System and method for assessing a cardiac arrhythmia

The instant text concerns a system and method for assessing a cardiac arrhythmia, in particular a bradycardia, a tachycardia, an asystole, or atrial fibrillation.

Active implantable medical devices (AIMD's) in cardiac rhythm management (CRM) applications generally collect ECG or other physiologic data and transmit these data to an external computation system, such as a physician device or to a cloud service. Some implantable medical devices also contain algorithms to detect cardiac arrhythmias from measured signals in the device themselves. Upon detection of a cardiac arrhythmia, these devices can transmit the measured ECG signal with the annotated arrhythmia detection to a secure server that allows trained medical personnel to review the signal and the detected arrhythmia in order to guide patient care. However, these devices may at times falsely detect arrhythmias, thereby sending false signals and detections that increase the burden of review by medical personnel. For example, false detections of atrial fibrillation (AF) by implantable cardiac monitors (ICMs) are a well-known issue in the cardiac rhythm management field.

One reason for false arrhythmia detections may lie in the fact that implantable medical devices generally are designed as low-power devices, and are therefore not well-suited for computationally-intensive arrhythmia-detection algorithms that may be more accurate.

There hence is a general demand for a system and a method for assessing a cardiac arrhythmia which use an implantable medical device, such as a pacemaker device, a defibrillation device or a sensor device implanted into or in proximity to a patient's heart, for sensing a cardiac signal in order to assess the presence of a cardiac arrhythmia based on said cardiac signal. It herein is desirable not to increase the computational demand within the implantable medical device in order to maintain a low power consumption within the - 2 implantable medical device, wherein generally it shall be enabled to reduce the number of false detections.

US 2013/0231947 A1 discloses a system for medical monitoring in which data processing capabilities are distributed among computing devices connected to a network to optimize usage of computational resources. Patient monitoring herein may generally take place by wearable devices, which transmit data to communication devices such as a smart phone, which forwards the data to an Internet cloud.

US 2019/0216350 A1 discloses a system and method for medical premonitory event estimation, the system including one or more processors to perform operations comprising acquiring a first set of physiological information of a subject, and a second set of physiological information of the subject received during a second period of time, and calculating first and second risk scores associated with estimating a risk of a potential cardiac arrhythmia event for the subject based on applying the first and second sets of physiological information to one or more machine learning classifier models. Data herein are detected at the patient by wearable devices.

It is an object of the instant invention to provide a system and method for assessing a cardiac arrhythmia based on a cardiac signal as sensed by an implantable medical device, the system and method herein increasing reliability of an arrhythmia detection while not increasing a computational intensity at the implantable medical device.

In one aspect, a system for assessing a cardiac arrhythmia comprises: an implantable medical device comprising a sensing device for sensing a cardiac signal and a communication device for transmitting a transmit signal comprising information relating to said cardiac signal; and a computation system configured to obtain a multiplicity of features relating to said cardiac signal based on said transmit signal and to derive a classification indication for a cardiac arrhythmia based on said multiplicity of features, wherein the computation system is configured to input said multiplicity of features into a classification model for deriving said classification indication. - 3

Accordingly, within the system an implantable medical device, for example in the shape of a pacemaker device, a defibrillator device, or a sensor device which is implanted into a patient in or in proximity to the patient's heart, is configured to sense a cardiac signal, such as an ECG signal. By means of a communication device the implantable medical device transmits a transmit signal relating to the sensed cardiac signal to an external device, such as an external monitoring device, using for example a close-range communication protocol such as an MICS, Zigbee or BLE protocol. The external monitoring device, which may for example be a user's smart phone or another portable device, serves as a link to forward the transmit signal or a signal derived from the transmit signal to a computation system, such as a server system, connected to the external monitoring device for example by means of a communication network, such as the Internet. The server system herein may represent a cloud system, wherein computational resources for example are distributed via multiple servers in order to provide computational capabilities for processing data as received from the implantable medical device.

For example, the implantable medical device may sense a cardiac signal and derive, from the sensed cardiac signal, an indication whether a cardiac arrhythmia potentially is present. The implantable medical device may then send a transmit signal relating to the sensed cardiac signal for reception and processing by the computation system, wherein the implantable medical device may in addition provide the indication of the identified cardiac arrhythmia to the computation system. The computation system then, based on the transmit signal as received from the implantable medical device, is enabled to assess the potential cardiac arrhythmia signal in order to provide a classification indication which confirms the cardiac arrhythmia or discards the cardiac arrhythmia as identified by the implantable medical device.

