Method for analyzing potential ablation therapies

Abstract

The present invention relates to a method for analyzing potential ablation therapies (1), in particular for analyzing potential ablation therapies for a patient (P) having atrial fibrillation, via a control system (2), comprising an analysis step (3) in which the control system (2) applies a trained machine learning model (4) to input data (5) to generate output data (6), the input data (5) comprising electrical biomarkers (7) derived from electrocardiogram data (8) from the patient (P) of at least one potential ablation therapy (1), the potential ablation therapy (1) including at least one potential ablation event (9) with a potential ablation location (10), and the output data (6) comprising a predicted change in the electrical biomarkers (7) following application of the potential ablation therapy (1) to the patient (P), the predicted change being derived from the input data (5) via the trained machine learning model (4).

Description

[Technical field]

[0001] The present invention relates to a method for analysing potential ablation therapies as claimed in claim 1, a computer readable medium storing a trained machine learning model as claimed in the generic claim 13, a control system configured to carry out the above-mentioned method as claimed in claim 14 and a surgical system as claimed in claim 15. [Background technology]

[0002] The method particularly relates to ablation therapy for atrial fibrillation. Atrial fibrillation is the unregulated excitation of atrial muscle cells, electrically. During atrial fibrillation, the atria contribute minimally to the function of the heart. Thus, atrial fibrillation reduces cardiac output, which is not an imminent danger. However, when atrial fibrillation becomes chronic, it is associated with increased morbidity and mortality. One of the treatment options for atrial fibrillation is ablation therapy. Ablation therapy is the destruction of cells that allow reentry of electrical waves, thereby reducing the unregulated excitation of atrial muscle cells.

[0003] The success rate of ablation therapy is related to the location of the ablation. In many cases of atrial fibrillation, the atrial myocytes around the pulmonary vein ostium are ablated. Although this standard therapy has shown good results for many patients, it is not always successful. For patients with complex atrial fibrillation or especially recurrent atrial fibrillation, atrial mapping of the electrical wave front and reentry points is available. However, such mapping is complex and difficult to interpret. Therefore, further methods for the identification and evaluation of potential targets for ablation are needed.

[0004] Known methods such as EP 3744282 use machine learning algorithms to generate cardiac ablation treatment plans. However, applying machine learning to a large number of input variables requires a large training data set. It is even more difficult to design a machine learning architecture that can generate a general treatment plan from rather heterogeneous input data. In theory, the accuracy of machine learning algorithms can be improved by considering a large number of input variables, but in practice, it is not feasible to collect a training data set from scratch by considering a large number of ablation therapies.

[0005] Another known method (WO 2019 / 217430) is related to learning excitation pathways and therapy locations based on cardiac measurements and functional models. Functional models are a good approach to reduce the complexity of machine learning algorithms and the required training data sets. However, the functional understanding of atrial fibrillation is still in progress, and combining functional models with machine learning algorithms is a complex task in itself. Summary of the Invention [Problem to be solved by the invention]

[0006] It is therefore an object of the present invention to provide a method for analyzing potential ablation therapies that is realistically trainable and usable, with manageable complexity. [Means for solving the problem]

[0007] The above mentioned problem is solved by a method as claimed in claim 1.

[0008] The main realization of the present invention is that cardiac pathologies treated by ablation are generally related to electrical malfunctions. Therefore, they are well represented by the results of electrical measurements. It is therefore possible to develop machine learning algorithms based on electrical data, i.e., electrocardiogram data. By deriving electrical biomarkers from electrocardiogram data and using trained machine learning models to predict changes in electrical biomarkers after applying potential ablation therapy, the complexity of the machine learning algorithm is significantly reduced. In particular, by shifting the focus from trying to learn a broadly defined general success of a treatment using a large number of data, the definition of a cost function for training the machine learning model becomes more easily manageable in the present invention. Being able to select a good cost function is key to successfully applying machine learning algorithms.

[0009] A further advantage is that electrical measurements before and after ablation therapy are already available for a large number of patients. By deriving the predicted changes in the electrical biomarkers, it is also possible to present these to the physician, who can then design a suitable overall treatment for the patient based on these changes in combination with his knowledge and experience. This can further improve the overall quality of the treatment as well as the acceptance of the proposed method by physicians and patients.

