Method for analysing a quasi-periodic biological signal

Abstract

The invention relates to a method for analysing a quasi-periodic biological signal which comprises measured values for a measurement interval. According to the invention, the method comprises the following steps: (a) determining from the quasi-periodic biological signal a first observation window, which has a first period, and a second observation window, which has a second period, the first observation window and the second observation window being within the measurement interval and the second period being shorter than the first period; (b) classifying the measured values using the first observation window by a first model to obtain a first classification result and classifying the measured values using the second observation window by a second model to obtain a second classification result, the first model and the second model each being based on an artificial neural network; and (c) interpreting the first classification result and the second classification result by determining the relevance of one or more of the measured values to generate the first classification result and the relevance of one or more of the measured values to generate the second classification result.

Description

Description Method for analyzing a quasi-periodic biological signal

[0001] The invention relates to a method for analyzing a quasi-periodic biological signal. It particularly relates to a method for analyzing an electrocardiogram (ECG).

[0002] The automated detection of characteristic phenomena in biological signals and their interpretable traceability is of utmost clinical relevance. Due to their architecture and limited interpretability, existing methods for classifying phenomena in biological signals are neither understandable nor comprehensible, and thus cannot be transferred to clinical monitoring.

[0003] The current standard in clinical practice is still the manual evaluation of rhythm strips and additional calculated parameters. The physician then makes a decision based on their expertise and experience. Depending on the physician's experience, this can lead to a high number of false positives. In recent years, research has increasingly presented powerful neural networks (classifiers) for various pathologies [1]. These networks can be created from features based on expert knowledge (machine learning) or self-learned features (deep learning).For both approaches, there are methods to understand the classification decision [2], [3]. However, especially in the area of ​​deep learning, which does not require expert knowledge, there are limitations in the area of ​​interpretability and traceability in the analysis of biological signals [4], which is why the classification decision, but not the basis of the decision, could be integrated into clinical monitoring.

[0004] Classifiers based on expert knowledge (machine learning) are easy to interpret, but they require expert knowledge. This applies to both the The creation of such classifiers and the interpretation of the decisions obtained using these classifiers are challenging. Furthermore, they exhibit low convergence and are not customizable. This means that these classifiers can only be trained on problems for which the expertise to solve them is available. Approaches without expertise that define features to characterize different classes (deep learning) can learn features themselves using large databases, but these models are not interpretable by default, or the basis of the decision is not comprehensible. Methods for understanding deep learning-based decisions that serve to interpret decision-making are limited and little researched to date.However, this is precisely what is necessary for (1) the comprehensibility of a decision for the clinician and (2) the investigation of new phenomena where the characteristics are not yet present in the expert knowledge and are therefore of great value for the understanding of the diagnosis or therapy.

[0005] In [5], the influence of different signal lengths on the accuracy of ECG classification is investigated. To interpret the decision, relevant features in the electrocardiogram (ECG) are inferred using interpretability methods and visualized. The authors apply the decomposition of the signals into viewing windows of different lengths solely with the goal of performance optimization while taking interpretability into account. However, in [5], no parallel structure is implemented for combining models with different viewing windows, such as long-term and short-term. Furthermore, different features, such as rhythm and morphology, are not inferred based on short-term or long-term behavior. Finally, there is no visualization based on either of these features.

[0006] In [6], a qualitative representation of the interpretation of the output signal of an ECG was presented. However, this does not apply a combined representation that would allow conclusions to be drawn about features of rhythm and morphology. Therefore, it is unclear to what extent both sets of features can be prepared for the clinician and subsequently interpreted by him. Likewise, no parallel analysis is presented here. Structure used to combine models with different viewing windows.

[0007] To date, no system is known that has integrated an interpretable display based on self-learning network architectures into clinical monitoring.

[0008] The object of the invention is to eliminate the disadvantages of the prior art. In particular, it is intended to provide a method for analyzing a quasi-periodic biological signal that enables the automated detection of characteristic phenomena of a quasi-periodic biological signal. Furthermore, the automated detection should be interpretable and traceable.

[0009] This object is solved by the features of claim 1. Expedient embodiments of the inventions emerge from the features of the subclaims.

[0010] According to the invention, a method is provided for analyzing a quasi-periodic biological signal comprising measured values ​​for a measurement period. The method comprises the steps: (a) determining a first viewing window having a first time duration and a second viewing window having a second time duration from the quasi-periodic biological signal, wherein the first viewing window and the second viewing window lie within the measurement period and the second time duration is shorter than the first time duration; (b) Classification of the measured values ​​using the first viewing window by a first model to obtain a first classification result and classification of the measured values ​​using the second viewing window by a second model to obtain a second classification result. nises, wherein the first model and the second model are each based on an artificial neural network; and (c) interpreting the first classification result and the second classification result by determining the relevance of one or more of the measured values ​​to generate the first classification result and the relevance of one or more of the measured values ​​to generate the second classification result.