By means of the computation system, computations for an accurate identification of a cardiac arrhythmia may be carried out externally from the implantable medical device, such that the implantable medical device may be designed as a low power device having reduced computational capability. Within the implantable medical device, hence, algorithms of limited complexity are carried out, wherein subsequently data relating to an identified - 4 cardiac arrhythmia are transmitted towards the computation system for confirmation of the cardiac arrhythmia.

The computation system, for identifying a cardiac arrhythmia, makes use of a classification model as it is known from artificial intelligence (AI) models. Artificial intelligence methods have the potential for providing powerful and accurate solutions to various applications, including arrhythmia detection. Artificial intelligence methods generally are difficult to implement on implantable medical devices due to their high computational demand (relative to an implantable medical devices resources), but have been found to be well-suited to cloud- based implementations in which computing resources are more abundant.

Within the instant disclosure, therefore, computations for identifying a cardiac arrhythmia are carried out by a computation system using a classification model according to the principles of artificial intelligence methods. By means of such classification model, for example a sensed cardiac signal in the shape of an ECG signal may be processed in order to confirm or reject the presence of arrhythmias in those signals as originally detected by the implantable medical device. Such artificial intelligence reassessment may help to reduce the number of false arrhythmia detections presented to medical personnel, thereby improving arrhythmia specificity of the overall medical device system and lessening the review burden for medical personnel.

In one embodiment, the computation system comprises a feature extraction module for extracting said multiplicity of features from the transmitted signal or a signal derived from the transmitted signal. Within the feature extraction module the multiplicity of features may for example be determined based on at least one statistical measure of different signal properties of the cardiac signal.

Generally, multiple features can be derived from a cardiac signal, such as the standard deviation of an RR interval, a mean change in the RR interval, a mean P-wave strength, the standard deviation of a QRS width or the like. RR intervals are the intervals between successive heartbeats. Such features are input into the classification model, and the classification model provides a classification indication as an output, the classification 5 indication indicating whether a cardiac arrhythmia is deemed to be present or not. The classification indication herein may have a Boolean value (true, false), or may be provided as a probability value indicating a likelihood that a true cardiac arrhythmia is present.

For example, the multiplicity of features may be determined based on at least one statistical measure of different signal properties of the cardiac signal. The signal properties may for example be selected from the group of an RR interval, a change in the RR interval, a relation of an RR interval to a prior RR interval, a signal strength such as a P-wave strength, a width of a signal waveform such as a QRS width, a computed power within a frequency range, and a spectral distribution. Generally, signal properties may be timing-based properties (such as the RR interval), morphological properties (such as the P-wave strength) or frequency domain properties (such as the spectral power within a frequency range). Features may be derived based on the signal properties using a statistical analysis, in particular by forming a mean value, a median value, a coefficient of variation, a percentile value, or a standard deviation value. For example, for multiple signal properties multiple statistical measures may be derived, such that for example from each of m signal properties n statistical measures may be derived, resulting in an m x /i feature matrix which may be fed into the classification model in order to derive the classification indication.

In addition to such features (which may be denoted as basic features and for example are based on statistical measures), advanced features may be derived and fed into the classification model. Such advanced features may for example include the presence of fibrillation waves and the presence of an ST segment depression, and hence relate to an overall characteristic within a sensed cardiac signal.

In one embodiment, the classification model comprises at least one decision tree using multiple features as inputs. For example, the classification model may comprise multiple decision trees to which the multiple features may be fed in order to derive said classification indication.

For example, at least one of the multiple decision trees defines at least one decision boundary to which at least one of the multiplicity of features is compared, wherein the output is 6 determined based on the comparison. For example, a decision tree may define a range which is bound by an upper boundary and by a lower boundary. By means of the decision tree, hence, it is determined whether a specific feature falls into the range as defined by the decision tree, or is outside of the range.

Herein, for different features different decision trees defining different ranges may be used, such that by means of the decision trees it is determined whether different features fall into different, associated ranges.

In another embodiment, at least one of the multiple decision trees may have a more complex shape, using one or multiple features as inputs in order to decide based on the one or the multiple features whether a cardiac arrhythmia is present.