[0010] In particular, a method is proposed for analyzing potential ablation therapies, particularly for patients with atrial fibrillation, via a control system, the method comprising an analysis step in which the control system applies a trained machine learning model to input data, thereby generating output data, the input data comprising electrical biomarkers derived from electrocardiogram data from a patient of at least one potential ablation therapy, the potential ablation therapy including at least one potential ablation event with a potential ablation location, and the output data comprising predicted changes in the electrical biomarkers following application of the potential ablation therapy to the patient, the predicted changes being derived from the input data via the trained machine learning model.

[0011] According to claim 2, the input data may consist mainly of ECG-derived data or of ECG-derived data and ECG data, which further enhances the above mentioned advantages. Claim 2 also relates to the possibility of analysing the output data to identify potential ablation targets for the patient. The predicted changes in the electrical biomarkers may be a good indication for identifying potential ablation targets and planning an ablation therapy.

[0012] A training step for the machine learning model is proposed in claim 3. The training data set may be derived from a training database that may be retrospectively labeled. This database may be based on an open or commercial database, which can significantly improve the usability of the proposed method.

[0013] In an embodiment as claimed in claim 4, the electrical biomarkers may be derived by the control system from electrocardiogram data in the training database.

[0014] Currently, biomarker responses to single ablation events and to the entire ablation therapy are of particular interest. In the embodiment described in claim 5, a first model may be trained in a training step to derive changes in electrical biomarkers after the application of one potential ablation event, and / or a second model may be trained in a training step to derive changes in electrical biomarkers. In some cases, a single ablation event may have a clearly defined biomarker response, while in other cases, the biomarker response of an ablation therapy is composed of two or more ablation events. In the latter case, a single ablation event may have a synergistic effect, which can be seen by considering the entire ablation therapy.

[0015] Claim 6 relates to the possibility that the first and second models can be applied separately or jointly, in particular to different electrical biomarkers.

[0016] In an embodiment as claimed in claim 7, the training data set may comprise focal source ablation events. The focal source ablations resulting from the mapping are more specific and provide a wider range of ablation locations, allowing the proposed method to identify potential focal sources.

[0017] Claim 8 relates to a derivation step for automatically deriving electrical biomarkers from electrocardiogram data. The electrical biomarkers may be derived from electrocardiogram data by non-machine learning algorithms. Here, known algorithms can be used to derive a number of electrical biomarkers from electrocardiogram data. In particular, what is cited in claim 9 applies in the proposed method. However, it is equally preferred to use a machine learning model for deriving one or more electrical biomarkers. This machine learning model may be part of a machine learning model for deriving the expected changes in the electrical biomarkers, which allows predictions based directly on electrocardiogram data.

[0018] The embodiment as defined in claim 10 considers in particular the possibility of using primarily the last ablation events of the ablation therapy from the training database or of using ablation events throughout the ablation therapy, depending on the available training data set.

[0019] The embodiment of claim 11 relates to determining a classification of potential ablation therapies by use of output data, which may also be based on machine learning algorithms. Additionally, analysis of electrical biomarkers based on non-machine learning algorithms may be utilized in this classification.

[0020] According to claim 12, the method may provide an output of a classification of potential ablation therapies, in particular from online ECG data measurements, which may be used to identify or verify potential ablation locations even during surgery.

[0021] Of equal importance, another teaching as claimed in claim 13 relates to a computer-readable medium on which the trained machine learning model is stored. All explanations given with regard to the proposed method, in particular with regard to the training step, are fully applicable.

[0022] Another teaching, also of the same importance, as set forth in claim 14, relates to a control system arranged to carry out the proposed method. All explanations given with respect to the proposed method are fully applicable.

[0023] Another teaching, also of the same importance, as set forth in claim 15, relates to a surgical system connected to or forming part of the proposed control system. All the explanations given with respect to the proposed method are fully applicable.

