HRP20140414B1 - System and computer implemented method of detection and recognition of wave forms in time series - Google Patents

Abstract

The present invention relates to a system and computer implemented method of detecting and recognizing waveforms in time series, in particular a physiological signal that includes a preprocessing of the original signal for the purpose of noise removal, extraction of morphological features of the signal, detection of interesting parts of the signal, extraction of dynamic features, recognition / classification waveforms and grouping / clustering of similar signal segments. The extraction of the morphological features of a signal is done by transforming the original one-dimensional signal into a series of characteristic vectors Vperceptions that summarize the morphology of the signal around the point at time T, which vectors All consist of a set of features m; m;…; signals, part is dynamically cumulative and part is the time-morphological determinant of VMD, whereby the series transformation is performed forward and backward relative to the time sequence. The signal transformation result is used as an input to the classifier for the purpose of building Model I for QRS section detection. Dynamic features, together with the morphological features of the signal (transformation result), form a set of extended features, that is, vectors with extended features, which, along with the corresponding source signal tags, become the basis for the supervised learning process for waveform recognition. The result of such a machine learning process is Model II, which is used to classify or recognize waveforms (Figure 10 c). Model II can then be used to recognize the waveforms of new signals (Figure 10 d) or the recognized ECG waves. Model I and Model II are stored on a computer or can be stored on any computer-readable medium.

Description

The field of technology

The subject invention relates to a system and a computer-implemented procedure for detecting and recognizing waveforms in time series, especially physiological signals. The procedure enables the detection of interesting shapes (predetermined shapes) of the curve and recognition (classification) of the shape of the curve. Based on the results of the procedure, automatic signal processing is achieved that does not require a human expert. The use of a computer-implemented procedure enables fast processing of large amounts of data. Automatic diagnostics based on models for the detection of interesting sections of the ECG signal (QRS complex) and the classification of waveforms enable faster diagnostics and facilitate the work of doctors. The present invention also relates to a computer program for the detection and recognition of waveforms in time series.

Technical problem

Although very precise and fast QRS detectors have been developed so far, and quite good results have been achieved in comparison to the existing tests for algorithms, automated ECG diagnostics have not yet taken root in practice to a sufficient extent. An electrocardiogram (ECG) is a drawing produced by an electrocardiograph, a device that records the electrical activity of the heart over time. The analysis of different waves and vectors of depolarization and repolarization leads to significant data in the diagnosis of diseases. ECG consists of P-wave, QRS-complex (the complex consists of Q, R and S-wave), T-wave and sometimes U-wave, see figure 1 and 2. In the subject field of technology, the most valuable achievements are manifested in methods which achieve very high detection results of the QRS complex of the ECG signal measured according to the given standards. "State of the art" methods are those that have sensitivity, positive predictivity and general accuracy of QRS complex detection above 95%. In addition to the detection of interesting segments (heartbeat, i.e. QRS complex), the challenge is also to recognize the shape of the waves, which enables a more precise diagnosis of disorders in the work of the heart. Waveform recognition (classification) methods are only approaching these thresholds.

The technical problem that is solved by the present invention is to provide a computer-implemented procedure for early diagnosis of heart diseases that provides greater accuracy and reliability of ECG diagnostics compared to known algorithms.

Another goal of the invention is to create a knowledge base that contains a set of extended features created by extracting dynamic and morphological characteristics of signals for the purpose of processing with machine unsupervised learning algorithms and finding the most similar signal segments for a new signal segment or the most similar waveforms for a new wave.