By means of the decision trees it may be determined whether the different features indicate that a cardiac arrhythmia is present. In one embodiment, each feature is associated with a particular decision tree. As output from each decision tree herein it is decided whether, based on the associated feature, a cardiac arrhythmia (likely) is present or not.

In one embodiment, based on the outputs from the different decision trees the classification indication is derived, in particular by combining the outputs to come up with the classification indication indicating whether a cardiac arrhythmia is present or not. In particular, in one embodiment the computation system is configured to derive the classification indication for a cardiac arrhythmia by computing a score value based on the outputs of the multiple decision trees. The score value may in particular be determined according to where s represents the score value, which is computed by dividing the numbers of decision trees indicating, in their output, that a cardiac arrhythmia is present ("# of trees voting for arrhythmia") by the total number of decision trees ("# of total trees"). Additionally, the votes of each tree need not be weighted equally, as indicated in the equation above; each tree may 7 have a different weight contributing to the score value. In general, various methods of calculating a score value may use an ensemble of decision trees.

Based on the score value, the classification indication may be set. For example, the score value may be compared to a predefined threshold value, wherein it is found that a cardiac arrhythmia is present if the score value exceeds the threshold value, and it is found that no cardiac arrhythmia is present if the score value does not exceed the threshold value. The threshold value herein may be fixed, or the threshold value may be user programmable or user-adjustable, such that dependent on the setting of the threshold value the sensitivity of the classification model may be adapted as desired.

In one embodiment, the classification model, in an initial training phase prior to actual operation of the system, is trained using a set of training data. For example, a training set of ECG signals may be used which have been graded beforehand as indicating a cardiac arrhythmia or not, for example by medical experts. From the set of training data features may be computed, and the graded signals and the derived features may be fed into a training function in order to generate an ensemble of decision trees. Herein, each decision tree may for example generate decision boundaries for the supplied features that best separate the classes (indicating "a cardiac arrhythmia to be present" or "a cardiac arrhythmia to be not present"). Using then, in later operation, the decision trees thus defined, for example a score value may be computed, and based on the score value, in particular based on a comparison of the score value to a threshold value, it may be determined whether a cardiac arrhythmia is present or not. Using a decision tree ensemble as the classification model according to known machine learning approaches allows to achieve a balanced trade-off between interpretability, complexity, and performance. Because the classification model is implemented on a computation system, for example a cloud-based computation system, connected (directly or indirectly) to the implantable medical device by means of a communication network such as the Internet, computationally intensive, high-performance artificial intelligence methods may be used which may identify and discard false arrhythmia detections presented to medical personnel and hence may allow to reduce a review burden. 8

In one embodiment, the implantable medical device is configured to identify, based on a sensed cardiac signal, a signal portion exhibiting a cardiac arrhythmia and to transmit such signal portion as the transmit signal. In this embodiment, the implantable medical device performs a preprocessing, in which an initial determination whether a cardiac arrhythmia is present or not is carried out. Based on the preprocessing of the implantable medical device, a signal portion of the sensed cardiac signal is transmitted towards the computation system, which re-assesses the signal portion and confirms or discards the cardiac arrhythmia.

In another embodiment, no substantial preprocessing is done at the implantable medical device, in particular no determination whether a cardiac arrhythmia is likely present or not. In this case, hence, the computation system solely assesses the cardiac signal in order to determine the presence of a cardiac arrhythmia.

In another aspect, a method for assessing a cardiac arrhythmia comprises: sensing, using a sensing device of an implantable medical device, a cardiac signal; transmitting, using a communication device of the implantable medical device, a transmit signal comprising information relating to said cardiac signal; obtaining, using a computation system, a multiplicity of features based on said transmit signal; and deriving, using the computation system, a classification indication for a cardiac arrhythmia based on said multiplicity of features, wherein said multiplicity of features are input into a classification model for deriving said classification indication.

The advantages and advantageous embodiments described above for the system equally apply also to the method, such that it shall be referred to the above in this respect.