[0024] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. [Brief description of the drawings]

[0025] [Figure 1] FIG. 1 shows a schematic diagram of the proposed method. [Diagram 2] FIG. 1 illustrates training of a machine learning model. [Diagram 3] FIG. 1 illustrates application of a trained machine learning model. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0026] The proposed method is used to analyze potential ablation therapies 1. In the illustrated preferred case, the proposed method is used for a patient P with atrial fibrillation. The potential ablation therapies 1 are analyzed via a control system 2. In FIG. 1, the control system 2 is shown diagrammatically as a dedicated control system 2 comprising hardware connected to the patient P. The control system 2 may be a general-purpose computer, a cloud-based computer system, etc., and may be configured to implement the proposed method.

[0027] The method comprises an analysis step 3 in which the control system 2 applies a trained machine learning model 4 to input data 5, thereby generating output data 6.

[0028] Currently, the term "trained machine learning model" describes at least a minimal amount of data necessary to apply the results of training a machine learning algorithm. The trained machine learning model 4 may exist in a compressed representation and does not have to be usable by any general-purpose computer. The trained machine learning model 4 may, for example, comprise weights of an artificial neural network predefined in another manner. The operations necessary to apply the trained machine learning model 4 may be stored in the control system 2 and / or may be part of the trained machine learning model 4. However, in a preferred embodiment, the trained machine learning model 4 comprises program code and is applicable to the input data 5 by a general-purpose operating system.

[0029] The input data 5 comprises electrical biomarkers 7 derived from electrocardiogram data 8 from a patient P of at least one potential ablation therapy 1. The potential ablation therapy 1 comprises at least one potential ablation event 9 with a potential ablation location 10. The input data 5 is shown diagrammatically on the left side of FIG. 1. The electrical biomarkers 7 are described in more detail below. A potential ablation therapy 1 may comprise only one potential ablation event 9 or may comprise two or more potential ablation events 9. The potential ablation therapy 1, the potential ablation events 9 and the potential ablation locations 10 may be stored in any suitable data structure. This data structure is preferably standardized. In FIG. 1, the values ​​x, y, z represent each potential ablation therapy 1, for example the coordinates of a potential ablation location 10 in the potential ablation therapy 1, which are shown as an example for illustrative purposes. Three potential ablation therapies 1 are shown in FIG. 1, one of which comprises two potential ablation events 9. The term "potential" means that this ablation therapy 1 does not necessarily have to be applied to patient P. The currently proposed method rather serves to predict whether it is promising to apply this potential ablation therapy 1 to patient P.

[0030] The ablation event 9 is formed by one or more regions within an area of ​​interest of a common focal source or by one or more regions that otherwise share a common area, for example a pulmonary vein ostium. An ablation event may be, for example, one application of a cryo-catheter.

[0031] The output data 6 comprises predicted changes in electrical biomarkers 7 after application of a potential ablation therapy 1 to a patient P. The predicted changes are derived from the input data 5 via a trained machine learning model 4. In Figure 1, this change is illustrated in a set of changed electrical biomarkers 11. Additionally, for illustrative purposes, changes in electrocardiogram data 8 are illustrated.

[0032] The electrocardiogram data 8 is preferably surface or intracardiac electrocardiogram data and may be measured by electrodes 12 connected to the control system 2 .

[0033] These electrical biomarkers 7, or some of these electrical biomarkers 7, may be time-related and / or time-independent. By selecting the electrical biomarkers 7 as the focus of the machine learning mechanism and looking at predicted changes in the electrical biomarkers 7, a number of machine learning architectures are made particularly suitable for the proposed method. It should be understood that the following list of preferred architectures is not exhaustive.

[0034] The trained machine learning model 4 is derived by training in a commonly known manner. The trained machine learning model 4 is part of a machine learning mechanism that comprises training and applying the trained machine learning model 4.

[0035] The underlying architecture may be a fixed-input architecture. Preferably, the fixed-input architecture is a convolutional neural network, a feed-forward network, a deep residual network, or a classification architecture such as a support vector machine or a Bayesian classifier. The fixed-input architecture may be coupled to an autoencoder.

[0036] To generate a fixed input, the electrocardiogram data 8 can be divided into time windows of fixed length, each comprising one or more QRS complexes. Additionally or alternatively, the electrical biomarkers 7 may not be time-related. For 2D input data 5, convolutional networks may be used. However, 1D implementations may also be applied, which are well suited to time series. Depending on the input data 5, feedforward networks are preferred, which may form a particularly simple approach. In particular, when the input data 5 comprises long time series, solutions to the vanishing gradient problem, such as deep residual networks, may be used. In general, the electrocardiogram data 8 before the ablation therapy 1 may be time-related, where early and late events are equally important. The classification architectures mentioned may be particularly advantageous for a large number of correlated electrical biomarkers 7.

