Computer systems and methods for enhancing atrial fibrillation treatment

Abstract

To provide computer systems and computer-implemented methods for the treatment of atrial fibrillation.SOLUTION: The presently disclosed subject matter includes computer methods and computer systems that make it possible to select and provide a suitable treatment for AF patients that increases the likelihood of long-term amelioration of AF conditions thereby enhancing the treatment of AF patients. The disclosure provides methods and systems for the analysis of the condition of AF patients, the tailoring of personalized treatment regimens based on various AF features obtained from the patients, and the determination of a treatment selected from at least PVI only and PVI plus.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] (CROSS REFERENCE TO RELATED APPLICATIONS) This application claims the benefit of U.S. Provisional Patent Application No. 63 / 524,712, filed July 3, 2023, and U.S. Provisional Patent Application No. 63 / 611,303, filed December 18, 2023, which are incorporated herein by reference.

[0002] FIELD OF THEINVENTION The subject matter of this disclosure relates to computer systems and computer-implemented methods for treating atrial fibrillation. [Background technology]

[0003] Atrial fibrillation (AF) is the most commonly diagnosed sustained arrhythmia and is characterized by rapid and irregular activation of the atria. Atrial fibrillation may be paroxysmal, lasting 7 days or less with or without intervention, or may continue for more than 7 days (persistent, PS-AF) or more than 12 months (long-standing, LS-PS-AF). Persistent AF is the term used for long-standing persistent AF when attempts to restore sinus rhythm have been abandoned or have proven impossible. See, e.g., 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).

[0004] Treatment strategies vary depending on the severity and duration of AF. Pulmonary vein isolation (PVI) is a medical procedure that aims to restore normal cardiac rhythm and reduce or eliminate symptoms associated with AF. However, in severe cases of AF, PVI alone may not be sufficient, and long-term failure of "PVI alone" has been observed.

[0005] Evidence suggests a consistent association between enlargement of the left atrial volume and a worsening of both the severity and chronic nature of atrial fibrillation (AF) in affected patients. The geometry of the atria is known to change as well, going from a tubular to a spherical shape as the disease progresses.

[0006] 1a and 1b show the imaging system output of the left atrium of two AF patients. In this example, the expansion value (in this case, the Euclidean expansion value) is calculated by dividing the width of the posterior wall by its length.

[0007] The methods and systems disclosed herein enable the selection and delivery of an appropriate therapy for an AF patient, which increases the likelihood of long-term improvement of the AF condition, thereby enhancing the treatment of the AF patient.

[0008] It is disclosed herein that changes (e.g., expansion) in different subcompartments (or "chambers") of the heart, observed and measured at specific anatomical locations of the different subcompartments, can be used to predict long-term success or failure of PVI-only and / or PVI-plus treatment in AF patients.

[0009] Considering the left atrium as an example, the geometric transformations, such as expansion, in each of its anatomical regions, including posterior, anterior, left lateral, right lateral, roof, and inferior wall, have been found by the inventors to have predictive quality with respect to the outcome of PVI-only treatment. According to the inventors' findings, the specific geometric characteristics (or specific combinations of geometric characteristics) of the left atrium observed in AF patients indicate whether long-term success or long-term failure is expected after PVI-only treatment, and thus help select the appropriate treatment to be administered to the patient, choosing between PVI-only and PVI-plus.

[0010] According to some examples, an index (and novel techniques utilizing the index) is proposed that indicates the likelihood of long-term success of PVI "only" treatment in atrial fibrillation subjects. The index can be used to select patients for PVI only or one of the different PVI plus. The index is generated based on cardiac-related features including geometric characteristics and possibly other cardiac-related features such as voltage-related features and cycle length. Additional details of statistical analyses performed on data obtained from AF patients that support the above principles are provided in the "Examples" section below.

[0011] One challenge in using the geometric characteristics of the cardiac subcompartments relates to the fact that the imaging output of cardiac imaging systems is often compromised or incomplete, including unmapped regions. This drawback creates difficulties and in some cases precludes the ability to determine changes in the overall volume and / or shape of the heart and / or atria, thus creating technical setbacks in the analysis of the cardiac condition and selection of appropriate therapy in AF patients.

[0012] The subject matter disclosed herein includes a new approach to predicting the success of PVI treatment, which is effective even in the case of incomplete imaging output. According to this approach, a set of anatomical prediction parameters (also referred to as "success prediction parameters") can be used, each parameter being calculated based on geometric data collected from a respective anatomical site of the heart. The multiplicity and diversity of parameters, and the fact that different anatomical prediction parameters are calculated based on different respective anatomical sites, provide data redundancy, and even if the imaging output of the heart is incomplete, the mapped region of the heart provides sufficient output that can be used to calculate parameters and provide the desired prediction based on these parameters.

[0013] The subject matter disclosed herein further includes a dedicated computer system and method for recommending / selecting an appropriate ablation procedure for an AF patient using a machine learning (ML) model. The machine learning classifier is trained using a training dataset that includes a plurality of features (hereinafter "AF-associated features") determined for different patients in an AF patient database, and correlates between various combinations of AF-associated features and respective probabilities of success of PVI-only and / or PVI-plus procedures.

[0014] According to one approach, the training data includes information indicating whether PVI-only or PVI-plus treatment was administered to each patient, and the observed long-term success of the AF treatment. The AF-related features are labeled according to the observed success or failure of the treatment, thus enabling the ML model to associate the AF-related features with the recommended AF treatment.

[0015] According to another example, the training data includes information indicative of the PVI treatment (ablation procedure) recommended by the physician treating the patient, rather than the actual observed outcome of the treatment. As explained further below, this approach is based on the wisdom of the crowd (WoC) principle and allows the ML model to make associations between AF-related features and the treatment options selected by the physician. This helps to overcome technical challenges resulting from the lack of sufficient data on the observed outcomes of AF treatments.

[0016] The present disclosure also relates to the field of personalized medical treatment. More specifically, the present disclosure provides methods and systems for analyzing the condition of an AF patient and tailoring a personalized therapy selected from at least PVI only and PVI plus based on various AF characteristics obtained from the patient. Summary of the Invention [Means for solving the problem]

[0017] According to a first aspect of the subject matter of the present disclosure, there is provided a computer-implemented method for selecting an atrial fibrillation (AF) therapy for a patient, the method comprising: obtaining a 3D anatomical map of at least one subcompartment of the patient's heart; Segmenting the anatomical (3D) map, thereby obtaining a plurality of anatomical regions of at least one subcompartment; determining at least one respective dimension for at least a portion of a plurality of anatomical sites; determining one or more anatomical prediction parameters, each anatomical prediction parameter being determined based on geometric data collected from an anatomical region of the patient's heart; and generating a prediction indicating the likelihood of long-term success of a pulmonary vein isolation (PVI) only procedure and / or the likelihood of long-term success of a PVI plus procedure applied to the patient's heart based on the one or more anatomical prediction parameters.