State of the art

Paper "Perception-based approach to time series data mining"; BATYRSHIN, I. Z., AND SHEREMETOV, L. B.. Appl. Soft Comput. 8, 3 (June 2008), 1211–1221. generally discloses a method of describing shapes from time series based on the principles of human shape descriptions and the use of such descriptions for time series data mining. Similar concepts of displaying time series based on the ideas of human perception and describing shapes are presented in the work of the author KLEPAC, G., Detecting regularities based on a unique model of time series transformation, doctoral dissertation, Faculty of Organization and Informatics, Varaždin, University of Zagreb, 2005. Paper describes a time series transformation model similar to that described in the subject invention (REF II). According to the subject invention, the described procedure provides a lower level of detail in the shape descriptions and includes smaller curve segments. The subject invention achieves a more precise description of the form of time series and thus a more reliable and precise application of the same in solving technical problems from the subject domain. Paper, below reference [1], "Automatic classification of heartbeats using ecg morphology and heartbeat interval features", DE CHAZAL, P., O'DWYER, M., AND REILLY, R. B., IEEE Transactions on Biomedical Engineering 51, 7 ( 2004), 1196–1206, shows a procedure for automatic processing of electrocardiograms (ECG) for heartbeat classification that includes three phases: pre-processing phase, processing phase and classification phase. The difference between the procedure according to the proposed invention and the procedure in the mentioned paper consists of an additional step in the signal processing phase by extracting morphological features by transforming the original one-dimensional signal into a series of characteristic VPT perception vectors that summarize the morphology of the signal in the vicinity of the point at time T. This step achieves greater accuracy (ACC) algorithm in relation to the results of the mentioned work (reference [1]). The work, below reference [2], "Heartbeat classification using feature selection driven by database generalization criteria", LLAMEDO, M., AND MARTÍNEZ, J. P., Biomedical Engineering, IEEE Transactions on 58, 3 (2011), 616–625, shows simple heart classifier operation based on ECG model features selected with an emphasis on improved generalization ability. As in the previous work (reference [1]), a higher accuracy of the algorithm was achieved by the procedure according to the subject invention. The paper, hereinafter referred to [3], "Heartbeat classification using morphological and dynamic features of ecg signals", YE, C., BHAGAVATULA, V., AND COIMBRA, M., presents a new approach for heartbeat classification based on morphological and dynamic characteristics.

According to current knowledge, the most successful approaches in published works, references [1], [2] and [3], which published results according to the AAMI (ANSI/AAMI EC57:1998/(R)2008) standard in comparison with the procedure according to the subject invention show lower accuracy (ACC) algorithm. Table 18 shows a comparison of three variants of the process according to the subject invention and some of the most successful approaches that have published results according to the AAMI standard. The computer-implemented procedure for detecting and recognizing waveforms in time series, especially the ECG physiological signal according to the subject invention, shows in all three ways of implementing the invention a higher accuracy (ACC) algorithm.

In addition to the mentioned works, the work, "A novel method for detecting r-peaks in electrocardiogram (ecg) signal", MANIKANDAN, M. S., AND SOMAN, K., Biomedical Signal Processing and Control 7, 2 (2012) can be cited as less relevant. ), 118–128 showing an algorithm for detecting QRS segments.

From the patent literature document US2008/0103403 shows a procedure for diagnosing silent/or symptomatic heart diseases in humans, based on the extraction and analysis of hidden factors, or a combination of hidden and known factors of the ECG signal. The diagnosis procedure uses the resting ECG signal of a group of diagnosed patients. The group consists of patients with a priori diagnosis as sick patients and patients with a priori diagnosis as healthy patients. Furthermore, the described method uses neural networks (NN) that receive a block of signals as input (after filtering without transformations). The main difference between the subject invention and document US2008/013403 consists in that said document describes a method that only classifies whether a patient is healthy or sick. The procedure according to the subject invention also distinguishes the types of disorders according to the AAMI standard. Furthermore, the procedure according to the subject invention does not take as input a sequential block of signals but a transformed signal, i.e. the result of the transformation of a part of the signal which then summarizes the morphological characteristics. Apart from the above, the document US2008/0103403 does not provide exact tests based on which the success of the algorithm could be compared with the algorithm according to the subject invention.