[0037] Also, recurrent architectures are preferred, in particular recurrent neural networks, and more preferably architectures based on long-short-term memory or gated recurrent units, which can be used, for example, depending on the length of the ECG data 8 and / or the electrical biomarkers 7. Another preferred architecture is an attention architecture, in particular a transformer architecture and / or a non-recurrent architecture with an attention mechanism. If it is unclear in which part of the ECG data 8 the relevant information for predicting the success of a potential ablation therapy 1 may be located, it may be advantageous to use low frequencies of the ECG data 8 and derive long sequences of electrical biomarkers 7. In particular, if the relevant information is sparsely distributed, attention architectures are well suited.

[0038] The cost function for training may be based on known responses to ablation therapy 1, as described below. The response may be compensated for by the time interval to ablation therapy 1 and / or the ablation event 9. It is preferable to predict long term changes in the electrical biomarkers 7 and / or immediate changes in the electrical biomarkers 7 immediately following ablation therapy 1.

[0039] Here, preferably, the input data 5 is mainly composed of electrocardiogram-derived data, in particular electrical biomarkers 7. Alternatively, the input data 5 may be mainly composed of electrocardiogram-derived data and electrocardiogram data 8. Here, further below, the term "mainly" means that the listed data form the core of the respective data or of the algorithm using the respective data, but other data may also be present. This takes into account that it is possible to include additional data in a machine learning algorithm without significantly affecting the result. The definition of the term "mainly" should therefore be understood as a functional definition.

[0040] Tailored use of non-electrical data 8 may be advantageous to the proposed method. The proposed method may comprise analysing the output data 6 to identify potential ablation targets for the patient P.

[0041] The training of the machine learning model 4 will now be described with reference to Fig. 2. Preferably, the trained machine learning model 4 has been trained or is trained in a training step 13 on a training data set 14 by the control system 2. The training data set 14 may be derived from a training database 15 comprising ablation therapy data 16, the ablation therapy data 16 comprising ablation events 9 and electrocardiogram data 8 determined before and after ablation therapy 1 and / or one or more ablation events 9.

[0042] Predicting changes in electrical biomarkers 7 has the advantage that a database with different ablation therapies 1 can be used for training. Therefore, preferably, the training step 13 comprises retrospectively labelling a training database 15, which generates at least a part of a training data set 14.

[0043] The labelling may in principle be performed at least partly manually. However, it is preferred that the control system 2 derives the electrical biomarkers 7 at least partly, in particular mainly or completely, from the electrocardiogram data 8 of the training database 15 before and after the ablation therapy 1 and / or one or more ablation events 9. The automatic derivation of the electrical biomarkers 7 from a database already encompassing the changes in the electrocardiogram data 8 after the ablation therapy 1 allows an efficient approach for unsupervised learning.

[0044] In a preferred embodiment, the potential ablation therapy 1 includes at least two potential ablation events 9 with different ablation locations 10. Then, in a training step 13, the control system 2 trains a first model 17 to derive a predicted change in the electrical biomarker 7 after application of one potential ablation event 9 and / or trains a second model 18 to derive a predicted change in the electrical biomarker 7 after application of a potential ablation therapy. In some cases, the outcome of the complete ablation therapy 1 may be different than the sum of the outcomes of the single ablation events 9. In other cases, a single ablation event 9 may be the source of the success of the ablation therapy 1. Thus, looking at the individual ablation events 9 and ablation therapies 1 may result in improved analysis.

[0045] It should be noted that the training does not have to be fully automatic. The training step 13, and thus the trained machine learning model 4, may comprise both an automatic training step and a manual fine-tuning step. In general, the trained machine learning model 4 may be part of a model that comprises the trained machine learning model 4 and further manually derived models.

[0046] The proposed method allows for an efficient definition of a cost function for training the trained machine learning model 4. This is advantageous because the cost function can be used with many different architectures. In one preferred embodiment, the architecture behind the machine learning itself is derived and / or modified during the training step 13. The training step 13 may include a neural architecture search and / or may use guided or automated hyperparameter tuning.