[0018] In addition to the features described above, the method according to this aspect of the presently disclosed subject matter may optionally include one or more of the following features i-xi in any technically possible combination or permutation: i. The method includes generating an AF therapy recommendation for the patient based on the prediction, the recommendation being selected from a pulmonary vein isolation (PVI)-only procedure and / or a PVI-plus procedure, and generating data to indicate the recommendation to a healthcare provider. ii. Generating a prediction obtaining AF-related features of the patient, the cardiac-related features characterizing various cardiac-related attributes of the patient, the cardiac-related features being of at least the following types: a. one or more anatomical prediction parameters; b. one or more voltage-related characteristics; applying a machine learning (ML) model to the patient's AF-related features, the ML model being trained using a training dataset including a plurality of data points, each data point corresponding to an AF patient associated with an AF-related feature and an observed long-term success of an AF treatment administered to the patient, the AF treatment including at least only PVI, thereby training the ML model to provide data indicative of a recommended AF treatment for the given patient based on the AF-related features of the given patient during inference; and generating a machine learning model output indicating a recommended AF treatment selected from a group including at least PVI only and PVI plus treatment. iii. The machine learning model output includes information indicating the importance of each of the different input features in determining the recommended AF treatment. The iv.ML model is trained to provide data indicative of the probability of long-term success of a pulmonary vein isolation (PVI) only procedure. The v.ML model is trained to provide data indicative of the probability of long-term success of a pulmonary vein isolation (PVI) plus procedure. The vi.ML model is trained to provide data indicating the probability of long-term success of pulmonary vein isolation (PVI)-only and PVI-plus procedures. vii. The ML output including a recommendation to administer a PVI plus treatment to a patient also includes information indicating the specific type of PVI plus treatment that was recommended. viii. AF-related characteristics further include one or more of cycle length, patient medical history, and patient demographics. ix. The method includes comparing each of the one or more anatomical prediction parameters to a respective threshold value and determining a prediction based on an output of the comparison. x. The method includes determining one or more anatomical prediction parameters by applying a mathematical operation between different dimensions of the anatomical region. xi. The method includes performing one or more procedures on the patient to obtain cardiac-related characteristics from the patient.

[0019] In accordance with another aspect of the subject matter of the present disclosure, a computer system includes at least one processing circuit configured to select an atrial fibrillation (AF) therapy for a patient, the at least one processing circuit comprising: obtaining a 3D anatomical map of at least one subcompartment of the patient's heart; Segmenting the anatomical (3D) map, thereby obtaining a plurality of anatomical regions of at least one subcompartment; determining at least one respective dimension for at least a portion of a plurality of anatomical sites; determining one or more anatomical prediction parameters, each anatomical prediction parameter being determined based on geometric data collected from an anatomical region of the patient's heart; and generating a prediction indicative of the likelihood of long-term success of a pulmonary vein isolation (PVI)-only procedure and / or a PVI-plus procedure applied to the patient's heart based on the one or more anatomical prediction parameters.

[0020] The subject matter of the present disclosure further contemplates a non-transitory computer readable program storage device tangibly embodying a program of instructions executable by a computer to perform the method disclosed according to the first aspect above.

[0021] The subject matter of the present disclosure further contemplates a computer-implemented method and computer system (and respective non-transitory computer storage devices) for providing long-term improvement to an atrial fibrillation (AF) patient, the method comprising performing the operations described by the first aspect above and administering PVI-only therapy or PVI-plus to the patient as prescribed by the recommendation.

[0022] The subject matter of the present disclosure further contemplates a computer-implemented method and computer system (and respective non-transitory computer storage devices) for predicting a response to pulmonary vein isolation (PVI) therapy administered to an atrial fibrillation (AF) patient, the method including operations according to the first aspect above.

[0023] The subject matter of the present disclosure further includes a dedicated computer-implemented method and system (and respective non-transitory computer storage devices) for training a machine learning (ML) model to select / recommend an AF (atrial fibrillation) therapy by providing data indicative of a recommended therapy selected between a pulmonary vein isolation (PVI) therapy and a PVI plus therapy, the method comprising: The method includes: generating a training dataset including a plurality of data points, each data point corresponding to an AF patient associated with an AF-related characteristic and an observed long-term success of an AF treatment administered to the patient, the AF treatment including at least only PVI, and optionally including PVI plus; and training an ML model using the generated dataset to associate a particular AF-related characteristic with a corresponding probability of long-term success of the AF treatment, thereby enabling the ML model to provide, during inference, data indicative of a recommended AF treatment for a given patient based on the AF-related characteristics of the given patient.

[0024] According to another aspect of the subject matter of the present disclosure, there is provided a computer program product stored on a (non-transitory) computer storage device, the computer program product including software executable by a computer, the software including a machine learning (ML) model for recommending an AF (atrial fibrillation) therapy to an AF patient, the software comprising: The AF treatment is generated by a process that includes: generating a training dataset including a plurality of data points, each data point corresponding to an AF patient associated with an AF-related characteristic and an observed long-term success of an AF treatment administered to the patient, the AF treatment including at least only PVI, and possibly including PVI plus; and training an ML model using the generated training dataset to associate the particular AF-related characteristic with a corresponding probability of long-term success of the AF treatment, thereby enabling the ML model to provide, during inference, data indicative of a recommended AF treatment for a given patient based on the given patient's AF-related characteristics.

[0025] According to a further aspect of the presently disclosed subject matter, there is provided a computer system and a computer implemented method (and respective non-transitory computer storage devices) for determining a personalized therapy for an AF patient, the method comprising the steps / acts defined above according to the first aspect above.

[0026] The above disclosed systems, methods, non-transitory program storage devices, and computer program products according to various aspects of the presently disclosed subject matter may optionally include one or more of the above-listed features (i)-(xi), mutatis mutandis, in any technically relevant combination or permutation. [Brief description of the drawings]

[0027] In order to better understand the subject matter disclosed herein, and to illustrate how it may be carried out in practice, embodiments will now be described, by way of non-limiting example only, with reference to the accompanying drawings, in which: [Figure 1a] FIG. 1 is a diagram of the distended left atrium of an AF patient. [Figure 1b] FIG. 1 is a diagram of the distended left atrium of an AF patient. [Diagram 2] FIG. 1 is a block diagram of a computer system for enhanced treatment of AF patients, according to some examples of the presently disclosed subject matter. [Figure 3a] 1 is a flowchart illustrating operations performed as part of enhanced recommendations and treatments for AF patients, according to some examples of the presently disclosed subject matter. [Figure 3b] 1 is a flowchart illustrating operations performed as part of machine learning based enhanced recommendations and treatments for AF patients, according to some examples of the subject matter of the present disclosure. [Figure 4] 1 illustrates various stages of a computational analysis according to some examples of the subject matter of this disclosure. [Diagram 5] 1 is an image showing four anchor points on a posterior wall according to some examples of the subject matter of this disclosure. [Figure 6] 1 is a flowchart illustrating operations performed to train multiple machine learning models according to a WoC approach, in accordance with some examples of the subject matter of this disclosure. [Figure 7] 1 illustrates a schematic of inference using an ensemble of classifiers, in accordance with some examples of the subject matter of this disclosure. [Figure 8] FIG. 2 is a block diagram of a computer system that combines the functionality of system 200 with additional functionality related to identifying ablation lines, according to some examples of the subject matter of the present disclosure. [Figure 9] 1 is a graph showing a specific example of the correlation between right wall or septum (RLAT) width parameters and long-term success of PVI-only treatment, according to some examples of the presently disclosed subject matter. [Figure 10] 1 presents a graph showing survival probability of AF patients over time, according to some examples of the presently disclosed subject matter. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0028] In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the present invention. However, it will be understood by those skilled in the art that the subject matter of the present disclosure may be practiced without these specific details. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to obscure the subject matter of the present disclosure.

[0029] Unless specifically stated otherwise, and as will be apparent from the description below, it is understood that throughout this specification, descriptions utilizing terms such as "obtain," "apply," "generate," "provide," and the like refer to computer operations and / or processes that manipulate and / or transform data into other data, where such data is represented as physical quantities, such as electronic quantities, and / or where such data represents physical objects.

[0030] The terms "computer," "computer system," "computing device," "computerized device," and the like, should be expansively interpreted to include any type of hardware-based electronic device having one or more data processing circuits. Each processing circuit may comprise, for example, one or more processors operably connected to computer memory loaded with executable instructions for performing operations, as described further below.