Subject of the invention

The subject invention relates to a system and a computer-implemented procedure for detecting and recognizing waveforms in time series, especially physiological signals. The procedure enables the detection of interesting shapes (predetermined shapes) of the curve and recognition (classification) of the shape of the curve. The procedure specifically refers to the transformation of the original time series signal into perception vectors, which is the basis for several applications. These are the construction of models for QRS detection and classification of QRS complexes, the construction of models for recognition of wave forms and classification of waves, and clustering (grouping) of similar waves and finding the waves most similar to the observed wave. After noise removal, the original signal is converted into a series of vectors by transformation according to the subject invention, using known signal filtering techniques. The transformed time series is the basis for learning a classifier (sorting algorithm) for the purpose of detecting QRS complexes in the signal. Given that it is a process of supervised learning, it is understood that the learning signal is marked (annotated), i.e. that it contains marks where the QRS complexes are located. The result of the machine learning process is model I for QRS detection (or classifier) which can be stored on the computer in the form of program code (Figure 10 - a)). Such model I can then be used for classification or detection of QRS segments in new signals (Figure 10 - b). The classification algorithm applied in the described invention is the Random Forest algorithm. Random Forest. The detected QRS complexes are the basis for extracting the dynamic features of the signal (based on the intervals between the QRS complexes). Dynamic features together with morphological features (result of transformation) form a set of extended features which, along with the corresponding labels of the original signal, become the basis for the process of supervised learning for the purpose of recognizing waveforms. The result of such a machine learning process is a model that is used for classification or recognition of waveforms (Figure 10 - c)). The model can then be used to recognize the waveforms of new signals (Figure 10 - d)).

A set of extended features created by extracting dynamic and morphological characteristics is stored as a knowledge base that is used for processing with machine unsupervised learning algorithms. Such algorithms result in the creation of groups of similar waves (clusters). Examples of such algorithms are the K-means algorithm, hierarchical clustering, self-organizing maps, etc. The set of extended features created by extracting dynamic and morphological characteristics is stored as a knowledge base that is used for the purpose of finding the most similar waves for a new wave. Such waves are found by calculating the mathematical distance between the vectors representing the waves in the knowledge base and the wave for which the most similar waves from the base are searched. Examples of distances are Euclidean distance, Manhattan distance, Mahalanobis distance, etc.

The development of a specialized classifier goes beyond the scope of this invention, and with regard to the problem of extremely unbalanced data, execution speed and some other parameters described below, existing classifiers were considered. After testing different classifiers on a limited data set, the Random Forest (RF) algorithm was selected. The RF algorithm is a relatively fast and very precise classifier, whose capabilities have been confirmed in earlier similar research. The random forest algorithm belongs to the group of so-called ensemble of classifiers (multiple separate classifiers connected in one model) and is based on decision trees.