[0047] For faster training of the first model 17 and the second model 18, a general model may be trained, and the first model 17 and the second model 18 may be trained by transfer learning from the general model.

[0048] Here, preferably in the analysis step 3, the control system 2 applies the first model 17 and the second model 18 to the input data 5 to derive predicted changes in the electrical biomarkers 7. Preferably in the analysis step 3, the control system 2 applies the first model 17 and the second model 18 to different subsets of the electrical biomarkers 7 to derive predicted partial changes in the electrical biomarkers 7 after one or more potential ablation events 9 and after a potential ablation therapy 1. The control system 2 can then derive predicted changes from the predicted partial changes in the electrical biomarkers 7.

[0049] With respect to the training data set 14, the ablation therapy 1 in the training data set 14 may mainly comprise at least one focal source ablation event. Preferably, the training data set 14 may comprise an ablation therapy 1 that mainly includes focal source ablation events. A focal source ablation event is an ablation event 9 in which at least one focal source is ablated, where mapping or other determination of the focal source has been performed prior to the ablation therapy 1 to identify the focal source. Nevertheless, an advantage of the proposed method is that mapping is not always necessary, or even rarely unnecessary. Also, even if the training data set 14 comprises focal source ablation events derived from mapping, it is not necessary to use actual information about the mapping in the training data set 14.

[0050] The method may comprise a derivation step in which the control system 2 applies an algorithm to the electrocardiogram data 8 to automatically derive at least a portion of the electrical biomarkers 7 from the electrocardiogram data 8. Preferably, in the derivation step, the control system 2 applies one or more non-machine learning algorithms to the electrocardiogram data 8 to derive the one or more electrical biomarkers 7. Additionally or alternatively, in the derivation step, the control system 2 may apply one or more secondary machine learning models to the electrocardiogram data 8 to derive the one or more electrical biomarkers 7.

[0051] Here, preferably the same algorithm or the same architecture is used for training and also for deriving the electrical biomarkers 7 in this derivation step.

[0052] The secondary machine learning model may be based on any suitable architecture, preferably based on a convolutional network or an autoencoder. It is also possible to include one or more secondary machine learning models in the machine learning model 4 and to apply and / or train these models together. For example, feature extraction layers are known.

[0053] The electrical biomarkers 7 may comprise at least one of the following: cardiac cycle length, electrocardiogram morphology classification, signal amplitude, cardiac rhythm classification, peak timing values, in particular R-peak timing values, peak variability, in particular R-peak variability, or quantitative changes between heart beats, in particular between consecutive heart beats, e.g. mean square error values. Additionally or alternatively, the input data 5 may comprise at least one of patient data, preferably biomarkers representative of age, sex, risk factors and / or biological biomarkers, preferably blood pressure data and / or blood measured biomarkers.

[0054] Cardiac cycle length can be derived from RR interval measurements and / or frequency analysis. Morphology can be identified by scores, values ​​such as TT voltage elevation, comparison to QRS templates, etc. A standardized learning mechanism is involved. Amplitude and power can be represented by average values, peak values, etc. Rhythm classification may include regularity scores and / or standard deviation of cycle length.

[0055] Non-electrical biomarkers may be included in the machine learning mechanism to interpret the results of the prediction. Non-electrical biomarkers may be mixed with the electrical biomarkers 7 or may be included only in the output layer, for example. Non-electrical biomarkers may also be used in the derivation step.

[0056] Here, preferably, the trained machine learning model 4, in particular the first model 17, is trained by using primarily the last ablation event 9 of the ablation therapy 1 of the training database 15 as the training dataset 14, or by using ablation events 9 throughout the ablation therapy 1. Here, as mentioned above with respect to the definition of the term "primarily", it would be possible to include in the training a large number of ablation events 9 that were not the last ablation event 9 without any significant adverse effect on the training dataset 14.

[0057] 3 shows a possible application of a first model 17 and a second model 18. First, the first model 17 is used to predict the change in biomarker 7 for an ablation event 9, and then the first model 17 and the second model 18 are used to predict the change in biomarker 7 after a second ablation event 9 that concludes the ablation therapy 1.

