit of this feature is [mm].

[0055] The average line stabilization velocity may be calculated by dividing the average distance between the ablation site and the estimated least-squares line by the duration of the ablation site. The unit of this feature is [mm / second].

[0056] The maximum effective impedance drop may be associated with a set of ablation sites that are less than 1 mm from the ablation line.

[0057] The maximum high-frequency index—the maximum effective ablation index—may relate to a set of ablation sites less than 1 mm from the ablation line.

[0058] The number of ablation cleavages may be the number of times the ablation index and impedance drop are cleaved (from the tissue, with a value of -10000).

[0059] After the minimum distance, the minimum Euclidean distance between the center of the ablation site and one of the ablation sites that will form the next ablation site may also be used.

[0060] The minimum distance before may be the minimum Euclidean distance between the center of the ablation site and one of the ablation sites that make up the previous ablation site.

[0061] The number of ablation sites that form ablation sites.

[0062] The number of effective ablations within a line can be estimated based on a least-squares model. A site is assumed to be effective if the distance to the ablation line is less than 1 mm.

[0063] The following set of features measures the density of ablation sites associated with a given ablation site: Number of ablation sites within a 3mm radius from the current ablation site Number of ablation sites within a 4mm radius from the current ablation site Number of ablation sites within a 5mm radius of the current ablation site Number of ablation sites with a radius of 6 mm from the current ablation site Number of ablation sites within a 7mm radius from the current ablation site Number of ablation sites within an 8mm radius from the current ablation site Number of ablation sites with a radius of 9 mm from the current ablation site Number of ablation sites within a 10mm radius of the current ablation site Number of ablation sites within a radius of 11 mm from the current ablation site Time difference from the next ablation site Time difference from the previous ablation site

[0064] Demographics and left atrial predictors for atrial fibrillation Continuous values ​​of demographic parameters such as age, body mass index (BMI), and left atrial (LA) diameter can function as features of the high-density layer network. In addition to the left atrial predictor for atrial fibrillation, categorical variables such as sex, use of specific medical history, and presence of hypertension can be converted into one-hot coded vectors and used as input to the high-density layer.

[0065] The left atrial fibrillation predictor is a set of 114 parameters calculated from the three-dimensional representation of the left atrium, cycle length, and voltage map. As atrial fibrillation progresses, the morphology of the left atrium changes from tubular to spherical, and its volume expands. Evidence suggests a consistent association between left atrial volume expansion and worsening of both the severity and chronic nature of atrial fibrillation in affected patients. Therefore, factors such as atrial sphericity, low-voltage regions, and cycle length across the seven anatomical structures of the left atrium (posterior wall, anterior wall, left lateral, right lateral, roof, mitral valve, and inferior wall) may be helpful in predicting atrial fibrillation.

[0066] The left atrial predictor in the atrial fibrillation report displays 114 parameters categorized as follows: 16 parameters in the atrial size parameter group Seven parameters in the atrial globulus parameter group 24 parameters in the low-voltage zone parameter group 36 parameters in the global voltage parameter group 16 parameters in the cycle length parameter group 114 total parameters within the parameter group

[0067] Atrial size parameters An automated segmentation algorithm, such as that disclosed in U.S. Patent Application Publication No. 2023 / 0397950, filed on June 9, 2023, which is incorporated in its entirety herein, can be used to measure the width and height of the seven anatomical structures of the left atrium (posterior wall, inferior wall, anterior wall, superior wall, septal wall, left lateral wall, and mitral valve), as well as the total volume and surface area of ​​the left atrium. The width (or height) of an anatomical structure is mathematically defined as the distance in millimeters between the maximum and minimum points along the x-axis (or y-axis) in a two-dimensional plane. A schematic example is shown in Figure 5, which demonstrates the calculation of the anterior wall dimension using atrial globe size parameters so that the width-to-height ratio of each of the seven structures of the left atrium is displayed based on the atrial size measurement. This approach assumes consistent properties across each of the anatomical structures, meaning that one dimension of the left atrium consistently exceeds the others. For example, the anterior portion is expected to consistently extend forward relative to its width.

[0068] The low-voltage zone parameters may also be calculated as a voltage map based on pre-ablation electroanatomical points that may be generated in the background, identifying three different low-voltage zones: Scar tissue with a voltage of less than 0.2 mV Boundary-zone voltage region with voltages in the range of 0.2 to 0.5 mV A region with a voltage of less than 0.5mV.

[0069] The automated segmentation algorithm quantifies these low-voltage zone regions and expresses them as percentages of eight specific subanatomical structures: namely, the posterior wall, inferior wall, anterior wall, superior wall, septal wall, left lateral wall, mitral wall, left atrial appendage, and the left atrium excluding the four pulmonary veins and atrial appendage. Thus, a total of 24 parameters can be derived by calculating three low-voltage zone (LVZ) regions as percentages of each of the eight aforementioned anatomical regions. An example of a background voltage map with ablation lines that can be used to calculate the low-voltage zone parameters is shown in Figure 6. As shown, the red and orange regions reflect low voltages less than 0.2 mV, the yellow and green regions reflect low voltages between 0.2 and 0.5 mV, the light blue region reflects voltages greater than 0.5 mV, the dark orange sphere reflects ablation on the left WACA line, the light orange sphere reflects ablation on the right WACA line, the purple sphere reflects ablation on the roof line, and the yellow sphere reflects ablation on the lower line.

[0070] Global voltage parameters The automated algorithm calculates four voltage parameters across eight subanatomical structures (posterior wall, inferior wall, anterior wall, superior wall, septal wall, left lateral wall, mitral wall, left atrial appendage, and the left atrium excluding the four pulmonary veins and atrial appendage). The four voltage parameters are the 25th, 50th, median, and 75th voltage percentiles, as well as the range between the 25th and 75th percentiles.