A decision tree is a simple classifier obtained by building a tree structure in such a way that in each branch of the tree the data set is divided according to the variable that best separates the subset in that branch with respect to the target variable. The calculation of the "best variable" in a particular branch is based on the entropy or Gini measure of impurity. The process is repeated recursively until the conditions for the tree to stop growing are met. The advantages of decision trees are the simple mathematics on which they are based, the possibility of interpretation, i.e. the reading of the decision-making process in the sequence of branches of the tree from the root to the leaves, and the possibility of interaction with individual parts of the tree. The pitfalls of decision tree models are the fact that they predict well the class but not the probabilities for the class, the instability of the model, the tendency to overtrain (which is why growth limiting rules or pruning are built in) and the fact that they handle many variables poorly. In order to overcome these problems, Leo Breiman proposed the Random Forest RF model. In general, ensembles of classifiers can be divided into groups of independent classifiers (eng. bagging) and sequential classifiers (eng. boosting). The model makes independent decisions based on a voting principle. classifiers, each individual classifier independently makes a decision and then the final Sequential classifiers depend on each other, i.e. each subsequent classifier that is built improves the general model based on the results (errors) of the previous one. The RF algorithm is an example of independent decision trees that were built in such a way that only a part of the data set was used for each individual tree in the forest (sampling with repetition, English bootstrap) and a random subset of attributes was considered for each branching of an individual tree. In this way, two sources of randomness are introduced into the model, which helps prevent overtraining. In this way, each tree of the forest carries part of the information, and the overall, i.e., average score during the prediction is very accurate. According to many authors, the RF algorithm is one of the best machine learning algorithms in general, it tolerates many variables well and estimates the future error of the model well (data that did not enter the bootstrap sample remains for testing - Out of bag error). Furthermore, the RF algorithm itself performs variable importance analysis through the process of randomly building many trees and internally measuring the performance of each tree on a subset of the data that was not used for training. In addition, RF is not sensitive to non-normality in the distributions of the predictor variables or to their different scales. The disadvantage of the RF algorithm is the impossibility of interpretation, but the overall accuracy of the model was more important for the technical problem of the invention. Neural networks are certainly one of the logical candidates when choosing a machine learning algorithm. Several test classifications were performed and the results were not better than the RF algorithm. Additionally, with neural networks, it is necessary to "sacrifice" part of the data from the learning set for the purpose of validating the model (insurance against the overtraining effect), and considering the relatively small number of observations belonging to some target classes in this invention, neural networks were not the final choice. In the subject invention, implementations of the RF algorithm were applied in the software packages for machine learning Orange and Weka. The implementation in the Orange tool was used in QRS detection of unfiltered data, and in later tests (subject-based testing and shape classification), the "Fast Random Forest" implementation was used due to the size of the data sets and the length of the execution. As for the unbalanced data, a combined approach of the previously mentioned solutions was applied in the test on the filtered data. In particular, the data were subsampled in such a way that 5 percent of the data representing negative observations were selected, and that sample, together with all positive observations (QRS complexes), formed the basic sample on which the subject-based test described in the following text was performed. Also, the classifier was created through the so-called meta-classifier, which means that in the process of learning the RF algorithm, the cost function that optimizes the algorithm took into account the changed values of the cost of errors in favor of the correct detection of the QRS complex. Furthermore, given the unbalanced data, the required probability threshold given by the classifier to conclude that a beat has occurred is lowered from the assumed 0.5 to 0.4.

For the purpose of comparison with the aforementioned studies, digital filters were implemented in the detection of QRS complexes and in the classification of arrhythmias. Median-filter to remove the base line shift, and "buterworth" filter to remove noise caused by city network interference. Most recent research describes the application of the same or similar preprocessing techniques, which enables a correct comparison in terms of feature extraction and classification of the processed signals.

Short description of the pictures

In the following, the invention will be described in detail with reference to the drawings, whereby:

Figure 1 shows the morphological and dynamic characteristics of different ECG waves,

Figure 2 shows the characteristic parts of a normal ECG section,

Figure 3 shows typical steps in computer processing of ECG signals,

Figure 4 shows several sections of the curve with illustrated morphological differences that are attempted to be captured by vectors as a result of the transformation,

Figure 5 shows a conceptual representation of the method according to the subject invention,

Figure 6 shows the basic steps of the procedure according to the subject invention,

Figure 7 shows the sliding window method,

Figure 8 shows the pop-up method,

Figure 9 shows the transformation of the ECG signal according to the present invention,

Figure 10 shows a flowchart of the process according to the subject invention, i

Figure 11 shows the flow chart of unsupervised learning.