[0011] The method according to the invention can be an automated method. It uses a self-learning network architecture to analyze the measured values ​​of the quasi-periodic biological signal. The self-learning network architecture is the artificial neural network mentioned in step (b), which is used for the first model and the second model. The first model is referred to as the long-term model with regard to the first observation window, which is longer in time than the second observation window. The second model is referred to as the short-term model with regard to the second observation window, which is shorter in time than the first observation window. The self-learning network architecture is an interpretable self-learning network architecture.It can be used, for example, for rhythm and morphology analysis of measured values ​​in clinical monitoring, with the long-term model recognizing rhythm characteristics and the short-term model recognizing morphology characteristics. The method according to the invention can be carried out using a data processing device, for example, a computer. A measuring device can be provided to record the quasi-periodic biological signal. The measuring device can be, for example, the electrodes of an electrocardiograph.

[0012] Step (a) of the method according to the invention enables preprocessing of the measured values. With regard to this preprocessing, the unprocessed quasi-periodic biological signal is also referred to as the raw signal. This raw signal consists of the original measured values ​​of the quasi-periodic biological signal as recorded by the measuring device. Therefore, the measured values collectively referred to as the raw signal. Step (b) involves applying the long-term model and the short-term model to enable classification of the preprocessed measured values. The long-term model and the short-term model can be applied in a parallel structure. Furthermore, to improve the classification result, the classification results of the long-term model and the short-term model from step (b) can be combined. Step (c) involves interpreting the results of step (b), i.e., the first classification result and the second classification result.

[0013] It can be provided that the method according to the invention further comprises the step: (d) combining the relevance of one or more of the measured values ​​for generating the first classification result and the relevance of one or more of the measured values ​​for generating the second classification result by representation in a two-dimensional plane. Step (d) is therefore aimed at presenting the results of step (c), namely the presentation of relevance.

[0014] The term "quasiperiodic biological signal" refers to a signal that is almost, but not exactly, periodic. A quasiperiodic biological signal can be viewed as a time series. An example of a quasiperiodic biological signal is an electrocardiogram (ECG). A quasiperiodic biological signal can be viewed as a time series, with measured values ​​being obtained within a measurement period in which the biological signal is or was recorded.

[0015] The first viewing window and the second viewing window can each be sliding viewing windows. Both viewing windows have a constant duration, but with a sliding viewing window, the temporal start of the viewing window shifts continuously. The second viewing window can lie within the first observation window. This means that the time range encompassed by the second observation window lies within the time range encompassed by the first observation window. For long-term and short-term observation, the first observation window preferably covers a time period of several quasi-periods, and the second observation window a maximum of one quasi-period of the quasi-periodic biological signal.

[0016] Step (a) of the method according to the invention enables preprocessing of the measured values ​​of the quasi-periodic biological signal. For this purpose, step (a) of the method according to the invention can comprise the following substeps: (ai) filtering the measured values ​​of the quasi-periodic biological signal using a filter to remove noise components from the measured values ​​and obtain filtered signals; (a2) windowing the filtered signal to obtain windowed signals; and (as) Padding the windowed signals while preserving padded signals.

[0017] The padded signals obtained in step (as) can then be used in step (b). Step (ai) involves the use of a filter. This filter can be a high-pass filter, for example. The filter is intended to eliminate noise components, such as low-frequency signal components. Step (a2) involves windowing the signal from step (ai), for example, using a Tukey window. This is intended to prevent edge effects in the following step (as). Step (as) is intended to ensure that all signal times are perceived equally often through the sliding viewing windows during step (b).

[0018] The artificial neural networks used in step (b) can each be based on a convolutional neural network. For example, an ID Convolutional Neural Network classifier (ID CNN) can be used. The convolutional neural network can each consist of several convolutional layers, where At least one feature map is obtained in each convolutional layer. It can be provided that in the temporally last convolutional layer, the mean of each feature map belonging to this convolutional layer is calculated. The mean of each feature map can be fed to a softmax function to obtain the first classification result and the second classification result.

[0019] In step (b), two observation windows are provided, namely the first observation window and the second observation window, which are used to classify the measured values. The shorter observation window, i.e., the second observation window, is not only observed once, but continuously in the short-term model.

[0020] Step (c) provides for the interpretation of the first classification result and the second classification result by determining the relevance of one or more of the measured values ​​for generating the first classification result and the relevance of one or more of the measured values ​​for generating the second classification result. For this purpose, it can be provided that the relevance of one or more of the measured values ​​for generating the first classification result and the relevance of one or more of the measured values ​​for generating the second classification result are each determined using relevance propagation rules. Preferably, the relevance of several of the measured values ​​or of all measured values ​​is determined, with the relevance of all measured values ​​being particularly preferably determined.If the relevance of all measured values ​​is not determined, the measured values ​​whose relevance is determined can be selected arbitrarily, based on prior knowledge, or based on technical conditions. The selection of measured values ​​can be automated. The relevance of the measured values ​​is determined by calculation.

[0021] Step (d) involves combining the relevance of one or more of the measured values ​​for generating the first classification result and the relevance of one or more of the measured values ​​for generating the second classification result by displaying them in a two-dimensional plane. In this way, a combined relevance is obtained. It may be provided that the combined relevance is the quasi-periodic biological signal is projected. The combined relevance can be visualized together with the quasi-periodic biological signal onto which it is projected. The visualization can be done, for example, on a digital monitor or a rhythm strip.