By means of the system and method, a cardiac arrhythmia may be identified or discarded as a false arrhythmia detection. Herein, various types of arrhythmias may be assessed, including for example bradycardia, tachycardia, asystole, and atrial fibrillation, and also pathological defects may be identified, including for example a long QT syndrome, and ST elevation, or an AV block. - 9 -

The various features and advantages of the present invention may be more readily understood with reference to the following detailed description and the embodiments shown in the drawings. Herein, Fig. 1 shows a view of an implantable medical device implanted in a patient and communicating with a monitoring device external to the patient;

Fig. 2 shows a schematic view of the implantable medical device; Fig. 3 shows a schematic view of the implantable medical device in communication with a cloud-based server for assessing a cardiac arrhythmia;

Fig. 4 shows an example of a cardiac signal and a feature derived from said cardiac signal;

Fig. 5 shows another example of a cardiac signal and a feature derived from said cardiac signal; Fig. 6 shows a schematic view of a feature matrix derived from signal properties by means of a statistical analysis;

Fig. 7 shows a distribution of training set data in dependence on two features as derived from training set cardiac signals;

Fig. 8 shows a view of an example of a decision tree of a classification model;

Fig. 9 shows a range as defined by the decision tree of Fig. 8 within the distribution of training set data according to Fig. 7; and 10

Fig. 10 shows a schematic view of a combination of multiple decision trees in order to derive a score value for determining a classification indication for a cardiac arrhythmia.

Subsequently, embodiments of the invention shall be described in detail with reference to the drawings. In the drawings, like reference numerals designate like structural elements.

It is to be noted that the embodiments are not limiting for the invention, but merely represent illustrative examples.

Fig. 1 shows a schematic drawing of an implantable medical device 1, for example in the shape of a pacemaker device, a defibrillator device, or a sensor device, which is implanted in a patient, for example in a patient's heart H. For example, the implantable medical device 1 may be a leadless pacemaker which is implanted in the right ventricle RV of the patient's heart H. In another embodiment, the implantable medical device 1 may for example be a stimulation device such as a pacemaker device or a defibrillator device which is subcutaneously implanted in a patient and comprises one or multiple leads extending in at least one of the right ventricle RV, the right atrium RA, the left ventricle LV, or the left atrium LA.

It shall be noted herein that the implantable medical device 1 may be any implantable medical device 1 which is functional to sense a signal relating to a cardiac activity, such as an electrocardiogram (ECG) signal or another signal, such as a pressure signal, a flow signal or the like, relating to a cardiac function.

As illustrated in Fig. 1, the implantable medical device 1 is in communication connection with an external device 2, for example a monitoring device, such as a mobile device which regularly is in proximity to the patient and hence may communicate with the implantable medical device 1 using a close-range communication technology, for example according to the MICS protocol, the Zigbee protocol, or the Bluetooth Low Energy (BLE) protocol. 11

The external device 2 generally serves as a link to an e.g. cloud-based computation system 3, which is in communication connection with the external device 2 for example via the Internet. The computation system 3, for example composed of an arrangement of one or multiple servers at a healthcare facility or at a remote service center, may process data received from the implantable medical device 1, may send data towards the implantable medical device 1, may provide information to medical personnel for review, and may allow for medical personnel to program a configuration in relation to the implantable medical device 1. Referring now to Fig. 2, in one embodiment the implantable medical device 1 may be a leadless device having a housing 10 and components enclosed in the housing 10. The implantable medical device 1 for example comprises a sensing device 11, for example in the shape of an arrangement of one or multiple electrodes arranged on the housing 10 at one or multiple different locations, for sensing a cardiac signal relating to a cardiac function, such as an electrocardiogram (ECG) signal. The implantable medical device 1 further comprises a processing device 12, an energy storage device 13 in particular in the shape of a battery, and a communication circuitry 14 for communicating with an external device 2, in particular a monitoring device for example in the context of a home monitoring, according to a standardized communication technology. The communication circuitry 14 can also be called communication device 14.

Referring now to Fig. 3, the implantable medical device 1 in particular shall be configured to sense signals in relation to a cardiac arrhythmia. In cooperation with the external device 2 and the external computation system 3, herein, a cardiac arrhythmia shall reliably be detected, including a bradycardia, a tachycardia, an asystole, an atrial fibrillation, a long QT syndrome, an ST elevation, or an AV block.

In one embodiment, the implantable medical device 1 performs a preprocessing, in which a (potential) cardiac arrhythmia is identified in a sensed cardiac signal, for example an ECG signal (step SI in Fig. 3). Following the identification of a cardiac arrhythmia by the implantable medical device 1, the implantable medical device 1 transmits information relating to the detected cardiac arrhythmia to the external device 2 and, via the external 12 device 2, to the computation system 3. The information may for example include an indication that a cardiac arrhythmia has been detected and a signal portion, such as a portion of an ECG signal in a certain time range, that relates to the detected cardiac arrhythmia (step S2 in Fig. 3).