[0031] The one or more processors referred to herein may represent one or more general-purpose processing devices, such as, for example, a microprocessor, a central processing unit, etc. More specifically, a given processor may be one of a complex instruction set computing (CISC) microprocessor, a reduced instruction set computing (RISC) microprocessor, a very long instruction word (VLIW) microprocessor, a processor implementing other instruction sets, or a processor implementing a combination of instruction sets. The one or more processors may also be one or more special-purpose processing devices, such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital signal processor (DSP), a graphics processing unit (GPU), a network processor, etc.

[0032] Memory as referred to in this specification may include one or more of the following: internal memory, such as processor registers and cache; main memory, such as read-only memory (ROM), flash memory, dynamic random-access memory (DRAM), such as synchronous DRAM (SDRAM) or Rambus DRAM (RDRAM).

[0033] Some of the operations according to the teachings herein may be performed by a computing device specially constructed for the desired purpose, or by a general-purpose computer specially configured for the desired purpose by a computer program stored on a computer-readable storage medium.

[0034] 2 and 7 show general schematic block diagrams of systems according to some examples of the subject matter of the present disclosure. The different components of FIG. 2 and FIG. 7 may be composed of any combination of software, hardware, and / or firmware performing the functions defined and described herein. The components of FIG. 2 and FIG. 7 may be centralized in one location and / or device, or distributed across two or more locations or devices. For example, FIG. 2 and FIG. 7 show computer system 220 and computer 210 as two separate computers, but in other examples, the two computer systems may be integrated into one computer system. In different examples of the subject matter of the present disclosure, the system may include fewer, more, and / or different modules than those shown in FIG. 2 and FIG. 7.

[0035] Figures 3a, 3b, and 6 are flow charts illustrating operations performed according to examples of the subject matter of the present disclosure. According to different examples of the subject matter of the present disclosure, fewer, more, and / or different steps may be performed than those shown in Figures 3a, 3b, and 6. According to different examples of the subject matter of the present disclosure, one or more of the steps shown in Figures 3a, 3b, and 6 may be performed in a different order and / or one or more groups of steps may be performed simultaneously for multiple commands.

[0036] It should be noted that any reference made in the description to the left atrium is made by way of non-limiting example only, and that the subject matter of the present disclosure also contemplates applying the same principles to other subcompartments of the heart as well, including, for example, the right atrium, the left ventricle, the right ventricle, and blood vessels connected to the heart, such as the pulmonary veins.

[0037] Reference is now made to FIG. 2, which illustrates a schematic block diagram of a computer system according to some examples of the subject matter of the present disclosure. The system 200 includes a computer 210 (also referred to as an "AF therapy recommendation and treatment computer system") configured to, among other things, analyze an anatomical map of an AF patient's heart and provide recommendations regarding a preferred method of PVI treatment. The computer system 210 illustrated in FIG. 2 may include at least one processing circuit 230 including one or more computer processors configured to execute several functional modules according to computer readable instructions embodied on a non-transitory computer readable storage medium. Such functional modules are hereinafter referred to as being included in the processing circuit 230. Notably, although the system 210 illustrates a single processing circuit, it may otherwise include two or more processing circuits, e.g., with different circuits including different functional modules.

[0038] The anatomical map (or model) of the heart may be generated by the computer 210 or may be received from another computer system. The anatomical map of the heart may be generated using anatomical images of the heart received from a suitable medical imaging (computer) system (220). Generation of the anatomical map may include using fast anatomical mapping (FAM) technology available, for example, in the CARTO™ system manufactured by Biosense Webster Inc. (Irvine, Calif.), or using any other suitable technology, or using any suitable combination of the above. The map of the heart may include any suitable type of three-dimensional (3D) anatomical map generated using any suitable technology. The map may be obtained from an intracardiac ECG (electrocardiogram), and may include both position information of the catheter moving around the atrium and ECG measurements such as voltage at each location.

[0039] In some examples, the processing circuitry 230 comprises an automatic map segmentation module 221 configured to automatically segment the 3D map of the heart into its various subcompartments and further segment one or more of the subcompartments into their respective anatomical sites, such as the posterior and anterior walls, the left side wall, the roof wall, the inferior wall, the septum, the mitral valve, the left atrial appendage (LAA), and the four pulmonary veins (LSPV, LIPV, RSPV, RIPV), etc. Notably, although all of the anatomical sites listed above may be provided during segmentation, not all of them are necessarily considered during the analysis and treatment recommendation process.

[0040] The processing circuitry 230 may further include a map dimensionality reduction module 223 configured to convert the 3D map into a 2D representation of the map, a cardiac subcompartment dimension calculator 225 configured to measure or calculate dimensions of various anatomical locations of various subcompartments of the heart, e.g., the left atrium, a geometric parameter calculator 227 configured to calculate at least one parameter (also referred to herein as an "anatomical prediction parameter"), and a treatment recommendation module 229 configured to determine and provide a recommendation as to whether a PVI-only treatment is likely to be successful if administered to a patient. The computer system 210 may further include a computer interface including, e.g., a keyboard and mouse and a computer display, and data may be displayed to, e.g., a health care provider, using a suitable graphical user interface program.

[0041] Reference is now made to Figure 3a, which is a flow chart illustrating operations performed as part of an analysis and treatment recommendation process according to some examples of the presently disclosed subject matter. For ease of understanding, and merely as a non-limiting example, the description of the operations disclosed in Figure 3a will be made with reference to components shown in system 200.

[0042] The process described below can be applied to analyze the heart of a patient suffering from AF and determine whether a PVI-only procedure may be the right choice of treatment for the patient, e.g., whether it may provide long-term success if administered to the patient. According to one example, the term "long-term success" is meant to include a period of AF and any arrhythmia free period after treatment that lasts from 3 to 18 months ("AF-free and arrhythmia-free period"). According to one specific example, long-term success refers to an AF-free and arrhythmia-free period of at least 3 months after treatment. According to another specific example, long-term success refers to an AF-free and arrhythmia-free period of at least 6 months after treatment. According to another specific example, long-term success refers to an AF-free and arrhythmia-free period of at least 9 months after treatment. According to another specific example, long-term success refers to an AF-free and arrhythmia-free period of at least 12 months after treatment. According to another specific example, long-term success refers to an AF-free and arrhythmia-free period of at least 15 months after treatment. According to another specific example, long-term success refers to an AF-free and arrhythmia-free period of at least 18 months after treatment.

[0043] The terms "improving," "augmenting," or "enhancing" an AF therapy refer to the determination of an AF therapy that results in longer AF-free and arrhythmia-free periods as compared to alternative AF therapies.

[0044] The term PVI only (or "PVI alone") refers to an ablation procedure applied only to the pulmonary veins of the left atrium, while PVI plus refers to additional ablation applied to additional anatomical sites of the heart. Additional areas targeted in PVI plus include, for example: These may include one or more of: RoofLine; LeftCarina; RightCarina; PosteriorLine; InferiorLine; MitralLine; AnteriorLine; LAAIsolation; CavoTricuspidIsthmus; RAAIsolation; InferiorVenaCavaIsolation; SuperiorVenaCavaIsolation; Substrate Modification; Substrate Isolation; Fractionation; Focal; Rotor; LA_PosteriorWall; CoronarySinus; RA_FreeWall; LA_AnteriorWall; LA_LateralWall; LA_InferiorWall; RA_Septum; RA_CristaTerminalis; Inferior_Mitral_Line; and LA_Septum.

[0045] In block 301, a 3D map of the patient's heart is obtained. As described above, the 3D map may be generated based on images of the heart received from a suitable imaging system (220). If the left atrium is being examined, the 3D map includes at least a portion of the heart including the left atrium.