[0071] Cycle length parameter The automated algorithm calculates four global cycle length parameters across eight subanatomical structures (posterior wall, inferior wall, anterior wall, superior wall, septal wall, left lateral wall, mitral wall, left atrial appendage, and the left atrium excluding the four pulmonary veins and atrial appendage). The four cycle length parameters are the 25th, 50th, median, and 75th cycle length percentiles, as well as the range between the 25th and 75th percentiles.

[0072] In one embodiment, a pulmonary vein isolation non-responder network can be trained using the features described above. A deep learning algorithm can be trained using a set of 1,300 cases encompassing both persistent and paroxysmal conditions treated solely by pulmonary vein isolation. The dataset can be divided into an 80% training set and a 20% test set. Each case can be labeled as either a responder or a non-responder to pulmonary vein isolation. The high-density layer shown at the bottom of Figure 1 can be trained using a binary cross-entropy loss function.

[0073] Vector velocity field divergence and conduction velocity: The vector velocity field can be defined with respect to the local activation time. The detection of the local activation time and the normalization of the local activation surface can be modeled as follows: T(x,y)=ax 2 +by 2 +cxy+dx+hole+f In the equation, (x, y) represents the surface of the catheter in 2D space after projecting 3D space into 2D space, assuming the catheter is located on a 2D surface. This assumption is verified by checking that two high eigenvalues ​​conserve 95% of the energy.

[0074] Vector velocity field divergence and conduction velocity. In response to this, the vector velocity field schematically shown in Figure 7 can be estimated as follows:

[0075]

number

[0076] In this figure, the surface is fitted to the local activation time, the XY surface represents the electrode position in two-dimensional space, and the Z axis represents the time difference of local activation from the reference point to the normal QRS time. Such a vector velocity field can provide raw data for obtaining rotation, divergence, and conduction velocity maps. Conduction velocity is measured as the average conduction velocity in a region of radius 2 mm within the velocity vector field. Rotation and divergence are mathematical operators that can be applied to velocity vector fields, as described in C, Dallet et al., *Cardiac Propagation Pattern Mapping With Vector Field for Helping Tachyarrhythmias Diagnosis With Clinical Tridimensional Electro-Anatomical Mapping Tools*, IEEE Trans Biomed Eng. 2019 Feb. 66(2):373-382. Rotation provides a measure of curvature in wavefront propagation by calculating the circulation of vector-valued velocities. The divergence field, on the other hand, is a mathematical tool for evaluating the tendency of conduction velocity (CV) vectors around each point to converge or diverge from there. Specifically, a negative divergence value indicates that the conduction velocity is moving inward toward a point, creating a "sink" in the conduction velocity field where the wavefront collides. Conversely, a positive divergence value suggests that the conduction velocity is moving outward from the point that forms the "source" in the conduction field, and the wavefront spreads out from that point.

[0077] Activation type map After pacing, the start and end points of the bipolar window can be determined, as schematically shown in Figure 8. Bipolar windowing refers to a set of signal processing operations to detect the active portion of the bipolar signal, which is performed after a set of operations (far-field reduction and pacing rejection). Each bipolar window is then decomposed into Hermitian basis functions, and the resulting coefficients and width parameters are used to represent the bipolar window. This method is well applied to clustering QRS groups. Each bipolar window is modeled as a combination of Hermitian basis functions, centered on the location of its maximum peak:

[0078]

number

[0079]

number

[0080]

number

[0081] Therefore, Figure 9 shows an example of modeling different waves using Hermitian basis functions (N=8) to reconstruct the bipolar windows of different waves with different sets of coefficients.

[0082] Next, unsupervised clustering of the pulses can be performed based on the similarity between coefficients to present "families" of pulses. The physician annotates each family with the type of pulse (normal, dual potential, fractionated, or noisy), as shown in Figure 10. As can be seen, this reduces the need to review annotations for each pulse. A random forest decision tree can then be run to obtain a set of Hermitian coefficients and annotated labels to label each pulse as one of the possible pulse categories.

[0083] Figure 11 is a flowchart of a computer-implemented method for using deep learning to identify unresponsive individuals in pulmonary vein isolation under one embodiment. Flowchart 1100 shows method actions, indicated as flowchart blocks for specific actions involving and / or between system elements 102-144 in Figure 1.

[0084] In block 1102, a set of machine learning models is trained to predict which atrial fibrillation patients will respond to which ablation line treatment. The system trains a neural network to distinguish which ablation line treatments each patient may respond to. For example, but not limited to, this could include the training system training a set of machine learning models to predict which atrial fibrillation patients will respond to which ablation line, treatment, for example, whether one patient will respond to pulmonary vein isolation treatment alone and another patient will respond to pulmonary vein isolation treatment plus treatment.

[0085] Machine learning models can be applications of artificial intelligence that provide systems with the ability to automatically improve from experience without being explicitly programmed. A patient with atrial fibrillation may be a person who is receiving or registered to receive medical treatment for the most commonly diagnosed persistent arrhythmia, characterized by rapid and irregular activation of the upper cavity of the heart. Treatment can be medical care given to a patient for a disease or injury. An ablation line can be a long, narrow mark or band where a wound is surgically formed that can block the abnormal electrical pathway causing the arrhythmia. Ablation line treatment can be medical care that surgically creates a wound in a long, narrow mark or band that can block the abnormal electrical pathway causing the arrhythmia in a patient.