The transmit signal received from the implantable medical device 1 by the external device 2 may be further processed within the external device 2 for forwarding to the computation system 3. Upon reception of the information at the computation system 3, the computation system 3 performs an algorithm to reassess the detected arrhythmia (step S3) making use of an artificial intelligence model, in particular a classification model, serving to identify the detected arrhythmia as true (step S4) or false (step S5) and providing a corresponding classification indication, e.g. in the shape of a label-like Boolean indication (True/False).

It shall be noted that the setup as depicted in Fig. 3 is only one example of a possible setup For example, it also is possible that the implantable medical device 1 does not perform a particular pre-processing in order to identify a cardiac arrhythmia, but merely transmits a cardiac sense signal towards the computation system 3, which assesses the signal in order to identify a cardiac arrhythmia. In this case, hence, the computation system 3 does not serve to reassess in order to confirm or discard a detected cardiac arrhythmia, but to identify a cardiac arrhythmia without any particular prior indication of a cardiac arrhythmia from the implantable medical device 1.

In the example of Fig. 3, in order for the computation system 3 to assess a cardiac arrhythmia detection as being true or false, it must be able to classify the cardiac signal (e.g. an ECG signal) as having cardiac arrhythmia or not. Therefore, it is proposed to use an independent arrhythmia classifier that receives the implantable medical device detection as input, and provides a “true cardiac arrhythmia” or “false cardiac arrhythmia” indication as the output. Classification herein may be carried out making use of artificial intelligence technologies, specifically machine learning technologies.

Prior to actual operation of a system for detecting a cardiac arrhythmia, a classification model needs to be identified and trained, such that in actual operation features may be fed - 13 - into the classification model as inputs in order to classify a sensed signal as exhibiting a cardiac arrhythmia or not.

As a first step in developing the classification model, appropriate signal features need to be identified that can be used to discriminate among the different classes (i.e., for classifying the processed signal as a “True cardiac arrhythmia” or “False cardiac arrhythmia” signal). Specifically, in the case of reassessing cardiac arrhythmia detections from an implantable medical device 1, features in the cardiac signal need to be identified that can be used by the model to classify the signal as having cardiac arrhythmia or not.

As a specific example, one feature might be the standard deviation of RR intervals in the signal. Figs. 4 and 5 demonstrate that for a “no cardiac arrhythmia” signal C (in this case a sinus rhythm, Fig. 4), the distribution of RR intervals (representing a signal property P) has little spread, and for that particular example a standard deviation of 53 milliseconds (the standard deviation representing a feature M derived as a statistical measure from the signal property P). On the other hand, for a “cardiac arrhythmia” signal (Fig. 5), in this particular example atrial fibrillation, the distribution of RR intervals has a wider spread and a resulting substantially larger standard deviation of 142 milliseconds in the shown example. This large difference indicates that the standard deviation of RR intervals may provide a suitable feature for a model to discriminate between “cardiac arrhythmia” and “no cardiac arrhythmia” classes.

In reality, and provided there is enough data for the classification model to learn from, many features can be incorporated into the model. In general, more useful features that are incorporated into the model leads to improved classification performance.

The classification model may primarily use statistical measures of various signal properties as basic features. In the example above, the RR intervals are a signal property P, while the standard deviation of RR intervals represents a statistical measure M of that signal property.

Examples of signal properties may include different timing-based properties (ARR, RR _n /RR _n -i, etc.), morphology properties (P-Wave strength, QRS width, etc.), and frequency- 14 domain properties (power within a range of frequencies, distribution of the spectrum, etc.). Examples of statistical measures include mean, median, coefficient of variation, percentiles, range, etc. In general, the basic feature scheme for the proposed classification model is shown in Fig. 6, wherein from each of m signal properties PI... Pm n statistical measures may be derived, such that a feature matrix of size m x n arises comprising features in the shape of statistical measures Mll...Mmn. Thus, each statistical measures Mll...Mmn can be a feature M, and each of signal properties PI ...Pm can be a signal property P.

In addition to the basic features illustrated in Fig. 6, advanced features may be used in the classification model. The advanced features may not follow the scheme of “a statistical measure of a signal property”, but instead represent additional summary measures of the entire signal. For the case of cardiac arrhythmia classification, examples of advanced features may include: “presence of fibrillation waves”, “presence of ST segment depression”, etc. The classification model hence, in one embodiment, includes both basic and advanced features.