[0046] At block 303, an automatic map segmentation process is applied to the 3D map to identify multiple anatomical locations of one or more subcompartments of the heart (e.g., the left atrium). In some examples, the 3D map is received by an automatic map segmentation module 221 configured to perform an automatic segmentation process to identify different portions of the heart within the 3D map. In some examples, the 3D image received by computer 210 is converted into a 3D map using suitable software, for example implemented as part of the automatic map segmentation module 221.

[0047] In some examples, the automated map segmentation process includes applying a machine learning module to the 3D map configured to automatically segment the heart or atrium (and possibly other sub-compartments of the heart) and provide data indicative of different anatomical regions (or structures) of the heart or atrium.

[0048] According to some examples, the automatic map segmentation process is the same process described in U.S. Patent Application No. 18 / 207854, entitled "SYSTEMS AND METHODS FOR RECOMMENDING ABLATION LINES," filed by the applicant on June 9, 2023, published on December 14, 2023, and having U.S. Patent Application Publication No. 2023-0397950. Application No. 18 / 207854 claims priority from U.S. Provisional Patent Application No. 63 / 350,983, filed on June 10, 2022.

[0049] US Patent Application No. 18 / 207854 discloses an automated map segmenter (22) configured to segment an anatomical map and output data identifying portions of the anatomy. The map segmenter utilizes a machine learning model (implementing a neural network and described in detail with reference to FIG. 3 of US Patent Application No. 18 / 207854) that receives as input a 3D map of the heart and provides as output a set of codes indicative of anatomical parts of the heart (e.g., anatomical parts of the atria). The machine learning model is trained using a training dataset that includes a sufficient number of paroxysmal and persistent AF cases (e.g., 5000 or more cases) with an anatomical map for each case. A more detailed description of the training and execution of the machine learning model dedicated to segmenting the heart is disclosed in US Patent Application No. 18 / 207854.

[0050] In some examples, the automatic map segmentation module 221 of FIG. 2 of the present application is configured similarly to the automatic map segmenter 22 disclosed in 18 / 207854. In other examples, certain, but not all, characteristics of the automatic map segmenter 22 may be retained. For example, the machine learning model described in 18 / 207854 is trained to provide both segmentation of portions of the heart or left atrium and ablation lines. In some examples, the automatic map segmentation module 221 disclosed herein may be trained to provide segmentation of anatomical portions of the heart and / or left atrium but not ablation lines.

[0051] In block 305, the 3D map (now segmented) undergoes dimensionality reduction to obtain a corresponding two-dimensional (2D) representation. In some examples, these operations are performed (e.g., by map dimensionality reduction module 223) by applying Principal Component Analysis (PCA) to project the 3D map onto a 2D plane, as is well known in the art.

[0052] In block 307, different dimensions of the 2D representation of various anatomical regions of one or more subcompartments of the heart are measured to obtain respective dimensional values, including, but not limited to, height and width of each region.

[0053] FIG. 4 shows an image of the left atrium on the left (a). In the middle (b), an anterior view of the atrium with the anterior design line (indicated by a dot) is shown, and its two-dimensional representation after PCA projection is shown on the right (c). In some examples, the anterior width (or height, if measured along the y-axis) is determined as the length in millimeters (mm) from the maximum point minus the minimum point along the x-axis (or y-axis). Mathematically, the width (or height) of an anatomical site (structure) is the length in mm from the maximum point minus the minimum point along the x-axis (or y-axis) of a point in a 2D plane. This calculation assumes that each of the anatomical sites of the heart has certain properties, which means that given two different dimensions of the atrium, one dimension will always be larger than the second dimension. For example, it may be expected that the anterior length of the atrium is larger than its width.

[0054] Additionally, according to some examples, the rear wall dimension is calculated based on four anchor points, as shown in FIG. - Anchor point 1 - the junction between the LSPV, the roof, and the back wall. - Anchor point 2 - the junction between the RSPV, the roof, and the rear wall. - Anchor point 3 - the junction between the RIPV, the inferior wall, and the posterior wall. - Anchor point 4 - the junction between the LIPV, inferior wall, and posterior wall.

[0055] Based on these anchor points, eight types of left atrial dimension measurements are calculated. a. Back Width 1 - Geodesic and Euclidean distance between anchor point 1 and anchor point 2. b. Back width 2 - Geodesic and Euclidean distance between anchor point 3 and anchor point 4. c. Back height 1 - Geodesic and Euclidean distance between anchor point 1 and anchor point 4. d. Back Height 2 - Geodesic and Euclidean distance between anchor point 2 and anchor point 3.

[0056] In block 309, one or more parameters (also referred to herein as "anatomical predictor parameters") indicative of the likelihood of long-term success (or long-term failure) of a PVI-only procedure administered to an AF patient are determined. Examples of these parameters are listed in Table 1, which is provided in the "Examples" section below.

[0057] In addition to the singular parameters ("singular anatomical prediction parameters" or "singular geometric parameters") determined based on measurements of a single dimension (e.g., RLAT width, RLAT length, POST width, POST height, etc.) of a single anatomical portion of the heart, additional parameters are determined based on two or more singular parameters (also referred to herein as "complex anatomical prediction parameters" or "complex geometric parameters"). This can be accomplished, for example, by applying a mathematical operation (e.g., multiplication) between pairs or triplets of singular parameters to obtain their respective values. In some examples, a product of all pairs of singular values ​​is calculated to obtain the additional parameter. In some examples, a product of all triples of singular values ​​is calculated (in addition to or instead of the product of the pairs) to obtain the additional parameter. A list of complex anatomical parameters generated by multiplying pairs of singular parameters is provided in the "Examples" section of Table 2 below. Notably, the subject matter of the present disclosure is not limited to pairs and triples of singular values, and in some examples, the complex anatomical prediction parameters can be calculated using more singular values.

[0058] The above calculations can be performed, for example, by a geometric parameter calculator 227 in the system 210, which can be configured to determine single geometric parameters and to calculate complex geometric parameters. In particular, if the anatomical prediction parameters are based on a single dimension, the dimension values ​​determined in the previous stage can be used in a simple manner. However, if the anatomical prediction parameters are based on two or more dimensions, corresponding calculations are performed.

[0059] In some examples, once one or more anatomical prediction parameters are calculated, the parameters are compared to respective predefined values, e.g., thresholds set to distinguish between long-term success and long-term failure (block 311). For example, considering the width of the right side wall of the left atrium (RLAT), according to the results described below, in order to achieve 75% sensitivity for detecting long-term failure, the threshold indicating long-term success may be equal to a value higher than 42.6 mm. Based on the results of the comparison, either long-term success or long-term failure of the PVI-only procedure can be determined, and a corresponding recommendation can be generated, where in case of predicted long-term success, PVI-only treatment is recommended, and otherwise, PVI-plus is recommended. In particular, although FIG. 3a illustrates the determination of a prediction regarding PVI-only, it is further contemplated to predict the long-term success of PVI-plus as well using appropriate AF-related features and thresholds. In some examples, a group of different anatomical prediction parameters is used to increase the robustness of the recommendation.

[0060] In some examples, the process continues by administering the appropriate treatment to the patient according to the recommendation (block 313). If long-term success of the PVI-only procedure is predicted, the patient is administered the PVI-only treatment, and if long-term failure of the PVI-only procedure is predicted, a different treatment (e.g., PVI plus) is administered to the patient. FIG. 8 below shows an example of a computer system that combines the functionality of system 200 with additional functionality including facilitating and guiding the appropriate PVI treatment. In some examples, once the recommended treatment is determined, the system can recommend suitable ablation lines and display the lines on the respective maps to guide the physician through the process.

[0061] The presently disclosed subject matter further contemplates a dedicated machine learning (ML) model for selecting an appropriate ablation technique (i.e., PVI or PVI plus) for AF patients. Figure 3b is a flow chart illustrating operations performed as part of a machine learning based analysis and treatment recommendation process according to some examples of the presently disclosed subject matter. The model is trained to make an association between AF-related features of AF patients and recommended treatments.