[0086] After training, the set of trained machine learning models identifies at least one ablation line treatment associated with a patient based on atrial fibrillation-related features associated with the patient (block 1104). The system uses medical imaging to identify the patient's ablation lines. As an example, but not limited to, this could include the set of trained machine learning models 100 extracting maps from medical imaging, such as specific intracardiac electrocardiogram data of a 61-year-old patient named Pat, and correlating the mapping data with treatments using 33 ablation sites across six ablation lines.

[0087] A trained machine learning model can be an artificial intelligence application developed to allow the system to automatically improve from experience without being explicitly programmed. The ablation site can be an area or location where damage is surgically formed that can block abnormal electrical pathways causing arrhythmias. The patient can be a person who is receiving or registered to receive medical treatment. Atrial fibrillation-related features can be specific attributes associated with an irregular, often rapid heart rhythm characterized by a heart that beats in a disorderly and irregular manner.

[0088] Once at least one ablation line treatment is identified, a set of trained machine learning models estimates the probability of reconnection of ablation sites associated with at least one ablation line treatment in block 1106, based on the ablation sites and corresponding ablation characteristics. The system determines which ablated sites are capable of spontaneously reconnecting. For example, but not limited to, this could include a set of trained machine learning models 100 extracting mapping data from data of 33 ablation sites and corresponding ablation characteristics, the mapping data including impedance drop, radio frequency index, catheter stability, catheter power, and / or catheter force, and then storing the mapping data in an accessible data structure.

[0089] The reconnection probability may be the estimated recurrence of atrial fibrillation after catheter ablation. The corresponding ablation characteristics may be related to features or qualities typically associated with the surgical generation of damage that can block the abnormal electrical pathways causing arrhythmias.

[0090] In addition to suggesting ablation lines and reconnection sites, block 1108 uses a set of trained machine learning models to identify patient-associated atrial fibrillation predictors based on patient-associated demographics and cardiac component dimensional parameters. The system identifies the patient's atrial fibrillation predictors. For example, but not limited to, this could include a set of trained machine learning models 100 receiving patient demographic data and right atrial parameters including atrial size, atrial sphericity, low voltage zone, global voltage, and / or cycle length, which the set of trained machine learning models uses to identify the patient's atrial fibrillation predictors.

[0091] A patient's atrial fibrillation predictor may be a predictor of the most commonly diagnosed persistent arrhythmia, characterized by rapid and irregular activation of the heart. Demographics may be the statistical characteristics of a human population. Cardiac dimension parameters may be numerical or other measurable factors that define the components of the heart.

[0092] After identifying at least one ablation line treatment, reconnection probability, and atrial fibrillation predictor, a set of trained machine learning models predict whether the patient will respond to treatment using pulmonary vein isolation only, based on at least one ablation line treatment, the estimated reconnection probability of the ablation site, and any patient atrial fibrillation predictor, block 1110. The system predicts whether the patient's atrial fibrillation is likely to respond to treatment using pulmonary vein isolation only. In an embodiment, this may include a set of trained machine learning models 100 receiving data including six ablation line treatments with 33 ablation sites, the highest estimated reconnection probability of the ablation site, and Pat's atrial fibrillation predictor, and predicting that Pat's atrial fibrillation condition will not respond to treatment using pulmonary vein isolation only. The system 100 can then output this prediction 144 to Pat's physician.

[0093] A prediction may be a prediction of what someone thinks will happen, or a statement about future events or future data. Pulmonary vein isolation alone may be a medical procedure aimed at restoring normal rhythm in the left atrium and reducing or eliminating symptoms associated with atrial fibrillation.

[0094] In response to a prediction that a patient will not respond to any treatment using pulmonary vein isolation alone, block 1112 enables the healthcare provider to offer the patient at least one ablation line treatment. This system enables pulmonary vein isolation plus treatment for patients who require more than pulmonary vein isolation alone. As an example, but not limited to, this could include, a set of trained machine learning models 100 outputting detailed coordinates of 33 ablation sites in six ablation lines within Pat's right atrium identified for Pat, thereby enabling Pat's physician to provide treatment that would otherwise not have been offered. The healthcare provider may be an individual or organization providing healthcare services. At least one ablation line treatment is a line of ablation sites that are organized as an ablation line when identified for the patient or when enabled for the healthcare provider. This means that the ablation line treatment may be formed as a line when output as an ablation line prediction 114, or may be formed as a line by the final high-density layer 130 after being processed as an ablation site by ARMA softmax.

[0095] In addition to predicting whether a patient will be a responder to treatment with pulmonary vein isolation using one of the trained machine learning models, the trained machine learning models can also predict which ablation approach is most likely to result in a better outcome for the patient. In other words, the trained machine learning models can be used to recommend ablation approaches that include pulmonary vein isolation treatment alone (typically with WACA ablation) or "pulmonary vein isolation plus" treatments that include ablation strategies beyond WACA ablation. These alternative ablation strategies may include, but are not limited to, 1) anatomical ablation to isolate or traverse anatomical segments / structures / locations within the atrium; 2) linear ablation according to EA mapping findings, regardless of map type, which may consist of any combination of ablation lines, including, but not limited to, roof line, inferior line, mitral valve line (different types), anterior line, LAA separation, SVC separation, etc.; 3) local ablation, including, but not limited to, fractional ablation, CARTOFINDER ablation, fast CL ablation, low voltage ablation and / or homogenization, Isuprel mapping and ablation / breakthrough ablation, slow conduction corridor and pivot point ablation, GP ablation, spatiotemporal dispersion ablation, etc.; and 4) local anatomical ablation, which may include, but are not limited to, debulking of specific segments / structures through local ablation. Therefore, if the model predicts that a patient's atrial fibrillation will not respond to treatment using pulmonary vein isolation alone, the model can recommend to that patient an ablation approach that leads to a successful long-term outcome, namely pulmonary vein isolation alone or pulmonary vein isolation plus, which includes the type that is most likely to result in a better outcome for the patient.