[0022] It can be provided that features of the quasi-periodic biological signal are determined by means of the analysis provided according to the invention. In one embodiment of the method according to the invention, the features are features of the rhythm and / or morphology of the quasi-periodic biological signal.

[0023] The method according to the invention is based on an interpretable, self-learning neural network architecture for detecting phenomena in quasi-periodic biological signals, such as the electrocardiogram (ECG). To this end, the method according to the invention relies on a combined long-term and short-term observation, which results in rhythm characteristics being learned using the long-term model and morphology characteristics being learned using the short-term model. These learned characteristics can be used for a precise, interpretable classification of biological signals or segments of biological signals, but can also be represented in the biological signal itself. This can provide clinicians with better interpretability in diagnostic and therapeutic decisions.

[0024] The method according to the invention can learn features from large data sets, a process also referred to as deep learning. The learning of features can be combined with a methodology for interpreting the features. The architecture of the underlying neural network is specifically designed for the analysis of quasi-periodic biological signals. A viewing window for a longer signal segment, referred to as the first viewing window, which enables long-term observation, and a short signal segment, referred to as the second viewing window, which enables short-term observation, are evaluated. This aims to identify features across the quasi-periodic, i.e., the long-term, and the form of the quasiperiod, i.e., the short-term. A combined decision is made from the learned long-term and short-term features. The combined decision can be determined by weighting the decision certainty of the classification results from the long-term and short-term observations. The interpretability of the combined decision allows the representation of the knowledge base directly in the observed signal and thus in clinical monitoring. This representation is comprehensible for clinicians, since long-term features correspond to the rhythm in the clinic, and short-term features to the morphology. Both feature types are highly relevant in diagnostic and therapeutic decisions. A non-separated consideration, as previously, led to a mixing of the feature types and thus to the incomprehensibility of the decision basis.Long-term features are the features of rhythm, short-term features are the features of morphology.

[0025] The method according to the invention is characterized by the use of a self-learning neural network that considers the long-term and short-term behavior of a quasi-periodic biological signal. It can also make a combined, comprehensible decision. The method according to the invention also enables the representation of the combined decision in clinical monitoring to support diagnosis and therapy.

[0026] The key advantages of the method according to the invention are highly efficient decision-making compared to the state of the art. The decision-making process does not require expert knowledge, can be learned independently, and is interpretable by the clinician. The method according to the invention can also be used for phenomena that cannot yet be explained. Furthermore, the method according to the invention can be integrated into existing clinical monitoring, offering added value to clinical routine. Furthermore, the method according to the invention is transferable to monitoring systems in non-clinical settings and offers a high degree of convergence.

[0027] The invention will be explained in more detail below with reference to the drawings and exemplary embodiments which are not intended to limit the invention. Fig. 1 a visualization of an interpretable network architecture for the rhythm and morphology analysis of a quasi-periodic biological signal in clinical monitoring; Fig. 2 schematic representations of a preprocessing of ECG raw signals showing a sinus rhythm (a) and atrial fibrillation (b); Fig. 3 shows the description of an architecture for a long-term model, wherein the selected viewing window is 10 seconds and the input for the long-term model is the padded ECG signal 204 obtained in step 1 (c); Fig. 4 shows the description of an architecture for a short-term model, wherein the selected viewing window is 0.6 seconds and the input for the short-term model is the padded ECG signal 204 obtained in step 1 (c); Fig. 5 shows an exemplary combined representation of the relevance of the long-term model and the short-term model for integration into a clinical ECG monitor; Fig. 6 shows an exemplary combined representation of the relevance of the long-term model and the short-term model in a clinical ECG rhythm strip; Fig. 7 a confusion matrix of a long-term model for the detection of atrial fibrillation (AF); Fig. 8 shows a confusion matrix of a short-term model for the detection of atrial fibrillation (AF); Fig. 9 a confusion matrix of a combined model for the detection of atrial fibrillation (AF); Fig. 10 shows an exemplary sinus rhythm ECG with feature display; and Fig. 11 an example atrial fibrillation ECG with feature display.

[0028] Fig. 1 is a visualization of the interpretable network architecture for rhythm and morphology analysis of a quasi-periodic biological signal in clinical monitoring. First, a quasi-periodic biological signal is acquired (1). This could, for example, be an ECG, which is illustrated in Fig. 1 on an ECG monitor. The acquired quasi-periodic biological signal is the raw signal, consisting of measured values ​​acquired during a measurement period.

[0029] The acquired raw signal is viewed in parallel over both long-term 2 and short-term 3. The long-term view is intended to capture features of the rhythm of the quasi-periodic biological signal, while the short-term view is intended to capture features of the morphology of the quasi-periodic biological signal. Long-term 2 is performed with a sliding viewing window of, for example, 10 s in length. This viewing window is the first viewing window. Short-term 3 is performed with a sliding viewing window of, for example, 0.6 s in length (see arrow C). This viewing window is the second viewing window.

[0030] The signals processed in the long-term view and the signals processed in the short-term view are now used for feature extraction 4.