[0062] During the training phase, a ML model is trained using a training dataset that includes a number of features (hereinafter "AF-related features") determined for many different AF patients. According to one example, the model is trained to correlate between various combinations of AF-related features and the respective probabilities of successful PVI-only treatment. In another example, the model is trained to correlate between various combinations of AF-related features and the respective probabilities of successful PVI-only treatment and / or PVI-plus treatment.

[0063] A training data set of AF-related features of many different patients is obtained (block 301b). The AF-related features include at least one type of cardiac-related feature that is an attribute that characterizes the patient's heart, related to a cardiac parameter. The AF-related features may also include other AF-related features that are not cardiac-related.

[0064] One example of cardiac-related features is anatomical prediction parameters, including both singular and complex geometric parameters (e.g., pairs and triples). Additional examples of cardiac-related features that can be used during training and running of the ML model include voltage-related features and cycle length.

[0065] The voltage-related AF features are based on voltage measurements obtained from the patient. The voltage measurements include, for example, measurements made in low-voltage zone (LVZ) regions, low-voltage divisions, etc. The voltage measurements can be applied in a variety of ways. For example, the LVZ can be measured at each anatomical site (e.g., posterior wall, anterior wall, left lateral, right lateral, roof, and inferior wall). The voltage measurements can include, for example, one or more of absolute low-voltage values ​​measured over the entire area of ​​each anatomical site, and relative low-voltage values ​​representing the ratio between the measured low-voltage and the voltage over the entire area of ​​the anatomical site (e.g., posterior wall).

[0066] Additional voltage-related features include global voltage-related parameters, such as different percentiles of the voltage distribution (voltage pattern) in a portion of the heart or in its entirety, which may be based on different thresholds, e.g., the 25th percentile, the 50th percentile, the 90th percentile, etc.

[0067] Voltage-related features can be measured, for example, using the CARTO system by Biosense Webseter described above. FAM technology is designed to rapidly create high-resolution 3D anatomical maps of a cardiac chamber of interest, such as the left atrium. This is accomplished by moving a mapping catheter along the endocardial surface while continuously recording its location and corresponding electrical activity. FAM can be combined with voltage mapping for identification of the LVZ within the rapidly acquired anatomical map. Notably, in some instances, global voltage can be used alone as the only feature to determine PVI only or PVI plus.

[0068] As discussed above, a particular voltage measurement may be applied to multiple individual anatomical sites of the heart and / or atrium, thus allowing multiple different voltage-related characteristics to be obtained from a single patient.

[0069] Cycle length refers to any type of measurement of the time interval between detected electrical activations, for example activations measured with a CS (Coronary Sinus) catheter, an HRA (High Right Atrium) catheter, a locally placed catheter, or others.

[0070] Non-cardiac related AF related characteristics include patient demographics, including various details of personal information that characterize the patient, such as, for example, gender, age, place of residence, etc.

[0071] It should be noted that while various examples of types of AF-related features are described above, in different examples, only a subset of these features are used during training and running of the machine learning model. It should be further noted that the examples of AF features disclosed above should not be construed as limiting, and other types of AF-related features are also contemplated to be within the scope of the present disclosure.

[0072] To obtain AF-related features for training the model, data from multiple AF patients (e.g., thousands to tens of thousands or more patients) are collected, and respective parameters are calculated from each patient's data and used as features in the training data set. The AF-related features can be obtained, for example, during an atrial fibrillation ablation procedure. If available, a database of AF patients containing parameters obtained during catheterization and / or ablation can be used.

[0073] The training data set is used to train the machine learning model (block 303b). Different types of classifiers can be generated, including, for example, a classifier that provides data indicative of the probability of long-term success of a pulmonary vein isolation (PVI)-only procedure (e.g., generated using a training data set including data obtained from AF patients treated with a PVI-only procedure), a classifier that provides data indicative of the probability of long-term success of a PVI-plus procedure (e.g., generated using a training data set including data obtained from AF patients treated with a PVI-plus procedure), and a classifier that provides data indicative of the probability of long-term success of both PVI-only and PVI-plus (e.g., generated using a training data set including data obtained from a population of AF patients treated with either a PVI-only or a PVI-plus procedure).

[0074] According to some examples, the machine learning model is implemented as a supervised model, and patients (data points) in the training data set are tagged according to the observed long-term success or failure of each of the PVI-only treatments, which represents the ground truth. This approach requires obtaining valid information from a sufficiently large cohort of AF patients who show the observed long-term success or failure of the PVI-only treatment. In other examples, the data points in the training data set are tagged according to the observed long-term success or failure of each of the PVI-only treatment and / or the PVI plus treatment.

[0075] The training can be performed, for example, by computer 210 or by another computer remotely connected to computer 210 (e.g., by ML classifier training module 231, which can be configured to receive a training data set and train the model). The trained ML model can be stored in computer data repository 235 and made available to be used to select an appropriate ablation technique for an AF patient (block 305b).

[0076] The above approach can be implemented when such a cohort of AF patients is actually available. However, when this information is not available, technical challenges arise that preclude the implementation of such machine learning models. To overcome this technical shortcoming, an alternative approach is proposed herein in which the machine learning model is trained not on the actual observed outcomes of treatment, but on the decisions made by the physicians who examine AF patients and recommend (and possibly administer) ablation procedures, thus overcoming technical setbacks due to insufficient training data. This alternative approach is based on an assumption consistent with the "wisdom of crowds" (WoC) principle. According to this principle, at a large scale (e.g., including hundreds of thousands or millions or more unrelated cases), the majority of individuals make the correct decision.

[0077] The use of physician recommendations in training ML models addresses the technical issue of lack of data related to real-world outcomes of AF treatment and leverages expert clinical judgment as a reliable surrogate for outcome data, thereby facilitating the generation of robust and effective ML models despite limited availability of outcome data.

[0078] In this context, the "individual" is a physician, and the "decision" at hand concerns a recommendation to perform PVI only or a PVI plus procedure, and in some instances, what type of PVI plus procedure to perform. Thus, here, the ground truth is the recorded medical recommendation, not the observed success or failure, but is believed to be a good predictor of success or failure, and can be used to select an appropriate ablation technique.

[0079] According to some examples, a database of patients that meet certain requirements is used for this purpose. These requirements may include, for example: Size: The database is large enough to provide statistically viable analyses and store information on many different AF patients, such as thousands or tens of thousands of AF patients and many different physicians. Diversity: The database contains many AF patients with various paroxysmal and PS-AF cases that are characterized by different features. These features may include various atrial geometries, different voltage patterns (i.e., distribution of low voltages over different regions of the atrium), as well as different ages and genders. The diversity of the database can be determined using known statistical methods. Independence: The database contains many patients with similar AF characteristics who were treated independently by different physicians at different medical institutions. The similarity between different AF patients in the database can be determined by applying clustering mechanisms such as Gaussian Mixture Model (GMM), spectral clustering, or K-means to the data.

[0080] According to one approach, training can be based on the actual opinions of a cohort of physicians presented with different AF patients to obtain the opinions of many physicians on each case. This approach can use either binary or multi-class classification. This approach can be employed when the data includes recommendations made by many physicians on specific "real" patients.

[0081] However, such data is often not available, so a different approach is proposed to overcome the lack of data. Since the database is characterized on the one hand by a high diversity in the recorded cases and on the other hand by similarities between the cases, the database can actually be considered to contain groups (or classes) of patients with similar AF characteristics, with each group corresponding to a specific AF patient ("virtual patient"). The inventors have found that independent medical opinions made by many different physicians on similar cases, all assigned to the same group, can be viewed as if they were made on the same AF patient. Thus, for the purpose of training machine learning models, the data can be viewed as if many different physicians had given medical opinions to the same patient. As mentioned above, it can be assumed that on a large scale, the majority vote of the medical opinions made by physicians on each group of patients is likely to be the option of choice for a concrete patient.