[0096] Figure 11 shows blocks 1102-1112 occurring in a specific order, although blocks 1102-1112 may occur in a different order. In other implementations, each of blocks 1102-1112 may be executed in combination with other blocks, and / or some blocks may be split into different sets of blocks.

[0097] System Overview An exemplary hardware device on which the subject may be implemented is described. Those skilled in the art will understand that the elements shown in Figure 12 may vary depending on the system implementation. Referring to Figure 12, an exemplary system for implementing the subject disclosed herein includes a hardware device 1200 which includes a processing unit 1202, memory 1204, storage 1206, data input module 1208, display adapter 1210, communication interface 1212, and a bus 1214 that connects elements 1204-1212 to the processing unit 1202.

[0098] Bus 1214 may include any type of bus architecture, such as a memory bus, peripheral bus, or local bus. Processing unit 1202 is an instruction execution machine, apparatus, or device, which may include a microprocessor, digital signal processor, graphics processing unit, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), etc. Processing unit 1202 may be configured to execute program instructions stored in memory 1204 and / or storage 1206 and / or received via data input module 1208.

[0099] Memory 1204 may include read-only memory (ROM) 1216 and / or random-access memory (RAM) 1218. Memory 1204 may be configured to store program instructions and data during the operation of hardware device 1200. In various embodiments, memory 1204 may include any of various memory technologies such as static random-access memory (SRAM) or dynamic RAM (DRAM), including, for example, variations such as dual data-rate synchronous DRAM (DDR SDRAM), error-correction code synchronous DRAM (ECC SDRAM), or RAMBUS DRAM (RDRAM). Memory 1204 may also include non-volatile memory technologies such as non-volatile flash RAM (NVRAM) or ROM. In some embodiments, memory 1204 is intended to include combinations of the technologies described above, as well as other technologies not specifically mentioned. When the subject matter of the present invention is implemented in a computer system, a basic input / output system (BIOS) 1220 containing basic routines that help transfer information between elements in the computer system, such as during startup, is stored in ROM 1216.

[0100] Storage 1206 may include a flash memory data storage device for reading from and writing to flash memory, a hard disk drive for reading from and writing to a hard disk, a magnetic disk drive for reading from and writing to a removable magnetic disk, and / or an optical disk drive for reading from and writing to a removable optical disk such as a CD-ROM, DVD, or other optical medium. The drives and their associated computer-readable media provide non-volatile storage for computer-readable instructions, data structures, program modules, and other data for hardware device 1200.

[0101] It should be noted that the methods described herein may be embodied in executable instructions stored on a computer-readable medium for use by or in connection with an instruction-executing machine, apparatus, or device, such as a computer-based or processor-based machine, apparatus, or device. Those skilled in the art will understand that in some embodiments, other types of computer-readable medium capable of storing computer-accessible data, such as magnetic cassettes, flash memory cards, digital video discs, Bernoulli cartridges, RAM, and ROM, may be used and may be used in exemplary operating environments. As used herein, “computer-readable medium” may include one or more of any suitable mediums for storing executable instructions of a computer program in one or more electronic, magnetic, optical, and electromagnetic formats, so that an instruction-executing machine, system, apparatus, or device can read (or fetch) instructions from the computer-readable medium and execute instructions for performing the methods described herein. A non-exclusive list of conventional exemplary computer-readable media includes portable computer diskettes, RAM, ROM, erasable programmable read-only memory (EPROM or flash memory), and optical storage devices, including portable compact discs (CDs), portable digital video discs (DVDs), high-definition DVDs (HD-DVD®), Blu-ray discs, and the like.

[0102] Several program modules, including an operating system 1222, one or more application programs 1224, program data 1226, and other program modules 1228, may be stored in storage 1206, ROM 1216, or RAM 1218. A user can input commands and information to the hardware device 1200 via a data input module 1208. The data input module 1208 may include mechanisms such as a keyboard, touchscreen, or pointing device.

[0103] Other external input devices (not shown) are connected to the hardware device 1200 via the external data input interface 1230. Examples, but not limited to, of the above, external input devices may include microphones, joysticks, gamepads, satellite receiving antennas, scanners, and the like. In some embodiments, external input devices may include video or audio input devices such as video cameras and still cameras. The data input module 1208 may be configured to receive input from one or more users of the hardware device 1200 and distribute such input to the processing unit 1202 and / or memory 1204 via the bus 1214.

[0104] Display 1232 is also connected to bus 1214 via display adapter 1210. Display 1232 may be configured to display the output of hardware device 1200 to one or more users. In some embodiments, a given device, such as a touchscreen, may function as both the data input module 1208 and the display 1232. An external display device may also be connected to bus 1214 via external display interface 1234. Other peripheral output devices, not shown, such as speakers and printers, may be connected to hardware device 1200.

[0105] The hardware device 1200 can operate in a networked environment using logical connections to one or more remote nodes (not shown) via the communication interface 1212. The remote nodes may be other computers, servers, routers, peer devices, or other common network nodes, and typically include many or all of the elements described above with respect to the hardware device 1200. The communication interface 1212 may interface with wireless and / or wired networks. Examples of wireless networks include, for example, Bluetooth networks, wireless personal area networks, wireless 802.11 local area networks (LANs), and / or wireless telephone networks (e.g., cellular, PCS, or GSM networks). Examples of wired networks include, for example, LANs, fiber optic networks, wired personal area networks, telephone networks, and / or wide area networks (WANs). Such network environments are common in intranets, the internet, offices, and enterprise-scale computer networks. In some embodiments, the communication interface 1212 may include logic configured to support direct memory access (DMA) transfers between the memory 1204 and other devices.