[0031] The signals processed in the long-term view are analyzed with a long-term model 41 (step (b)), which provides a first classification The signals processed in the long-term observation can be signals obtained in step (a), for example, padded signals from step (a^). These are the measured values ​​after the preprocessing step. The long-term model has a model architecture described later. The long-term model enables the extraction of features, namely features of the rhythm of the quasi-periodic biological signal.

[0032] The signals processed in the short-term view are analyzed with a short-term model 42 (step (b)), yielding a second classification result. The signals processed in the short-term view can be signals obtained in step (a), for example, padded signals from step (aa). These are the measured values ​​after preprocessing. The short-term model has a model architecture described later. The short-term model enables the extraction of features, namely features of the morphology of the quasi-periodic biological signal.

[0033] Feature processing 5 now takes place. For this purpose, a combination 51 of the first classification result and the second classification result is carried out (step (c)). The interpretation method 52 determines the relevance of individual contents of the raw signal, and this content can be displayed directly there, i.e., in the raw signal. Alternatively or additionally, this display can also take place in the signal after the preprocessing step (ai) in order to present the clinician with a noise-free signal. This is shown in the box "Feature Display" 6. A combined display can be carried out in the sense of short-term or long-term viewing. This corresponds to an interpretable display of the relevance of rhythm features and morphology features for the classification. The direct feature display can, for example, take place directly on the ECG monitor 61. It can also take place on a rhythm strip 62.The feature representation can be used for continuous adaptation 7 of the feature extraction.

[0034] Feature extraction 4 and feature processing 5 are based on the architecture of a self-learning neural network, which considers the long-term and short-term behavior of a quasi-periodic biological signal and makes a combined and comprehensible decision (Box A in Fig. 1). The feature representation (Box B in Fig. 1) illustrates the representation of the combined decision in clinical monitoring.

[0035] An embodiment of the method according to the invention is described below. In a first step, corresponding to step (a), the raw signal is preprocessed. This preprocessing includes preparing the measured values ​​of the raw signal for the interpretable, self-learning network architecture for rhythm and morphology analysis in clinical monitoring. In a second step, corresponding to step (b), the long-term model and the short-term model are applied. In a third step, corresponding to step (c), the classification results obtained in step (b) are interpreted. In a fourth step, corresponding to step (d), the features obtained in step (c) are displayed. Step (a): Preprocessing

[0036] The preprocessing steps are described chronologically below. These are illustrated in Fig. 2 as examples for two raw ECG signals. The first raw ECG signal 101, shown under (a) in Fig. 2, shows sinus rhythm, while the second raw ECG signal 201, shown under (b) in Fig. 2, shows atrial fibrillation. The preprocessing comprises the following steps: (ai) filtering the respective raw ECG signal to obtain a filtered signal, (a2) windowing the filtered signal to obtain a windowed signal, and (as) padding the windowed signal.

[0037] The first raw ECG signal 101 is a non-pathological sinus rhythm ECG, while the second raw ECG signal 201 is an ECG with atrial fibrillation. Figure 2 shows the chronological sequence (from top to bottom): the raw signal 101, 201, the signal after filtering 102, 202, the filtered signal 103, 203 after windowing, the windowed signal after padding 104, 204. Then steps (b), (c) and (d) are applied.

[0038] Steps (a2) and (a5) make it possible to avoid edge effects and features of technical origin. Thus, the generalizability of the method according to the invention is ensured. (ai) Filtering

[0039] The raw ECG signals 101, 201 are each filtered to obtain filtered signals 102, 202. To eliminate low-frequency signal components, the raw ECG signals 101, 201 are filtered using a high-pass filter with a cutoff frequency of 0.3 Hz. This filter is a 4th-order Butterworth filter. The signals are then transformed into filter banks using an 8-level discrete wavelet transform. These filter banks were cleaned of high-frequency noise using soft thresholding. Finally, the signals are reconstructed using an inverse discrete wavelet transform (DWT) based on the cleaned filter banks [7]. This produces a noise-free signal familiar to the clinician, which is the filtered signal 102, 202. (a ) Fenestration

[0040] The filtered signals 102, 202 obtained in step (ai) are now windowed, resulting in windowed signals 103, 203. Windowing is intended to prevent edge effects in the subsequent signal padding step. For this purpose, the signals are multiplied by a Tukey window with alpha = 0.06. This reduces the signal amplitude toward the edge to the zero line. (a ) Padding

[0041] The windowed signals 103, 203 obtained in step (a2) are now subjected to signal padding, resulting in the padded signals 104, 204. In this way, during the application of the network (step (b)) To ensure that all signal time points are perceived equally often through the sliding analysis windows, the windowed signals 103 and 203 were padded on the left and right with the first and last amplitude values ​​of the respective observation window lengths, according to the long-term model and the short-term model. The signals are then normalized to the range 0 to 1. Step (b): Application of the long-term model and short-term model

[0042] Step (b) is characterized by the application of the two model architectures and their subsequent combination, i.e., the combination of rhythm and morphology features. In step (b), the padded signals obtained in step (aa) are used.