Within the classification model, a selection of basic and advanced features used as inputs to reassess arrhythmia detections may vary depending on the specific arrhythmia targeted. For example, mean p-wave strength may be an important feature for atrial fibrillation (AF) classification, but not for bradycardia classification. Therefore the specific basic and advanced features used may change based on the type of arrhythmia being assessed. In other words, different classification models utilizing different features may be implemented for different types of arrhythmia (bradycardia, tachycardia, asystole, atrial fibrillation) or pathology (long QT syndrome, ST elevation, AV block).

Once appropriate signal features have been identified the next step is to train a model to use these features to classify a signal as exhibiting “cardiac arrhythmia” or “no cardiac arrhythmia”. This takes place in a training phase prior to actual operation of the system. For example, in one embodiment a decision tree ensemble may be used as the classification model. Decision tree ensembles are a classic machine learning approach that offers a balanced trade-off between interpretability, complexity, and performance. The classification - 15 - model using the decision tree ensemble is trained in the initial training phase making use of a set of training data, according to which the decision tree ensemble is defined in order to allow to classify, during later operation, a signal by feeding features derived from the signal into the model.

For example, a training set of cardiac (e.g. ECG) signals that have been graded as “cardiac arrhythmia” or “no cardiac arrhythmia” by experts is used to generate an ensemble of trees with different decision boundaries based on the supplied features. As a simple example, for each signal in the training set two features may be calculated: xl = RR Median, and x2 = RR _n /RR _n -i 25th Percentile. The space of these two features in a training set is shown in Fig. 7. Note that there is some clustering of “cardiac arrhythmia” and “no cardiac arrhythmia” signals in the training set using just these two features. “Cardiac arrhythmia” can be abbreviated as "AF", and “no cardiac arrhythmia” can be abbreviated as "Not AF".

Next, the graded class (“cardiac arrhythmia” or “no cardiac arrhythmia”) of all signals and their derived features are fed into a training function that generates an ensemble of decision trees. Each decision tree for example generates decision boundaries for the supplied features that attempt to best separate the classes.

This is depicted in Fig. 8, which shows a simple decision tree T taking features derived using a feature extraction module F from a sensed signal C as input and comparing the features to boundary values of boundaries Bl, B2 (having values of 0.7 and 0.9, respectively, in the shown example) at nodes Nl, N2. Fig. 9 shows the decision boundaries Bl, B2 in feature- space corresponding to the decision tree of Fig. 8. The decision tree T votes on whether a particular signal is “cardiac arrhythmia” or “no cardiac arrhythmia” based on the signal’s features and the decision boundaries Bl, B2, and consequently these decision boundaries Bl, B2 split the signals into decision regions of “Vote cardiac arrhythmia” (range A in between lower boundary Bl and the upper boundary B2) and “Vote no cardiac arrhythmia” (elsewhere). The signal C, in this and in other embodiments, can also be called cardiac signal C. - 16 -

Referring now to Fig. 10, when the training function generates an ensemble of decision trees T1-T3, each tree T1-T3 will vote for “cardiac arrhythmia” or “no cardiac arrhythmia”, as it is illustrated in Fig. 10 on the left. The outputs of the different decision trees T1-T3 herein may be combined by deriving a cardiac arrhythmia score according to, for example

S represents the score value, which is computed by dividing the number of decision trees indicating, in their output, that a cardiac arrhythmia is present ("# of trees voting arrhythmia") by the total number of decision trees ("# of total trees"). By combining the outputs of all trees T1-T3 in the ensemble, a map of the score value is created based on the combined decision regions of each tree, as shown in Fig. 10 on the right. For example, on the left are the decision boundaries and “Vote cardiac arrhythmia”/“Vote no cardiac arrhythmia” regions of the three unique trees T1-T3. When the decision regions of all three trees T1-T3 are combined, a cardiac arrhythmia score map is generated based on the number of trees voting for “cardiac arrhythmia” or “no cardiac arrhythmia” in a particular region of feature-space, indicated by the overlapping decision regions of all trees, as shown in Fig. 10 on the right. Note that the score values can be different in each overlapping area, and in this particular example there were three overlapping regions Dl, D2 and D3 in which a majority of trees voted for “cardiac arrhythmia”.