[0106] In a networked environment, the program modules, or parts thereof, shown for hardware device 1200 may be stored on a remote storage device, such as on a server. It will be understood that other hardware and / or software may be used to establish a communication link between hardware device 1200 and other devices.

[0107] It should be understood that the arrangement of the hardware device 1200 shown in Figure 12 is only one possible implementation, and other arrangements are also possible. Furthermore, it should be understood that the various system components (and means) defined by the claims, described below, and shown in various block diagrams represent logical components configured to perform the functions described herein. For example, one or more of these system components (and means) may be realized, in whole or in part, by at least some of the components shown in the configuration of the hardware device 1200.

[0108] In addition, at least one of these components may be implemented at least partially as an electronic hardware component and thus constitute a machine, while the other components may be implemented as software, hardware, or a combination of software and hardware. More specifically, at least one component as defined by the claims may be implemented at least partially as an electronic hardware component such as an instruction execution machine (e.g., a processor-based machine or a machine containing a processor) and / or as a dedicated circuit or circuit configuration (e.g., individual logic gates interconnected to perform a dedicated function) such as those shown in Figure 12.

[0109] Other components may be implemented as software, hardware, or a combination of software and hardware. Furthermore, some or all of these other components may be combined, some may be omitted entirely, and additional components may be added while still achieving the functions described herein. Thus, the subject matter described herein may be embodied in many different variations, and all such variations are considered to be within the scope of the claims.

[0110] In accordance with the foregoing, in at least one case referred to herein as Example A-1, a system for using deep learning to identify ablation line treatments for pulmonary vein isolation only comprises one or more processors and a non-temporary computer-readable medium for storing a plurality of instructions, which, when executed, cause one or more processors to train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatments using ablation lines, and by the set of trained machine learning models to identify at least one ablation line treatment for a patient based on atrial fibrillation-related features associated with the patient, and by the set of trained machine learning models to identify at least one ablation line treatment associated with at least one ablation line treatment The probability of reconnection at the ablation site is estimated based on the ablation site and the corresponding ablation characteristics; a set of trained machine learning models identifies patient-associated atrial fibrillation predictors based on patient-associated demographics and cardiac component dimension parameters; a set of trained machine learning models predicts whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on at least one ablation line treatment, reconnection probability, patient atrial fibrillation predictor, and patient demographics; and in response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone, healthcare providers are enabled to provide the patient with at least one ablation line treatment.

[0111] In another case referred to herein as Example A-2, the system of Example A-1, or any other exemplary embodiment described herein, may be further limited to training a set of machine learning models, which includes using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome.

[0112] In another case referred to herein as Example A-3, the system of Example A-1, or any other exemplary embodiment described herein, may be further limited to training a set of machine learning models in response to the number of records not exceeding a training record threshold for observed outcomes, which includes using a training dataset containing a certain number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, where the number of records exceeds a training record threshold for a healthcare provider recommended ablation approaches, and identifying at least one ablation line treatment for a patient based on a healthcare provider recommended ablation approaches.

[0113] In another case referred to herein as Example A-4, the system of Example A-1, or any other exemplary embodiment described herein, may be further limited, and predicting whether a patient will not respond to any treatment using pulmonary vein isolation alone is based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone.

[0114] In another case referred to herein as Example A-5, the system of Example A-1, or any other exemplary embodiment described herein, may be further limited to including identifying at least one ablation line for a patient, and predicting the probability of treatment success regardless of the ablation approach.

[0115] In another case referred to herein as Example A-6, the system of Example A-1, or any other exemplary embodiment described herein, may be further limited to the fact that at least one ablation line treatment is directed to a location other than the left atrium, and therefore is pulmonary vein isolation plus treatment, rather than pulmonary vein isolation treatment alone.

[0116] In another case referred to herein as Example A-7, the system of Example A-1, or any other exemplary embodiment described herein, may be further limited to the fact that at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider.

[0117] Consistent with the foregoing, in at least one case referred to herein as Example B-1, a method for using deep learning to identify non-responders of pulmonary vein isolation alone includes: training a set of machine learning models to predict which atrial fibrillation patients will respond to which treatment using ablation lines; using the set of trained machine learning models to identify at least one ablation line treatment for a patient based on atrial fibrillation-related features associated with the patient; and using the set of trained machine learning models to determine the probability of reconnection of the ablation site associated with at least one ablation line treatment, the ablation site and the corresponding ablation The method includes: estimating based on characteristics; identifying patient-associated patient atrial fibrillation predictors based on patient-associated demographics and cardiac component dimension parameters using a set of trained machine learning models; predicting whether a patient will not respond to any treatment using pulmonary vein isolation alone, based on at least one ablation line treatment, reconnection probability, patient atrial fibrillation predictors, and patient demographics using a set of trained machine learning models; and enabling healthcare providers to offer a patient at least one ablation line treatment in response to a prediction that the patient will not respond to any treatment using pulmonary vein isolation alone.

[0118] In another case referred to herein as Example B-2, the methods of Example B-1, or any other exemplary embodiment described herein, may be further limited to training a set of machine learning models, which includes using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome.

[0119] In another case referred to herein as Example B-3, the methods of Example B-1, or any other exemplary embodiment described herein, may be further limited to training a set of machine learning models in response to the number of records not exceeding a training record threshold for observed outcomes, using a training dataset containing a certain number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, the number of records exceeding a training record threshold for a healthcare provider recommended ablation approaches, and identifying at least one ablation line treatment for patients based on a healthcare provider recommended ablation approaches.