[0043] An ID Convolutional Neural Network classifier (1D-CNN) is capable of independently learning features in time series signals, such as an ECG time series, in order to perform a classification. To do this, the parameters of the two network models—the long-term model and the short-term model—are adjusted in an iterative process to minimize an error function. This error function describes the distance between a given class and the classification output of the two network models.

[0044] The 1D CNN consists of several convolutional layers 301, 401, in this example, the 9 convolutional layers, namely the convolutional layers 301-1 to 301-9 (see Figures 3 and 4). In each convolutional layer 301, 401, the input signal of the convolutional layer 301, 401 is convolved with at least one kernel. The kernel slides over the entire input signal, and the convolution result is fed to a nonlinear activation function for each shift time. This creates at least one feature map 302, 402, which indicates the degree to which the shape of the kernel matches the shape of the signal at the corresponding time, i.e., the degree of activation of the kernel. The first convolutional layer 301-1, 401-1 operates directly on the time series signal 204 to be analyzed, while the other convolutional layers 301-2 to 301-9 operate on the outputs of their respective preceding convolutional layers. This hierarchy allows complex Feature shapes are learned automatically. Following the final convolutional layer, the networks described here perform global average pooling 303, 403.

[0045] In the global average pooling 303, 403, the mean of the individual feature maps 302-9-1 to 302-9-24, 402-9-1 to 402-9-32 of the last convolutional layer 302-9, 402-9 is calculated. In this example, 24 feature maps are calculated. This is intended to assess the extent to which a feature is present in the entire input signal. In the last convolutional layer 302-9, 402-9, the presence of abstract features in the signal is evaluated. In this case, the convolutional layer 302-9, 402-9 itself only operates on the feature maps of the previous convolutional layer 302-8, 402-8 (not shown in Figures 3 and 4) and not on the input signal. Therefore, it is not directly checked to what extent the kernel of the last convolutional layer 302-9, 402-9 is present in the entire input signal, but to what extent the kernel of the last convolutional layer 302-9, 402-9 occurs in the feature map of the penultimate convolutional layer 302-8, 402-8.The output is a single number per feature map, which is then weighted and fed to a softmax function 304, 404 to perform the classification. The features are learned by adapting the kernel shape in the convolutional layers. By previously specifying the kernel size, it is possible to determine how large the perceptual range of a sample of a feature map should be with respect to the feature map of the previous convolutional layer or, in the first convolutional layer 301-1, 401-1, with respect to the time series signal 204. From this, the perceptual range of the last convolutional layer with respect to the time series signal 204, i.e., the input to the network, can also be calculated. This is given by the formula: [1], L is the number of convolutional layers, k is the kernel size of the convolutional layer and s is the stride, ie the offset between the initial times of the perception areas, whereby stride-1 samples are omitted. The use of Global Average Pooling 303 makes it possible to last convolutional layer 301-9 small. In the case of the classic CNN design, the perception region is chosen to be very large, since the entire feature maps of the last filter layer are input as a vector into a multilayer perceptron or other architecture for classification. A large perception region generates a small vector, and a small perception region generates a large vector, making the input for the subsequent classification network high-dimensional. This results in a large number of weights to be adjusted in the additional classification network, which complicates training. Global average pooling 303, 403 makes it possible to construct different networks that analyze an input signal in parallel on different time scales.

[0046] The long-term model is shown in Fig. 3. This is a network with a viewing window of 10 s. The short-term model is shown in Fig. 4. This is a network with a viewing window of 0.6 s. Both models are combined to improve classification performance by averaging the softmax outputs 304, 404 of both networks and using the argmax value as the classification result 305, 405. To evaluate the performance of the long-term model, the short-term model, and the combined model, the Fl-Score metric is applied as an example for a binary classification: 2 ■ TP Fl — Score: - , 2 ■ TP + FP + FN where TP stands for true positives, FP for false positives, and FN for false negatives. TP, FP, and FN can be read from the confusion matrix.

[0047] Details of the architecture of the long-term model are shown in box 306 (see Fig. 3). This shows that 96 feature maps f are generated from the first convolutional layer 309-1, where the input signal of the convolutional layer 309-1 was convolved using kernels of length 34 - this value describes the kernel size k - and a stride of 1 - this is the stride size s. 96 feature maps f are generated from the second convolutional layer 309-2, where the input signal of the convolutional layer 309-2 was convolved using kernels of length 35 - this The value is determined by the kernel size k and a stride of 1, which is the stride size s. The data for the third to ninth convolutional layers 302-3 to 303-9 are given in Box 306.

[0048] Details of the architecture of the short-term model are shown in Box 406 (see Fig. 4). This shows that 128 feature maps f are generated from the first convolutional layer 409-1, with the input signal of the convolutional layer 409-1 being convolved using kernels of length 16 – this value results from the kernel size k – and a stride of 1 – the stride size s. The second convolutional layer 409-2 produces 128 feature maps f, with the input signal of the convolutional layer 409-2 being convolved using kernels of length 17 – this value results from the kernel size k – and a stride of 1 – the stride size s. The details for the third to ninth convolutional layers 402-3 to 403-9 are given in Box 406. Step (c): Interpretation methodology

[0049] Step (c) is characterized by the combination of interpretation on a two-dimensional plane that has rhythm and morphology as coordinates.