[0082] According to the WoC approach, the machine learning model can be implemented, for example, as a supervised model, where patients in the training dataset are tagged according to their respective medical opinions made by different doctors recommending PVI only or PVI plus treatment, representing the ground truth.

[0083] Regardless of whether the first approach is used, in which the probability of success is based on the observed long-term success of PVI alone or PVI plus treatment, or the second approach (WoC), in which treatment recommendations are based on the treatment selection made by the physician, a machine learning classifier is trained using a training dataset to correlate between various combinations of AF-related features and the respective predictions. The training dataset can be used to train a suitable classifier machine learning model, such as a support vector machine (SVM), random forest, or logistic regression. If the training dataset is large enough, a deep neural network, such as a convolutional network with dense layers, can be used for classification.

[0084] The ML model can be trained to provide one or more training outputs. According to one example, the ML model can be trained to provide a binary classification that provides a recommendation indicating which approach is likely to be effective for treating AF for a particular patient, PVI only or PVI plus.

[0085] When implementing the ML model according to the WoC approach, according to one example, each patient is individually tagged according to the actual treatment given to the patient or according to the recommended treatment, e.g., 0 for PVI only, 1 for PVI plus.

[0086] If a multi-class classifier is trained that includes a PVI-only class and several PVI-plus classes, each corresponding to a different type of additional ablation line (e.g., additional ablation lines at anatomical sites of the heart other than the pulmonary veins in the left atrium), tagging can similarly be done according to each patient's individual treatment and / or recommendations.

[0087] According to an example where the patients in the training dataset are clustered into groups (using the clustering mechanism described above) according to the similarity between their AF features, with each group representing a type of patient (a "virtual patient"), the tagging can be done according to the population distribution of decisions made by all physicians on the different virtual patients.

[0088] Consider a simplified example: assume there are two classes of patients (virtual patients), each corresponding to patients with a common AF-related characteristic, and there are two physicians with the following recommendation distributions:

[0089] [Table 1]

[0090] According to the collective tagging approach, the tagging of a patient is determined according to the combined information from all physicians (e.g., by calculating a weighted average). This collective tagging approach reflects a consensus or averaged decision and leverages the combined expertise of the physicians for a more balanced and robust tagging of patient clusters.

[0091] According to some examples, the machine learning model is implemented as an ensemble of supervised machine learning classifiers, with each classifier in the ensemble representing a virtual doctor that provides a respective recommendation. In some cases, the classifiers in the ensemble may all be the same type of ML model, while in other cases, they may be different types of models (e.g., random forest, XGBoost, etc.). The training set may be split (e.g., randomly) into groups, and each classifier is trained using a subset of the data points (patients) in the group. During inference, each classifier provides a respective recommendation, and the final recommendation is determined based on a majority vote.

[0092] If the training dataset is large enough and contains enough recommendations made by different doctors, laboratories, or other medical institutions (occupying multiple doctors), the training dataset can be split into groups according to doctors or laboratories. This allows for the creation of separate ML models, each representing a "virtual doctor" and associated with a real doctor or laboratory.

[0093] An example of training multiple models with the WoC method is shown in the flow chart in Figure 6. Of note, the same type of ensemble can be used using a cohort of patients with recorded observations of PVI treatment success or failure, if such data is available.

[0094] In block 601, a training dataset of AF patients is generated. Each patient is associated with a respective AF-related characteristic and a PVI treatment recommendation. The patients are divided into subgroups, and in some examples, each subgroup includes patients with similar AF characteristics (block 603). As discussed above, in some examples, each subgroup includes patients who have been recommended AF therapy by a particular physician or laboratory.

[0095] Patients within each subgroup are tagged according to observed recommendations made by their respective physicians or laboratories (block 605). Individual or collective tagging techniques can be implemented.

[0096] These subgroups are then used to train respective classifiers for each physician-related or laboratory-related subgroup of patients (block 607), resulting in multiple classifiers, each corresponding to a specific physician or laboratory decision-making process, effectively acting as virtual physicians providing predicted recommendations for patients.

[0097] This approach allows for personalized and context-specific predictions, as each model is fine-tuned based on the medical preferences and opinions of an individual physician or laboratory. By leveraging these specialized models, the system can enhance the accuracy and relevance of its recommendations, ultimately leading to better patient outcomes and more efficient healthcare delivery, while overcoming the technical challenges mentioned above that stem from the lack of data on completed AF treatments.

[0098] Moreover, in both approaches (observed success and WoC), an additional output of the model includes data indicating the importance of different AF-related features in predicting the final ML output related to the recommended treatment. In the context of ML models, individual features or any combination of AF-related features (e.g., a specific single geometric parameter, a specific complex geometric parameter, a combination of two or more parameters of any type, voltage AF-related features, etc.) may be important for determining the recommended treatment. The ML models disclosed herein can be used to identify one or more features that are most important in providing a recommendation. Providing this information to the physician helps the physician streamline the recommendations made by the model, thereby further assisting the physician in making the final decision on how to treat the patient. This can be achieved by applying feature importance analysis. For example, if a decision tree or random forest model is used, the Gini index can be used to quantify the importance of the features.

[0099] Returning to Figure 3b, in the execution phase (inference), the trained ML model can be applied to determine a recommended treatment for a particular AF patient. Notably, although the training and execution phases are shown sequentially, as is well known in the art, the two phases can be executed asynchronously.

[0100] During execution, the AF-related features obtained for one or more AF patients are applied to the ML model to obtain a respective recommendation for each patient regarding whether to subject the patient to PVI only or PVI plus therapy.

[0101] The execution of the ML model may be part of the process described above with reference to Figure 3a. Assuming that a particular AF patient is a candidate for an ablation procedure, and that the physician would like to receive a recommendation as to which ablation procedure would be more suitable for the patient, the physician may use system 220 to apply the ML model for this purpose.

[0102] For this purpose, AF-related characteristics of the patient are obtained (307b). The AF-related characteristics (e.g., cardiac-related characteristics) can be obtained by applying to the patient a suitable procedure dedicated to measuring the relevant values ​​and, if necessary, performing further processing on the measurements. For example, anatomical prediction parameters can be obtained by applying the procedure described with reference to blocks 301 to 309 of FIG. 3.

[0103] A number of anatomical prediction parameters, both singular and composite, may be calculated (e.g., by the geometric parameter calculator 227) to provide a number of AF-related features that are recorded as part of the input dataset of AF-related features for the ML model.

[0104] Other types of AF-related features can be obtained by applying additional procedures to the patient or related data. For example, different types of voltage-related features can be obtained by performing different voltage measurements on the patient's heart and adding the obtained voltage values ​​to the input data set. Voltage and cycle length measurements can be performed using appropriate medical equipment, for example, using an intracardiac electrocardiogram operably connected to the system 200. Patient demographics can be obtained, for example, from the patient's medical records.

[0105] The ML model is applied to the input dataset of AF-related features obtained from the patient to obtain an ML model output (block 309b). The ML model output is provided including a recommendation indicating whether the patient should be administered PVI or PVI plus (block 311b). As described further above, in some examples, the ML model provides additional information indicating which additional treatment is the recommended treatment for the patient with respect to the PVI plus recommendation. As described above, in some examples, the ML model also provides information indicating which of the input features contributed most significantly to the model's output. In some examples, the process continues by administering the appropriate treatment to the patient according to the recommendation (block 313b).