[0120] In another case referred to herein as Example B-4, the methods of Example B-1, or any other exemplary embodiment described herein, may be further limited to predicting whether a patient will not respond to any treatment using pulmonary vein isolation alone, and training a set of machine learning models using a training dataset consisting only of records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone.

[0121] In another case referred to herein as Example B-5, the methods of Example B-1, or any other exemplary embodiment described herein, may be further limited to including identifying at least one ablation line for a patient, and predicting the probability of treatment success regardless of the ablation approach.

[0122] In another case referred to herein as Example B-6, the method of Example B-1, or any other exemplary embodiment described herein, may be further limited to the fact that at least one ablation line treatment is directed to a location other than the left atrium, and therefore is pulmonary vein isolation plus treatment, rather than pulmonary vein isolation alone.

[0123] In another case referred to herein as Example B-7, the methods of Example B-1, or any other exemplary embodiments described herein, may be further limited, and at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or enabled to the healthcare provider.

[0124] In accordance with the foregoing, in at least one case referred to herein as Example C-1, a computer program product comprising a non-temporary computer-readable medium having computer-readable program code embodied to be executed by one or more processors comprises program code including instructions, the instructions include training a set of machine learning models to predict which treatments using ablation lines will respond to which atrial fibrillation patients, the set of trained machine learning models to identify at least one ablation line treatment for a patient based on atrial fibrillation-related features associated with the patient, and the set of trained machine learning models to identify at least one ablation line treatment associated with the ablation line treatment The aim is to estimate the reconnection probability of the ablation site based on the ablation site and corresponding ablation characteristics, to identify patient-associated atrial fibrillation predictors based on patient-associated demographics and cardiac component dimension parameters using a set of trained machine learning models, to predict whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on at least one ablation line treatment, reconnection probability, patient atrial fibrillation predictors, and patient demographics, and to enable healthcare providers to offer the patient at least one ablation line treatment in response to a prediction that the patient will not respond to any treatment using pulmonary vein isolation alone.

[0125] In another case referred to herein as Example C-2, the computer program product of Example C-1, or any other exemplary embodiment described herein, may be further limited to training a set of machine learning models, which includes using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome.

[0126] In another case referred to herein as Example C-3, the computer program product of Example C-1, or any other exemplary embodiment described herein, may be further limited to training a set of machine learning models in response to the number of records not exceeding a training record threshold for observed outcomes, which includes using a training dataset containing a certain number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, where the number of records exceeds a training record threshold for a healthcare provider recommended ablation approaches, and identifying at least one ablation line treatment for a patient based on a healthcare provider recommended ablation approaches.

[0127] In another case referred to herein as Example C-4, the computer program product of Example C-1, or any other exemplary embodiment described herein, may be further limited to predicting whether a patient will not respond to any treatment using pulmonary vein isolation alone, based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone.

[0128] In another case referred to herein as Example C-5, the computer program product of Example C-1, or any other exemplary embodiment described herein, may be further limited to including identifying at least one ablation line for a patient, and predicting the probability of treatment success regardless of the ablation approach.

[0129] In another case referred to herein as Example C-6, the computer program product of Example C-1, or any other exemplary embodiment described herein, may be further limited, wherein at least one ablation line treatment is directed to a location other than the left atrium and is therefore a pulmonary vein isolation plus treatment, not a pulmonary vein isolation treatment alone, and at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider.

[0130] In the above description, unless otherwise indicated, the subject matter refers to the symbolic representation of actions and operations performed by one or more devices. Thus, such actions and operations, sometimes referred to as computer-executed, are understood to involve operations by units processing data in a structured form. These operations either transform data or maintain data in a location within the computer's memory system, thereby reconfiguring or otherwise modifying the operation of the device in a manner well understood by those skilled in the art. The data structure in which the data is maintained is a physical location in memory having specific characteristics defined by the data format. However, the subject matter, while described in context, is not intended to be limiting, as those skilled in the art will understand that various actions and operations described below can also be performed in hardware.

[0131] To facilitate understanding of the subject matter described above, many embodiments are described in relation to a series of actions. At least one of these embodiments, as defined by the claims, is performed by an electronic hardware component. For example, it will be recognized that various actions may be performed by a dedicated circuit or circuit configuration, by program instructions executed by one or more processors, or by a combination of both. Any description of a series of operations in this specification is not intended to mean that the specific order in which the series of actions must be performed must be followed. All methods described herein may be performed in any preferred order unless otherwise indicated herein or unless it is clearly inconsistent with the context.

[0132] While one or more implementations have been described as examples with respect to specific embodiments, it should be understood that the one or more implementations are not limited to the disclosed embodiments. Conversely, as will be apparent to those skilled in the art, it is intended to encompass a variety of modifications and similar configurations. Therefore, the appended claims should be given the broadest possible interpretation to encompass all such modifications and similar configurations.