[0050] Several methods are available in the literature for interpreting the classification results 305, 405. One group of these methods is based on tracing the activation of the softmax layer 304, 404 back through the network to the first filter layer 202 and from there to the time series signal 201 using so-called relevance propagation rules. This makes it possible to highlight which signal region (e.g., which region of the ECG) leads to the classification result 305, 405. One of these methods is Deep Taylor Decomposition [8]. The starting value for the calculation is the activation value of the softmax neuron with the highest activation. This relevance value is traced back to the results of the global average pooling 303, 403 using a relevance propagation rule (RFR), depending on the weights between the global average pooling 303, 403 and the softmax 304, 305 and the size of the global average pooling values, whereby these Relevance values ​​are then obtained. These relevance values ​​are then distributed among the feature maps 302-9-1 to 302-9-24, 402-9-1 to 402-9-32 of the last convolutional layer 302-9, 402-9 using an RFR, the kernel values, which are the weights, and the activation values ​​of the feature maps 302-9-1 to 302-9-24, 402-9-1 to 402-9-32. Finally, the relevance values ​​are propagated from feature map to feature map until they reach the network input, reaching the time series signal 201. Different RFRs can be used at different points in the network.

[0051] To combine the interpretation of the long-term and short-term models, the explanations of the classifications of both networks are plotted on a two-dimensional plane, and their coordinates provide information about the combined relevance of the signal content. The weighting of both models along these planes can be determined based on the decision accuracy weighting of the combined model. Step (d): Presentation in clinical monitoring

[0052] The presentation in clinical monitoring is characterized by a combination of interpretation on a two-dimensional level. Both rhythm features (long-term) and morphological features (short-term) are presented.

[0053] The relative relevance of the long-term model and the relative relevance of the short-term model in the combined two-dimensional plane 105, 205 can be projected onto the original signal 101, 201 (e.g., ECG) or the signal eliminated from interference 102, 202 using different colors in both axes, resulting in diagrams 106, 206 (see also Figures 1 and 5). This provides a direct representation of the individual relevances, namely long-term (rhythm) and short-term (morphology), visually visible and distinguishable. This enables feature representation in digital clinical monitoring.

[0054] Fig. 5 shows an exemplary combined representation of the relevance of the long-term model and the short-term model for integration into a clinical ECG monitor. The relevance is shown directly in the filtered signal 102, whereby the Diagram 106 is obtained. The combination is performed in a two-dimensional projection—this is diagram 106—with the two-dimensional plane 105 serving as the legend. In plane 105, a color is assigned to each relative relevance. These same colors are used in diagram 106 to indicate the relevance of each signal value.

[0055] The two-dimensional plane 105 shown in Fig. 5 shows color gradients within a square, with the "long-term" color changing from white (w) to violet (v) along the abscissa. Along the "short-term" ordinate, the color changes from white (w) to light blue (hb). Along the diagonal, which begins at the intersection of the abscissa and the ordinate, the color changes from white (w) to dark blue (db). The colors of these color gradients are used in diagram 106 depending on their relevance. Here, three colors fi, fi, and fi randomly selected in plane 105 are labeled in diagram 106 merely for illustrative purposes. The colors of the color gradient shown in the two-dimensional plane 205 are used in diagram 206 in the same way.

[0056] Fig. 6 shows an exemplary combined representation of the relevance of the long-term model and the short-term model in a clinical ECG rhythm strip 501. The relevance is shown directly in the filtered signal. The combination takes place in a two-dimensional projection – this is the ECG rhythm strip 501. The legend, which is a representation of the relative relevance of the long-term model and the short-term model in a combined two-dimensional plane 505, is shown in the top right. The two-dimensional plane 505 shows color gradients as already explained in connection with plane 105 (see description of Fig. 5). The colors of these color gradients are used in the ECG rhythm strip 501 depending on their relevance, in the same way as explained in connection with diagram 106 (see description of Fig. 5).

[0057] Both the clinical monitor (see Fig. 5) and the rhythm strip 501 shown in Fig. 6 are established methods for clinicians to make relevant diagnoses based on of the ECG. The integration of additional relevance in terms of special features in morphology and rhythm supports faster and more targeted clinical decisions in diagnostics and therapy. Example 1

[0058] One embodiment of the method according to the invention is explained in more detail below. This involves the application of the method according to the invention for the detection of atrial fibrillation (AF). (a) Data basis

[0059] Four ECG databases were used for comprehensive training of networks for detecting atrial fibrillation: Shapman Shaoxing [9], CPSC 2018