Detailed description of the invention

The human ECG signal shown in Figures 1 and 2 consists of several characteristic points or waves that reflect the activities of physiological processes in the heart. The EKG consists of a P wave, a QRS complex (the complex consists of Q, R, and S waves), a T wave, and sometimes a U wave. Most algorithms that are currently known aim to recognize the QRS complex of the ECG signal. Such algorithms are called QRS detectors because their recognition of the R wave actually implies the recognition of a QRS complex that includes local minima to the left and right of the R wave. The QRS complex actually reflects one heartbeat, and the correct detection of the QRS complex is the basis for heart rhythm analysis. Current methods achieve very high results in QRS complex detection and are all performed over filtered signals and implementing a search back phase. The backward lookup phase is actually a response to several problems that occur during beat detection. The first is a negative consequence of filtering the original signal, which is a change in the morphology of the original signal after filtering, where the extremes of the curve are very often shifted by several time units after filtering, most often in the direction of the filter passage. Because of this shift in the filtered signal, after identifying the potential R wave of the QRS complex, the ANSI standard allows searching for the actual R wave in the original signal in the 150 ms interval before or after the point found in the filtered signal. Another reason is the existence of sharp and high-amplitude T waves, and on the other hand, small and wide QRS complexes in some patients. Therefore, it is necessary to incorporate additional logic in order to avoid the detection of T waves during QRS detection, and enable the successful detection of small QRS complexes. In general, most methods of computer processing of ECG signals, whether it is QRS detection or arrhythmia classification, involve two basic steps before the actual detection or classification. The first step is signal preprocessing (noise removal), and the next step is feature extraction (shown in Figure 3). Figure 6 shows the basic steps of the process according to the subject invention, which includes the process of pre-processing the original signal for the purpose of removing noise, extracting morphological features of the signal, detecting interesting parts of the signal, i.e. QRS detection, extracting dynamic features, recognizing/classifying waveforms and grouping/clustering similar segments of the signal . According to the subject invention, a signal with reduced noise from the time domain will be transformed by a series of calculations in such a way as to obtain a multidimensional signal from a one-dimensional signal, the dimensions of which will actually speak of different morphological characteristics of the original signal. The extraction of the morphological features of the signal takes place by transforming the original one-dimensional signal into a series of characteristic VPT perception vectors that summarize the morphology of the signal in the vicinity of the point at time T, which VPT vectors consist of a set of features m1T ;m2T ;…;mnT , where part of the mentioned features describe geometric characteristics of the signal segment, part of which is dynamically cumulative, and part of the temporal morphological determinant VMD, whereby the transformation of the series is carried out forward and backward in relation to the time sequence, shown in Figure 9.

The result of the signal transformation is used as an input to the classifier for the purpose of building a model I for the detection of interesting shapes of signal segments. Model I is used to detect interesting segments of the signal for the purpose of detecting such segments in new signals, Figure 10 a and b shows the construction and use of Model I. The procedure for building Model I is as follows. The source or training signal is pre-processed to remove noise. The original one-dimensional signal with reduced noise is transformed into a series of characteristic vectors of VPT perception that summarize the morphology of the signal in the vicinity of the point at time T according to the above procedure, which will be described in more detail later. The result of the transformation is a series of characteristic VPT vectors or features that are used for learning for the purpose of QRS detection. The result is the construction of a classifier - model I, that is, a program code for QRS detection or generally interesting signal segments. Model I for QRS detection is used to classify segments in new signals that have previously been subjected to the denoising process and the above transformation. Given that it is a process of supervised learning, it is understood that the learning signal is marked (annotated), i.e. that it contains marks where the QRS complexes are located. The result of the machine learning process is model I, which is stored on the computer in the form of program code (Figure 10 a). Such a model is then used for the classification or detection of QRS signal segments in new signals (Figure 10 b). Furthermore, the result of the above transformation and the result of the detection of interesting shapes, i.e. QRS detection (Model I), is used as an input to the classifier for the purpose of building model II for distinguishing the shape of signal segments of the original signal or the learning signal and the new signal. The procedure for building model II is as follows. The source or training signal is pre-processed to remove noise. The de-noised signal is subjected to segment classification using I-model and then the dynamic features are extracted. At the same time, the signal with reduced noise is subjected to the above-mentioned transformation into a series of characteristic vectors of VPT perception. The result of these two simultaneous steps are vectors with extended features and the construction of a classifier or model II for recognizing the shape that is stored on the computer in the form of program code. The use of model II for recognizing the shape of signal segments for the purpose of recognizing the shape in new signals, i.e. recognized ECG waves, takes place in the following way. The new signal is subjected to the noise removal procedure and then to the above-mentioned transformation into a series of characteristic vectors of VPT perception. By using model I, the detection of interesting signal segments, or the classification of signal segments (QRS detection) takes place. The detected QRS segments are subjected to the extraction of dynamic signal features resulting in extended features or vectors with extended features. By using the previously built model II, the classification of waveforms takes place, which results in recognized ECG waves. The detected QRS complexes are the basis for extracting the dynamic features of the signal (based on the intervals between the QRS complexes). Dynamic features together with morphological features (result of transformation) form a set of extended features, ie vectors with extended features which, along with the corresponding labels of the original signal, become the basis for the process of supervised learning for the purpose of recognizing waveforms. The result of such a machine learning process is model II, which is used for the classification or recognition of waveforms (Figure 10 c). Model II can then be used to recognize the waveforms of new signals (Figure 10 d) or recognized ECG waves.