[0133] [Implementation Method] (1) A system for using deep learning to identify unresponsive individuals in pulmonary vein isolation, wherein the system is One or more processors, A non-temporary computer-readable medium that stores multiple instructions, wherein, when the multiple instructions are executed, one or more processors, We train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatment using ablation lines. The set of trained machine learning models is used to identify at least one ablation line treatment for the patient based on atrial fibrillation-related features associated with the patient. The set of trained machine learning models is used to estimate the probability of reconnection of the ablation site associated with the at least one ablation line treatment, based on the ablation site and the corresponding ablation characteristics. The set of trained machine learning models identifies patient-associated atrial fibrillation predictors based on the demographic and cardiac component dimension parameters associated with the patient. The set of trained machine learning models is used to predict whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on the at least one ablation line treatment, the reconnection probability, the patient atrial fibrillation predictor, and the patient's demographics. A system that enables a healthcare provider to offer the patient the at least one line of ablation treatment in response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone. (2) Training the set of machine learning models comprises using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome, as described in Embodiment 1 of the system. (3) Training the set of machine learning models in response to the number of records not exceeding the training record threshold for the observed outcome, comprising using a training dataset containing a certain number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, wherein the number of records exceeds the training record threshold for a healthcare provider recommended ablation approaches, and identifying the at least one ablation line treatment for the patient is based on the healthcare provider recommended ablation approaches, as described in Embodiment 2 of the system. (4) The system according to Embodiment 1, wherein predicting whether the patient will not respond to any treatment using pulmonary vein isolation alone is based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone. (5) The system according to Embodiment 1, wherein identifying the at least one ablation line for the patient includes predicting the probability of success of the treatment regardless of the ablation approach.

[0134] (6) The system according to Embodiment 1, wherein the at least one ablation line treatment is directed to a location other than the left atrium, and therefore is pulmonary vein isolation plus treatment, rather than pulmonary vein isolation alone. (7) The system according to Embodiment 1, wherein the at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider. (8) A computer-aided method for using deep learning to identify unresponsive individuals in pulmonary vein isolation, the computer-aided method being: The goal is to train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatments using ablation lines, and The set of trained machine learning models identifies at least one ablation line treatment for the patient based on atrial fibrillation-related features associated with the patient. The set of trained machine learning models estimates the probability of reconnection of the ablation site associated with at least one ablation line treatment, based on the ablation site and the corresponding ablation characteristics. The set of trained machine learning models identifies patient-associated atrial fibrillation predictors based on the demographics and cardiac component dimension parameters associated with the patient, The set of trained machine learning models predicts whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on the at least one ablation line treatment, the reconnection probability, the patient atrial fibrillation predictor, and the patient's demographics. In response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone, enabling healthcare providers to offer the patient at least one line of ablation treatment, A computer implementation method, including (9) The computer implementation method according to Embodiment 8, wherein training the set of machine learning models comprises using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcomes. (10) Training the set of machine learning models in response to the number of records not exceeding the training record threshold for the observed outcome, comprising using a training dataset comprising a number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, wherein the number of records exceeds the training record threshold for a healthcare provider recommended ablation approach, and identifying the at least one ablation line treatment for the patient is based on the healthcare provider recommended ablation approach, as described in Embodiment 9.

[0135] (11) A computer implementation method according to Embodiment 8, wherein predicting whether the patient will not respond to any treatment using pulmonary vein isolation alone is based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone. (12) The computer-aided method according to Embodiment 8, comprising identifying the at least one ablation line for the patient, predicting the probability of success of the treatment regardless of the ablation approach. (13) The computer-assisted method according to Embodiment 8, wherein the at least one ablation line treatment is directed to a location other than the left atrium, and is therefore a pulmonary vein isolation plus treatment, rather than a pulmonary vein isolation treatment alone. (14) The computer-aided method according to Embodiment 8, wherein the at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider. (15) A computer program product comprising a non-temporary computer-readable medium having computer-readable program code embodied to be executed by one or more processors, wherein the program code is Train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatment using ablation lines. The set of trained machine learning models identifies at least one ablation line treatment for the patient based on atrial fibrillation-related features associated with the patient. The set of trained machine learning models estimates the probability of reconnection of the ablation site associated with at least one ablation line treatment, based on the ablation site and the corresponding ablation characteristics. The set of trained machine learning models identifies patient atrial fibrillation predictors associated with the patient based on the demographic and cardiac component dimension parameters associated with the patient. The set of trained machine learning models predicts whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on the at least one ablation line treatment, the reconnection probability, the patient atrial fibrillation predictor, and the patient's demographics. A computer program product including instructions that enable a healthcare provider to provide the patient with the at least one ablation line treatment in response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone.

[0136] (16) Training the set of machine learning models comprises using a training dataset which includes a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome, as described in Embodiment 15. (17) Training the set of machine learning models in response to the number of records not exceeding the training record threshold for the observed outcome, comprising using a training dataset containing a certain number of records of atrial fibrillation patients recording a provider ablation approaches for the treatment of atrial fibrillation, wherein the number of records exceeds the training record threshold for a provider-recommended ablation approach, and identifying the at least one ablation line treatment for the patient is based on the provider-recommended ablation approach, as described in Embodiment 16 of the computer program product. (18) A computer program product according to Embodiment 15, which predicts whether the patient will not respond to any treatment using pulmonary vein isolation alone, by training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone. (19) The computer program product according to Embodiment 15, which includes identifying the at least one ablation line for the patient and predicting the probability of success of the treatment regardless of the ablation approach. (20) The computer program product according to Embodiment 15, wherein the at least one ablation line treatment is directed to a location other than the left atrium and is therefore not a pulmonary vein isolation treatment alone but a pulmonary vein isolation plus treatment, and the at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider.

Claims

1. A system for using deep learning to identify unresponsive individuals in pulmonary vein isolation, wherein the system is: One or more processors, A non-temporary computer-readable medium that stores multiple instructions, wherein, when the multiple instructions are executed, one or more processors, We train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatment using ablation lines. The set of trained machine learning models is used to identify at least one ablation line treatment for the patient based on atrial fibrillation-related features associated with the patient. The set of trained machine learning models is used to estimate the probability of reconnection of the ablation site associated with the at least one ablation line treatment, based on the ablation site and the corresponding ablation characteristics. The set of trained machine learning models identifies patient-associated atrial fibrillation predictors based on the demographic and cardiac component dimension parameters associated with the patient. The set of trained machine learning models is used to predict whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on the at least one ablation line treatment, the reconnection probability, the patient atrial fibrillation predictor, and the patient's demographics. A system that enables a healthcare provider to offer the patient the at least one ablation line treatment in response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone.