As a result, Fig. 10 shows an example of a trained model that outputs a cardiac arrhythmia score for any input signal given its features. The trained model can now be used to classify a signal as “cardiac arrhythmia” or “no cardiac arrhythmia” by setting a user-defined score threshold. Any signal that has a cardiac arrhythmia score higher than this threshold will be classified as “cardiac arrhythmia”, and will be classified as “no cardiac arrhythmia” otherwise.

For example, to be highly-sensitive to cardiac arrhythmia, a user may decide that a signal should be classified as “cardiac arrhythmia” if at least one tree makes a “Vote cardiac arrhythmia” decision for the signal. This corresponds to a score threshold of 1/3 in the above example with three trees T1-T3. - 17 -

In another scenario, if a user decides to be more specific to cardiac arrhythmia, the threshold may be set such that at least two trees T1-T3 make a “Vote cardiac arrhythmia” decision in order to classify the signal as “cardiac arrhythmia”. In this case the score threshold would be 2/3 in the above example.

Consequently, the user can adjust the sensitivity and specificity of the model by adjusting the score threshold as desired. Note that in the above example, each tree T1-T3 is considered to have equal weight. In practice, however, each tree may be weighted differently according to how well their decision boundaries split the “cardiac arrhythmia” signals from the “no cardiac arrhythmia” signals. In general, various methods of calculating a score value may use an ensemble of decision trees.

The above example uses two features and three trees T1-T3, but in practice many, e.g. hundreds, of features and trees may be used for classification.

Furthermore, while in the specific example above a decision tree ensemble was used as the classification model, the algorithm may incorporate any type of classification model that can be trained on user-selected features (e.g. k-nearest neighbor, support vector machine, naive Bayes, neural networks, etc.).

Once the basic and advanced features to classify cardiac arrhythmia are determined, and a classification model is trained on a set of graded signals for which those features have been extracted, the solution can be implemented on the computation system 3, e.g. a cloud server, which then, during actual operation, receives cardiac arrhythmia detections from an implantable medical device 1. On the computation system 3, upon receipt of a cardiac arrhythmia detection from an implantable medical device 1, the computation system 3 extracts the basic and advanced features from the ECG signal associated with the detection using the feature extraction - 18 - module F as shown in Fig. 8. The features are subsequently used as input to the trained model (stored on the computation system 3), which produces a cardiac arrhythmia score as output. If the cardiac arrhythmia score is above a user-defined threshold, the model confirms the cardiac arrhythmia detection and e.g. presents it to medical personnel.

However if the cardiac arrhythmia score is below the user-defined threshold, the model rejects the cardiac arrhythmia detection and can do one of the following to filter the false detection: 1) Flag the detection as false cardiac arrhythmia and store for optional review. The optional review method can be implemented in several ways, including (but not limited to): i. Placing the detection on a separate “False cardiac arrhythmia” page on the computation system 3; ii. Placing the detection on the same page as other (true) detections but labeling it as “False” and visually highlighting, underemphasizing, or minimizing it.

2) Discard the false cardiac arrhythmia detection entirely, making medical personnel unable to review it. For the first option, the review burden is reduced by de-prioritizing review of implantable medical device detections that the classification model has deemed to be false. For the second option, the review burden is reduced by completely removing detections deemed to be false without allowing medical personnel to review them. Furthermore, the user-defined score threshold can be set by the user (i.e. the medical personnel) on a patient-by-patient basis. This allows the user to tune the sensitivity/specificity of the classification model for each implantable medical device 1 / patient. For example, the user may set a low score threshold for an implantable medical device 1 / patient that rarely transmits cardiac arrhythmia detections in order to see any cardiac arrhythmia detection that may occur (true or false). On the other hand, the user may set a high score threshold for an implantable medical device 1 / patient that frequently transmits false cardiac arrhythmia detections in order to reduce review burden and focus on 19 detections that have a higher likelihood of being true (as determined by the classification model).

- 20

List of reference numerals

1 Implantable medical device

10 Housing

11 Sensing device

12 Processing device

13 Energy storage device

14 Communication device 2 External device 3 Computation system A Range

Bl, B2 Boundary C Cardiac signal

D1-D3 Overlapping area

F Feature extraction module

M, Ml l...Mmn Feature

P, PI...Pm Signal property

H Heart

LA Left atrium

LV Left ventricle

N1, N2 Node

RA Right atrium

RV Right ventricle s Score value

T _{T 1 T3} Decision tree