[0058] The proposed method may comprise an ablation therapy classification step, in which the control system 2 determines a classification of one or more potential ablation events 9 and / or potential ablation therapies 1, in particular by determining a success score on the probability of successful treatment of the patient P by applying the potential ablation therapies 1 from the output data 6. Preferably, the method may comprise an ablation therapy determination step, in which the control system 2 determines a classification of at least two potential ablation events 9 and / or at least two potential ablation therapies 1 based on the output data 6 and / or in which the control system 2 determines an optimized ablation therapy 1 based on the output data 6.

[0059] The success of a potential ablation event 9 may be classified differently from the success of a potential ablation therapy 1. In particular, the control system 2 or the physician may design a potential ablation therapy by adding potential ablation events 9 targeting different electrical biomarkers. The proposed method may comprise a step of classifying the potential ablation events 9 into predefined classes of potential ablation events 9. These predefined classes may then be defined as classes having one of a group of predefined outcomes for one or more electrical biomarkers 7. It is one of the advantages of the proposed method that it is possible to determine the outcome of a single potential ablation event 9, possibly without the context of a complete potential ablation therapy 1.

[0060] The control system 2 may receive electrocardiogram data 8, preferably from online measurements of the patient P, in particular during surgery, and output a classification of a potential ablation therapy 1. Preferably, the potential ablation therapy 1 is at least partially defined by measuring electrocardiogram data 8 at or near a potential ablation location 10 during surgery. In this way, it becomes possible to identify and / or verify the ablation therapy 1 during surgery. In particular, it becomes possible to include or exclude at least one potential ablation event 9 in the ablation therapy 1.

[0061] The control system 2 may receive electrocardiogram data 8 from the surgical system 19 .

[0062] According to another teaching, a computer-readable medium is proposed on which a trained machine learning model 4 is stored. All references are made to the above description. The trained machine learning model 4 on the computer-readable medium is derived in a training step 13.

[0063] According to another teaching, a control system 2 is proposed which is adapted to implement the proposed method, all of which are referred to above.

[0064] According to another teaching, a surgical system 19 is proposed, which is connected to or forms part of the control system 2. The control system 2 is configured to carry out the proposed method. Reference is made to all the above descriptions. The surgical system 19 is adapted to be used during surgery, adapted to receive online electrocardiogram data 8 during surgery and adapted to display the results of the ablation therapy classification step and / or the ablation therapy decision step. The surgical system may be connected to electrodes 12 for measuring the online electrocardiogram data 8. It may include hardware for processing the electrocardiogram data 8, a display and / or an input unit.

Claims

1. A method for analyzing potential ablation therapies (1), in particular for a patient (P) with atrial fibrillation, via a control system (2), comprising: The method comprises an analysis step (3) in which the control system (2) applies a trained machine learning model (4) to input data (5) thereby generating output data (6); The input data (5) comprises electrical biomarkers (7) derived from electrocardiogram data (8) from a patient (P) of at least one potential ablation therapy (1), the potential ablation therapy (1) including at least one potential ablation event (9) with a potential ablation location (10); the output data (6) comprising predicted changes in the electrical biomarkers (7) following application of the potential ablation therapy (1) to the patient (P); The predicted change is derived from the input data (5) via the trained machine learning model (4). method.

2. the input data (5) is mainly composed of electrocardiogram-derived data, in particular the electrical biomarkers (7), or the input data (5) is mainly composed of electrocardiogram-derived data and electrocardiogram data (8), and / or The method comprises analysing the output data (6) to identify potential ablation targets for the patient (P). The method of claim 1.

3. The trained machine learning model (4) is or has been trained in a training step (13) on a training data set (14) by the control system (2), Preferably, said training data set (14) is derived from a training database (15) comprising ablation therapy data (16) comprising ablation events (9) and electrocardiogram data (8) determined before and after said ablation therapy (1) and / or one or more of said ablation events (9), More preferably, said training step (13) comprises retrospectively labelling said training database (15), thereby generating at least a portion of said training data set (14).

3. The method according to claim 1 or 2.

4. the control system (2) derives the electrical biomarkers (7) at least in part, in particular mainly or completely, from the ablation therapy (1) and / or the electrocardiogram data (8) of the training database (15) before and after one or more ablation events (9); The method of claim 3.