[0010] , PTB XL

[0011] ,

[0012] , and Georgia 12 lead

[0013] —

[0015] . Only channel II was used in each case. For a balanced dataset, all available 4927 AF ECGs were used, as were 4927 n-AF signals (where n-AF stands for "non-AF"). Furthermore, the number of sinus rhythm signals was set to 492 of the 4927 n-AF signals (10%) to generate the most diverse and balanced dataset possible. Of all ECGs, 90% were used for training and 10% for a test dataset. The training data is further split into training and validation datasets in a 5-fold cross-validation. (b) Combined long-term model and short-term model The long-term model was created with a viewing window of 10 seconds, and the short-term model with a viewing window of 0.6 seconds. Both models were combined with the same weighting. 10 seconds corresponds to the typical length of an ECG with AF in various available databases (see section (a) "Database," above), which provides the maximum range for detecting a rhythm. For the short-term observation, a range of 0.6 seconds was chosen to ensure that, in normal sinus rhythm, a maximum of one ECG heartbeat is captured in the viewing window, thus eliminating any beat-to-beat information. (c) Performance of the combined architecture

[0060] The performance of the long-term model, the short-term model, and the combined model is evaluated according to step 2 of Example 1 above, using confusion matrices and Fl scores for the respective models. The long-term model achieves an Fl score of 94.72%. The short-term model achieves an Fl score of 91.97%. The combination of the long-term model and the short-term model achieves an Fl score of 95.4%. The term "performance" describes the performance in terms of the classification of the respective model.

[0061] Fig. 7 shows the confusion matrix of a long-term model for detecting atrial fibrillation (AF). The "label" information corresponds to the truth, and the "classification" information corresponds to the prediction of the long-term model. Fig. 8 shows the confusion matrix of a short-term model for detecting atrial fibrillation (AF). The "label" information corresponds to the truth, and the "classification" information corresponds to the prediction of the short-term model. Fig. 9 shows the confusion matrix of a combined model for detecting atrial fibrillation (AF). The "label" information corresponds to the truth, and the "classification" information corresponds to the prediction of the combined model. (d) Exemplary feature representations with clinical relevance

[0062] The network architecture for rhythm and morphology analysis even learns features that have previously been considered clinical markers relevant for the detection of atrial fibrillation. Therefore, the interpretation of the feature representation in clinical monitoring is possible and clinically comprehensible. Two examples are presented below, highlighting clinically relevant markers for rhythm and morphology. These correspond to a clinician's textbook knowledge for diagnosing atrial fibrillation from the ECG.

[0063] Figure 10 shows an example sinus rhythm ECG with feature representation. Features with clinical relevance to morphology are highlighted in orange (ellipses 601) and features with clinical relevance to rhythm are highlighted in green (ellipses 602). The typical P wave (marked by orange ellipses 601 in Fig. 10), which represents atrial excitation after sinus node excitation, can be seen. Furthermore, a regular rhythm (distances between two R waves) is evident, which is highlighted in green in Fig. 10 by ellipses 602.

[0064] Fig. 11 shows an example atrial fibrillation ECG with feature representation. Features with clinical relevance for morphology are highlighted in orange (ellipses 701) and for rhythm in green (ellipses 702). It can be seen that the fibrillation waves typical for atrial fibrillation (marked by the orange ellipses 701 in Fig. 11) are morphologically relevant. Furthermore, an irregular Rhythm (distances between two R-waves) visible through the insistent coloration of the ECG in the green highlighted areas (ellipses 702). literature [1] N. Strodthoff, P. Wagner, T. Schaeffler, and W. Samek, “Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL,” IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 5, pp. 1519-1528, May 2021, doi: 10.1109 / JBHI.2020.3022989. [2] K. Gu, T. Prioleau, and S. Vosoughi, “Going Beyond Accuracy: Interpretability Metrics for CNN Representations of Physiological Signals,” in 2022 26th International Conference on Pattern Recognition (ICPRf Aug. 2022, pp. 4507-4513. doi: 10.1109 / ICPR56361.2022.9956192. [3] S. Vijayarangan, B. Murugesan, R. Vignesh, S. Preejith, J. Joseph, and M. Si- vaprakasam, “Interpreting Deep Neural Networks for Single-Lead ECG Arrhythmia Classification,” in 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBCf Jul. 2020, pp. 300-303. doi: 10.1109 / EMBC44109.2020.9176396. [4] T. Bender et al., “Analysis of a Deep Learning Model for 12-Lead ECG Classification Reveals Learned Features Similar to Diagnostic Criteria,” IEEE Journal of Biomedical and Health Informatics, pp. 1-12, 2023, doi: 10.1109 / JBHI.2023.3271858. [5] H. Honarvar et al. , “Enhancing convolutional neural network predictions of electrocardiograms with left ventricular dysfunction using a novel sub-waveform representation,” Cardiovascular Digital Health Journal, vol. 3, no. 5, pp. 220-231, Oct. 2022, doi: 10.1016 / j.cvdhj.2022.07.074. [6] D. Zhang, S. Yang, X. Yuan, and P. Zhang, “Interpretable deep learning for automatic diagnosis of 12-lead electrocardiogram,” iScience, vol. 24, no. 4, p. 102373, Apr. 2021, doi: 10.1016 / j .isci.2021.102373. [7] H.-Y. Lin, S.-Y. Liang, Y.-L. Ho, Y.-H. Lin, and H.-P. Ma, “Discrete-wavelet- transform-based noise removal and feature extraction for ECG signals,” IRBM, vol. 35, no. 6, pp. 351-361, Dec. 2014, doi: 10.1016 / j .irbm.2014.10.004. [8] G. Montavon, S. Lapuschkin, A. Binder, W. Samek, and K.-R. Müller, “Explaining nonlinear classification decisions with deep Taylor decomposition,” Pattern Recognition, vol. 65, pp. 211-222, May 2017, doi: 10.1016 / j .patcog.2016.11.008. [9] J. Zheng, J. Zhang, S. Danioko, H. Yao, H. Guo, and C. Rakovski, “A 12-Lead electrocardiogram database for arrhythmia research covering more than 10,000 patients,” Scientific Data, vol. 7, no. 1, p. 48, Feb. 2020, doi: 10.1038 / s41597- 020-0386-x.