2. The system according to claim 1, wherein training the set of machine learning models comprises using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome.

3. Training the set of machine learning models in response to the number of records not exceeding the training record threshold for the observed outcome comprises using a training dataset comprising a number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, wherein the number of records exceeds the training record threshold for a healthcare provider recommended ablation approaches, and identifying the at least one ablation line treatment for the patient is based on the healthcare provider recommended ablation approaches, according to claim 2.

4. The system according to claim 1, wherein predicting whether the patient will not respond to any treatment using pulmonary vein isolation alone is based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone.

5. The system according to claim 1, wherein identifying the at least one ablation line for the patient includes predicting the probability of success of the treatment, regardless of the ablation approach.

6. The system according to claim 1, wherein the at least one ablation line treatment is directed to a location other than the left atrium, and therefore is not pulmonary vein isolation treatment alone, but pulmonary vein isolation plus treatment.

7. The system according to claim 1, wherein the at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider.

8. A computer-aided method for using deep learning to identify unresponsive individuals in pulmonary vein isolation, wherein the computer-aided method is: The goal is to train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatments using ablation lines, and The set of trained machine learning models identifies at least one ablation line treatment for the patient based on atrial fibrillation-related features associated with the patient. The set of trained machine learning models estimates the probability of reconnection of the ablation site associated with at least one ablation line treatment, based on the ablation site and the corresponding ablation characteristics. The set of trained machine learning models identifies patient-associated atrial fibrillation predictors based on the demographics and cardiac component dimension parameters associated with the patient, The set of trained machine learning models predicts whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on the at least one ablation line treatment, the reconnection probability, the patient atrial fibrillation predictor, and the patient's demographics. In response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone, enabling healthcare providers to offer the patient at least one line of ablation treatment, A computer implementation method, including

9. The computer implementation method according to claim 8, wherein training the set of machine learning models comprises using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcome.

10. Training the set of machine learning models in response to the number of records not exceeding the training record threshold for the observed outcome comprises using a training dataset comprising a number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, wherein the number of records exceeds the training record threshold for a healthcare provider recommended ablation approaches, and identifying the at least one ablation line treatment for the patient is based on the healthcare provider recommended ablation approaches, according to claim 9.

11. The computer implementation method according to claim 8, wherein predicting whether the patient will not respond to any treatment using pulmonary vein isolation alone is based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone.

12. The computer-aided method according to claim 8, comprising identifying the at least one ablation line for the patient, predicting the probability of success of the treatment regardless of the ablation approach.

13. The computer-assisted method according to claim 8, wherein the at least one ablation line treatment is directed to a location other than the left atrium, and therefore constitutes pulmonary vein isolation plus treatment, rather than pulmonary vein isolation treatment alone.

14. The computer-aided method according to claim 8, wherein the at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider.

15. A computer program product comprising a non-temporary computer-readable medium having computer-readable program code embodied to be executed by one or more processors, wherein the program code is Train a set of machine learning models to predict which atrial fibrillation patients will respond to which treatment using ablation lines. The set of trained machine learning models identifies at least one ablation line treatment for the patient based on atrial fibrillation-related features associated with the patient. The set of trained machine learning models estimates the probability of reconnection of the ablation site associated with at least one ablation line treatment, based on the ablation site and the corresponding ablation characteristics. The set of trained machine learning models identifies patient atrial fibrillation predictors associated with the patient based on the demographic and cardiac component dimension parameters associated with the patient. The set of trained machine learning models predicts whether the patient will not respond to any treatment using pulmonary vein isolation alone, based on the at least one ablation line treatment, the reconnection probability, the patient atrial fibrillation predictor, and the patient's demographics. A computer program product including instructions that enable a healthcare provider to provide the patient with the at least one ablation line treatment in response to the prediction that the patient will not respond to any treatment using pulmonary vein isolation alone.

16. Training the set of machine learning models comprises using a training dataset containing a certain number of records of atrial fibrillation patients recording observed outcomes from atrial fibrillation treatment, wherein the number of records exceeds a training record threshold for the observed outcomes, as described in claim 15.

17. Training the set of machine learning models in response to the number of records not exceeding the training record threshold for the observed outcome comprises using a training dataset comprising a number of records of atrial fibrillation patients recording a healthcare provider ablation approaches for the treatment of atrial fibrillation, wherein the number of records exceeds the training record threshold for a healthcare provider recommended ablation approaches, and identifying the at least one ablation line treatment for the patient is based on the healthcare provider recommended ablation approaches, according to the computer program product of claim 16.

18. The computer program product according to claim 15, wherein predicting whether the patient will not respond to any treatment using pulmonary vein isolation alone is based on training a set of machine learning models using a training dataset that includes only records of atrial fibrillation patients who have recorded treatment using pulmonary vein isolation alone.

19. The computer program product according to claim 15, wherein identifying the at least one ablation line for the patient includes predicting the probability of success of the treatment regardless of the ablation approach.

20. The computer program product according to claim 15, wherein the at least one ablation line treatment is directed to a location other than the left atrium and is therefore not a treatment of pulmonary vein isolation alone but a treatment of pulmonary vein isolation plus, and the at least one ablation line treatment comprises a line of ablation sites that is organized as an ablation line when identified to the patient or when enabled to the healthcare provider.

Images