[0106] FIG. 7 illustrates a schematic of inference using an ensemble of "virtual doctor" classifiers, where an ML model containing the ensemble of classifiers is used to generate treatment recommendations for AF patients. Each of the machine learning models in the ensemble is applied to the AF-related features of the patient to obtain a respective output. Once the outputs from all classifiers in the model are obtained, a majority vote is applied to determine the appropriate recommendation accordingly.

[0107] 3b, blocks 309b and 311b are performed by the treatment recommendation module 229. The treatment recommendation module 229 can include, for example, an ML classifier module configured to apply an ML classifier (or an ensemble of classifiers) to the patient's AF-related features, as described above.

[0108] It is further noted that U.S. Patent Application No. 18 / 207854 describes (e.g., with reference to Figs. 2 and 3) a system (hereinafter "rendering system") for rendering a recommended ablation line on an anatomical structure. In some examples, the system 200 disclosed herein can be incorporated as part of the rendering system, providing the rendering system with the ability to generate a recommendation as to whether a PVI-only treatment is suitable for the patient. In other examples, the system 200 can be implemented as a stand-alone system separate from the rendering system described in U.S. Patent Application No. 18 / 207854. In any event, the system 200 and the rendering system can share certain components in some examples. For example, the medical imaging system 20 and the automatic map segmenter 22 shown in Fig. 2 of U.S. Patent Application No. 18 / 207854 correspond to the medical imaging system 220 and the automatic map segmentation module 221, respectively, of Fig. 2 of the present application.

[0109] 8 is a schematic diagram of a computer system combining the functionality of the system 200 disclosed herein with the deep learning system 10 disclosed in WO 18 / 207854 and WO 63 / 350983. In FIG. 8, primed reference numbers identify elements disclosed in WO 18 / 207854 and WO 63 / 350983, and non-primed reference numbers identify corresponding elements disclosed in the present application. As described in detail in WO 18 / 207854, the system of FIG. 8 includes three sections: an ablation line trainer 12 that builds trained deep learning units for an ablation line proposer 13; an ablation line recommender 14 that uses the ablation line proposer 13 to generate ablation line recommendations; and an ablation line guider 16 that provides ablation line guidance during a procedure. The ablation line trainer 12 may receive an anatomical map of the heart on which recommended ablation lines may be rendered, and may train a machine learning model utilized by the ablation line proposer 13 to output suggested ablation lines according to the patient's cardiac anatomy.

[0110] In some examples, the ablation line recommender 14 receives data indicative of a recommended ablation procedure from the treatment recommendation module 229 and uses this information when recommending an ablation line, where the recommended ablation line is selected to fit the recommended ablation procedure. For example, in the case of treatment of PVI only, the recommended ablation is appropriate for treatment of PVI only. The ablation line guider 16 provides ablation line guidance to the physician during the procedure. It can receive an anatomical map of the heart from the medical imaging system 20 during the procedure and can include an automatic map segmenter 22 to indicate various elements of the heart during the procedure. EXAMPLES

[0111] Table 1 shows the results of the statistical analysis carried out by the inventors, which supports the correlation between certain geometric characteristics of the (left) atrium and the prediction of long-term success or failure of PVI-only treatment. The data includes 95 cases of AF patients who underwent PVI-only treatment. Long-term failure of PVI-only treatment was defined according to a composite of the following three outcomes: 1) repeat ablation, 2) hospitalization for direct current cardioversion, or 3) AF-related hospitalization. Patients who encountered any one of the above three conditions within 12 months after PVI-only treatment were classified into the long-term failure subset. According to the above definition, 16 subjects out of 95 showed long-term failure.

[0112] Table 1 shows the basic atrial parameters (singular parameters) and their relationship to long-term success prediction. Table 1 (and Table 2) shows the P-values ​​for each parameter, indicating the difference between long-term success and long-term failure values. For example, posterior height and posterior width based on geodesic distance are important in distinguishing long-term success from long-term failure in persistent atrial fibrillation (PS-AF) subjects based on PVI alone.

[0113] [Table 2] Geo = Geodesic Euc = Euclid ANT=front wall POST=back wall RLAT = right lateral wall or septum LLAT=left wall postWidth - Back wall width

[0114] In the left column of the table, different parameters (characteristics) of the left atrial expansion are listed. Each parameter corresponds to a specific anatomical part of the heart and is calculated based on one or more dimensions of the anatomical part. The correlation between the values ​​of the different parameters (obtained before the treatment) and the long-term success or failure of the PVI only procedure was investigated. The listed parameters include: anterior (ANT) length and width, left atrial posterior wall (POST) length and width, roof length and width, right side wall (RLAT) length and width, inferior length and width, LLAT width and length.

[0115] Table 2 shows the complex anatomical prediction parameters generated by multiplying pairs of singular parameters.

[0116] [Table 3]

[0117] Based on this study, the most notable specific parameter was the right wall width, which increased significantly (p-value < 0.005) from 37.5 [34.3-42.6] mm in patients with successful PVI procedures to 43.2 [39.2-47.4] mm in patients with unsuccessful procedures. This is further demonstrated in Figure 9, which shows a comparison of the RLAT width in patients with long-term (LT) failure and long-term success of PVI only procedures.

[0118] We also found an increase in posterior wall height and width, as well as an increase in roof width, for subjects who failed PVI procedures.By multiplying each pair of statistically significant parameters, we obtained a new set of parameters that were more significant in distinguishing between long-term failure and long-term success.

[0119] Experiments to verify WoC-based ML models An experiment was conducted to verify the effectiveness of the ML model trained using the WoC approach. A cohort of real AF patients who received either PVI only or PVI plus treatment was collected along with their outcomes, specifically AF-free and arrhythmia-free duration. The ML model was then applied to the AF-related features of this cohort of patients, and the respective recommendations (PVI or PVI plus) were obtained from the model. The patients were divided into four groups based on the recommendations of the ML model and the actual treatments the patients received. 1.Patients recommended for PVI by the ML model who underwent PVI. 2. Patients recommended for PVI by the ML model who underwent PVI plus. 3. Patients recommended PVI plus by the ML model who underwent PVI plus. 4. Patients who underwent PVI and were recommended PVI plus by the ML model.

[0120] Figure 10 presents a graph showing the survival probability of AF patients over time, defined as AF- and arrhythmia-free periods. Table 3 below details the probability of achieving a one-year AF- and arrhythmia-free period for each group of patients based on observed outcomes.

[0121] [Table 4]

[0122] Analysis reveals that alignment between the ML model recommendations and the actual therapy administered correlates with improved outcomes in terms of 1-year freedom from AF and arrhythmia. Patients for whom the ML model recommended PVI and who underwent PVI had the highest 1-year AF- and arrhythmia-free probability (70.5%). Patients recommended for PVI by the model, but those who underwent PVI plus had a slightly lower 1-year AF- and arrhythmia-free probability (66.1%). Patients who were recommended and treated with PVI plus had a 50.4% one-year AF- and arrhythmia-free probability. Patients who received PVI plus but not PVI alone had the lowest 1-year AF- and arrhythmia-free odds (46.1%).

[0123] Compare the results: Group 1 vs. Group 2: Patients who were recommended PVI and received PVI (Group 1) had a higher 1-year AF- and arrhythmia-free probability (70.5%) compared with 66.1% for patients who were recommended PVI but received PVI plus (Group 2). This indicates that following the ML model recommendation for PVI results in better outcomes than deviating from the recommendation and performing PVI plus.

[0124] Group 3 vs. Group 4: Patients who were recommended PVI plus and received PVI plus (Group 3) had a higher 1-year AF- and arrhythmia-free probability (50.4%) compared with 1-year AF- and arrhythmia-free probability of 46.1% for patients who were recommended PVI plus but received PVI (Group 4), suggesting that following the ML model recommendation for PVI plus provides better outcomes than performing PVI instead.