[0010] F. F. Liu et al., “An open access database for evaluating the algorithms of ECG rhythm and morphology abnormal detection,” Journal of Medical Imaging and Health Informatics, vol. 8, no. 7, pp. 1368-1373, Sep. 2018, doi: 10.1166 / jmihi.2018.2442.

[0011] P. Wagner et al., “PTB-XL, a large publicly available electrocardiography dataset,” Scientific Data, vol. 7, no. 1, p. 154, May 2020, doi: 10.1038 / s41597-020- 0495-6.

[0012] P. Wagner, N. Strodthoff, R.-D. Bousseljot, W. Samek, and T. Schaeffter, “PTB- XL, a large publicly available electrocardiography dataset.” PhysioNet, 2022. doi: 10.13026 / kfzx-aw45.

[0013] E. A. Perez Alday et al., “Classification of 12-Lead ECGs: The PhysioNet / Com- puting in Cardiology Challenge 2020,” Physiological Measurement, vol. 41, no. 12, p. 124003, Dec. 2020, doi: 10.1088 / 1361-6579 / abc960.

[0014] E. A. Perez Alday et al., “Classification of 12-Lead ECGs: The PhysioNet / Com- puting in Cardiology Challenge 2020.” PhysioNet. doi: 10.13026 / dvyd-kd57.

[0015] A. L. Goldberger et al., “Physiobank, Physiotoolkit, and Physionet: Components of a New Research Resource for Complex Physiologic Signals,” Circulation, vol. 101, no. 23, pp. e215-e220, 2000.

Claims

Patent claims 1. A method for analyzing a quasi-periodic biological signal having measured values for a measurement period, the method comprising the steps of: (a) determining a first viewing window having a first time duration and a second viewing window having a second time duration from the quasi-periodic biological signal, wherein the first viewing window and the second viewing window lie within the measurement period and the second time duration is shorter than the first time duration; (b) classifying the measured values using the first viewing window by a first model to obtain a first classification result, and classifying the measured values using the second viewing window by a second model to obtain a second classification result, wherein the first model and the second model are each based on an artificial neural network; and (c) interpreting the first classification result and the second classification result by determining the relevance of one or more of the measured values to generate the first classification result and the relevance of one or more of the measured values to generate the second classification result.

2. The method according to claim 1, characterized in that it further comprises the step: (d) combining the relevance of one or more of the measured values for generating the first classification result and the relevance of one or more of the measured values for generating the second classification result by representation in a two-dimensional plane.

3. Method according to claim 1 or claim 2, characterized in that the first observation window comprises a time duration of several quasi-periods and the second observation window comprises a maximum of one quasi-period of the quasi-periodic biological signal.

4. Method according to one of the preceding claims, characterized in that the quasi-periodic biological signal is an electrocardiogram.

5. Method according to one of the preceding claims, characterized in that step (a) comprises the substeps: (ai) filtering the measured values of the quasi-periodic biological signal using a filter to remove noise components from the measured values and obtain filtered signals; (a2) windowing the filtered signal to obtain windowed signals; and (as) Padding the windowed signals while preserving padded signals.

6. Method according to one of the preceding claims, characterized in that the artificial neural networks used in step (b) are each based on a convolutional neural network.

7. The method according to claim 6, characterized in that the convolutional neural network consists of a plurality of convolutional layers, wherein at least one feature map is obtained in each convolutional layer.

8. Method according to claim 7, characterized in that in the temporally last convolutional layer the mean value of each feature map belonging to this convolutional layer is formed.

9. The method according to claim 8, characterized in that the mean value of each feature map is fed to a softmax function to obtain the first classification result and the second classification result.

10. Method according to one of the preceding claims, characterized in that the relevance of the first classification result and the relevance of the second classification result are each determined by means of relevance propagation rules.

11. Method according to one of claims 2 to 10, characterized in that the combined relevance is projected onto the quasi-periodic biological signal.

12. Method according to claim 11, characterized in that the combined relevance is visualized together with the quasi-periodic biological signal onto which it is projected.

13. Method according to one of the preceding claims, characterized in that characteristics of the quasi-periodic biological signal are determined by means of the analysis.

14. The method according to claim 13, characterized in that the characteristics determined by means of the analysis are characteristics of the rhythm and / or morphology of the quasi-periodic biological signal.

15. Method according to claim 13 or claim 14, characterized in that the determined features contain clinical relevance for diagnostic or therapeutic decisions.

16. Method according to one of the preceding claims, characterized in that the method is a self-learning method.

Images