5. the potential ablation therapy (1) includes at least two potential ablation events (9) with different ablation locations (10); In the training step (13), the control system (2) trains a first model (17) to derive a predicted change in the electrical biomarker (7) after application of one of the potential ablation events (9) and / or trains a second model (18) to derive a predicted change in the electrical biomarker (7) after application of one of the potential ablation therapies (1). The method according to claim 3 or 4.

6. In the analysis step (3), the control system (2) applies the first model (17) and / or the second model (18) to the input data (5) to derive the predicted changes in the electrical biomarkers (7); Preferably, in the analyzing step (3), the control system (2) applies the first model (17) and the second model (18) to different subsets of the electrical biomarkers (7) to derive predicted fractional changes in the electrical biomarkers (7) after one or more of the potential ablation events (9) and after the potential ablation therapy (1); the control system (2) deriving the predicted change from the predicted fractional change of the electrical biomarker (7); The method of claim 5.

7. the ablation therapy (1) in the training data set (14) primarily comprises at least one focal source ablation event; Preferably, the training data set (14) comprises an ablation therapy (1) including primarily focal source ablation events.

7. The method according to any one of claims 3 to 6.

8. The method comprises a deriving step in which the control system (2) applies an algorithm to the electrocardiogram data (8) to automatically derive at least a portion of the electrical biomarkers (7) from the electrocardiogram data (8); Preferably, during the deriving step, the control system (2) applies one or more non-machine learning algorithms to the electrocardiogram data (8) to derive one or more of the electrical biomarkers (7); and / or In the deriving step, the control system (2) applies one or more trained secondary machine learning models to the electrocardiogram data (8) to derive one or more of the electrical biomarkers (7).

8. The method according to any one of claims 1 to 7.

9. said electrical biomarkers (7) comprising at least one of the following: a cardiac cycle length, an electrocardiogram morphology classification, a signal amplitude, a signal power, a cardiac rhythm classification, a peak timing value, in particular an R-peak timing value, a peak variability, in particular an R-peak variability, or a quantitative change between heartbeats, in particular between successive heartbeats, e.g. a mean square error value; and / or said input data (5) further comprising at least one of patient data, preferably age, sex, risk factors and / or biological biomarkers, preferably biomarkers representing blood pressure data and / or blood measured biomarkers, 9. The method according to any one of claims 1 to 8.

10. The trained machine learning model (4), in particular the first model (17), is trained by using primarily the last ablation events (9) of the ablation therapy (1) of the training database (15) as the training data set (14) or by using ablation events (9) throughout the ablation therapy (1); 10. The method according to any one of claims 1 to 9.

11. The method comprises an ablation therapy classification step, in which the control system (2) determines a classification of one or more potential ablation events (9) and / or the potential ablation therapies (1), in particular by determining from the output data (6) a success score relating to the probability of successful treatment of the patient (P) by applying the potential ablation therapy (1), Preferably, the method comprises an ablation therapy determination step, in which the control system (2) determines a classification of at least two potential ablation events (9) and / or at least two potential ablation therapies (1) from the output data (6) and / or in which the control system (2) determines an optimized ablation therapy (1) from the output data (6).

11. The method according to any one of claims 1 to 10.

12. said control system (2) receiving electrocardiogram data (8), preferably from online measurements of a patient (P), in particular during surgery, and outputting said classification of potential ablation therapies (1); Preferably, the potential ablation therapy (1) is defined at least in part by measuring electrocardiogram data (8) at or near a potential ablation location (10) during surgery. The method of claim 11.

13. A computer-readable medium having stored thereon a trained machine learning model (4), The trained machine learning model (4) has been derived in a training step (13) according to claim 3 and possibly any one of claims 4 to 12.

23. A computer readable medium comprising:

14. 13. The method according to claim 1, Control system.

15. A surgical system connected to or forming part of a control system (2), The control system (2) is adapted to implement the method according to claim 11 and possibly claim 12, the surgical system (19) is adapted for use during surgery, adapted to receive online electrocardiogram data (8) during surgery, and adapted to display results of the ablation therapy classification step and / or the ablation therapy decision step. Surgery system.

Images