[0125] These results, as shown in the survival probability graph in Figure 10 and detailed in Table 3, suggest that the ML model recommendations are strongly correlated with treatment success. Adherence to the model's guidance can enhance clinical outcomes, highlighting the critical importance of following the ML model recommendations to achieve optimal clinical outcomes.

[0126] The various example logic blocks, modules, and algorithm steps described with respect to the examples disclosed herein may be implemented as electronic hardware, computer software, or a combination of both. To clearly illustrate this interchangeability of hardware and software, the various example components, blocks, modules, and steps have been described above generally in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the particular application and design constraints imposed on the overall system. The described functionality may be implemented in various ways for each particular application, and such implementation decisions should not be interpreted as causing a departure from the scope of the present disclosure.

[0127] It will also be appreciated that the system according to the present disclosure may be implemented, at least in part, on a suitably programmed computer. Similarly, the present disclosure contemplates a computer program readable by a computer for performing the methods of the present disclosure. The present disclosure further contemplates a non-transitory computer readable memory tangibly embodying a program of instructions executable by a computer for performing the methods of the present disclosure.

[0128] Those skilled in the art will readily appreciate that various modifications and changes can be made to the embodiments of the invention described above without departing from the scope of the invention as defined in the appended claims.

[0129] [Embodiment] (1) A computer-implemented method for selecting an atrial fibrillation (AF) therapy, the method comprising: obtaining a 3D anatomical map of at least one subcompartment of the patient's heart; Segmenting the anatomical (3D) map, thereby obtaining a plurality of anatomical regions of the at least one subcompartment; and determining at least one respective dimension for at least a portion of the plurality of anatomical sites; determining one or more anatomical prediction parameters, each anatomical prediction parameter being determined based on geometric data collected from an anatomical region of the patient's heart; generating a prediction indicative of the likelihood of long-term success of a pulmonary vein isolation (PVI) only procedure and / or the likelihood of long-term success of a PVI plus procedure applied to the patient's heart based on the one or more anatomical prediction parameters; and generating an AF therapy recommendation for the patient based on the prediction, the AF therapy recommendation being selected from a pulmonary vein isolation (PVI)-only procedure and a PVI-plus procedure. (2) generating the prediction, obtaining AF-related features of the patient including cardiac-related features characterizing various cardiac-related attributes of the patient, the cardiac-related features being of at least the following types: a. the one or more anatomical prediction parameters; and b. one or more voltage-related characteristics; applying a machine learning (ML) model to the AF-related characteristics of the patient, the ML model being trained using a training dataset including a plurality of data points, each data point corresponding to an AF patient associated with an AF-related characteristic and an observed long-term success of an AF treatment administered to the patient, the AF treatment including at least only PVI, thereby training the ML model to provide data indicative of a recommended AF treatment for the given patient based on the AF-related characteristics of the given patient during inference; and generating a machine learning model output indicating a recommended AF treatment selected from the group including at least PVI only and PVI plus treatment. (3) The method of embodiment 2, wherein the machine learning model output includes information indicating the importance of each of the different AF-related features in determining the recommended AF therapy. (4) The method of embodiment 2, wherein the ML model is trained to provide data indicative of the probability of long-term success of a pulmonary vein isolation (PVI) only procedure. (5) The method of embodiment 2, wherein the ML model is trained to provide data indicative of the probability of long-term success of a pulmonary vein isolation (PVI) plus procedure.

[0130] (6) The method of embodiment 2, wherein the ML model is trained to provide data indicative of the probability of long-term success of pulmonary vein isolation (PVI) only treatment and PVI plus treatment. (7) The method of any one of embodiments 1 to 6, wherein the ML output including the recommendation to apply a PVI plus treatment to the patient also includes information indicating the specific type of PVI plus treatment recommended. (8) A method according to any one of embodiments 1 to 7, comprising performing a procedure on the patient to obtain the cardiac-related features. (9) The method of any one of embodiments 2 to 8, wherein the AF-related characteristics further include one or more of cycle length, patient medical history, and patient demographics. (10) The method of embodiment 1, further comprising comparing each of the one or more anatomical prediction parameters to a respective threshold value and determining the prediction based on a comparison output.

[0131] (11) The method of any one of embodiments 1 to 10, comprising determining the one or more anatomical prediction parameters by applying a mathematical operation between different dimensions of the anatomical site. (12) A computer program product comprising a computer-readable storage medium carrying a program of instructions that, when read by a computer processor, causes the computer processor to perform the method according to any one of embodiments 1 to 11. (13) A computer system comprising at least one processing circuit configured to perform the operations of the method for enhancing atrial fibrillation (AF) therapy described in any one of embodiments 1 to 11.

Claims

1. 1. A computer-implemented method for selecting an atrial fibrillation (AF) therapy, the method comprising: obtaining a 3D anatomical map of at least one subcompartment of the patient's heart; Segmenting the anatomical (3D) map, thereby obtaining a plurality of anatomical regions of the at least one subcompartment; determining at least one respective dimension for at least a portion of the plurality of anatomical sites; determining one or more anatomical prediction parameters, each anatomical prediction parameter being determined based on geometric data collected from an anatomical region of the patient's heart; generating a prediction indicative of the likelihood of long-term success of a pulmonary vein isolation (PVI) only procedure and / or the likelihood of long-term success of a PVI plus procedure applied to the patient's heart based on the one or more anatomical prediction parameters; generating an AF treatment recommendation for the patient selected from a pulmonary vein isolation (PVI) only procedure and a PVI plus procedure based on the prediction.

2. Generating the prediction comprises: obtaining AF-related features of the patient including cardiac-related features characterizing various cardiac-related attributes of the patient, the cardiac-related features being of at least the following types: a. the one or more anatomical prediction parameters; b. one or more voltage related characteristics; applying a machine learning (ML) model to the AF-related features of the patient, the ML model being trained using a training dataset including a plurality of data points, each data point corresponding to an AF patient associated with an AF-related feature and an observed long-term success of an AF treatment administered to the patient, the AF treatment including at least only PVI, thus training the ML model to provide data indicative of a recommended AF treatment for the given patient based on the AF-related features of the given patient during inference; and generating a machine learning model output indicating a recommended AF treatment selected from a group including at least PVI-only and PVI-plus treatment.

3. The method of claim 2 , wherein the machine learning model output includes information indicating the importance of each of different AF-related features in determining the recommended AF treatment.

4. The method of claim 2 , wherein the ML model is trained to provide data indicative of the probability of long-term success of a pulmonary vein isolation (PVI) only procedure.

5. The method of claim 2 , wherein the ML model is trained to provide data indicative of the probability of long-term success of a pulmonary vein isolation (PVI) plus procedure.

6. The method of claim 2 , wherein the ML model is trained to provide data indicative of the probability of long-term success of pulmonary vein isolation (PVI) only and PVI plus procedures.

7. The method of claim 1 , wherein the ML output including the recommendation to administer a PVI plus treatment to the patient also includes information indicating the particular type of PVI plus treatment recommended.

8. The method of claim 1 , comprising performing a procedure on the patient to obtain the cardiac-related features.

9. The method of claim 2 , wherein the AF-related characteristics further include one or more of cycle length, patient medical history, and patient demographics.

10. The method of claim 1 , further comprising: comparing each of the one or more anatomical prediction parameters to a respective threshold; and determining the prediction based on an output of the comparison.

11. The method of claim 1 , comprising determining the one or more anatomical prediction parameters by applying a mathematical operation between different dimensions of an anatomical portion.

12. A computer program product comprising a computer readable storage medium carrying a program of instructions which, when read by a computer processor, causes said computer processor to perform the method of any one of claims 1 to 11.

13. A computer system comprising at least one processing circuit configured to perform the operations of the method for enhancing atrial fibrillation (AF) therapy according to any one of claims 1 to 11.

Images