Model I and Model II are stored in the form of program code on a computer or may be stored on any computer-readable medium. The extraction of dynamic and morphological characteristics into a series of characteristic VPT vectors, i.e. the result of the transformation is used for the purpose of grouping by shape similar segments of the signal, the result of which is the creation of groups or clusters of similar segments of the signal. Grouping by shape similar signal segments with the original signal results in new associated information being stored. The extraction of dynamic and morphological characteristics into a series of characteristic VPT vectors, i.e. the transformation result is further used for the purpose of finding the most similar signal segments for a new signal segment or the most similar waveforms for a new wave.

The set of extended features created by extracting dynamic and morphological characteristics is stored as a knowledge base and is used for the purpose of processing with machine unsupervised learning algorithms. Such algorithms result in the creation of groups of similar waves (clusters). Examples of such algorithms are the K-means algorithm, hierarchical clustering, self-organizing maps, etc. Furthermore, a set of extended features created by extracting dynamic and morphological characteristics is stored as a knowledge base and is used for the purpose of finding the most similar waves for a new wave. Such waves are found by calculating the mathematical distance between the vectors representing the waves in the knowledge base and the wave for which the most similar waves from the base are searched. Examples of distances are Euclidean distance, Manhattan distance, Mahalanobis distance, etc.

VPT perception vectors summarizing the morphology of the signal in the vicinity of the point at time T consist of a set of features m1T ;m2T ;...;mnT , where part of the mentioned features describe the geometric characteristics of the signal segment, part are dynamic cumulative, and part are temporal morphological determinants of VMD. The dynamic cumulative components of the x vector [image] are calculated separately for the duration of the trend of the signal and for the duration of the concavity of the signal.

Below is a detailed description of a series of calculations that turn a one-dimensional signal into a multidimensional signal, the dimensions of which will actually speak about the different morphological characteristics of the original signal. Also, the procedure for calculating the cumulative component of the x vector [image] for the duration of the trend, the procedure for calculating the cumulative component of the x vector [image] for the duration of the concavity and the procedure for calculating the VMD value of the component x vector [image] for the duration of the trend are also described below.

A time series represents the value of one continuous variable over time. The fact that one and the same property is measured at different times means that the value measured at one time is directly dependent on the value measured at an earlier time. This property of time series prevents the application of standard statistical methods, given that measurements at different moments (observations) are not mutually independent. This mutual dependence of all observations will be the basis for the development of the new method presented below. As described earlier, human interpretation of time series is manifested in the description of the shapes we notice, that is, segments of the series or, more precisely, adjacent observations and the interrelationship of their values. Modern findings from the field of neuroscience prove that the neurons in the brain of higher mammals function in a similar way, that is, that groups of neurons behave in harmony and that their activation takes place simultaneously when precisely defined patterns are found in the subject's field of vision. These ideas were the starting point for the development of a concept that would somehow encompass the basic morphological characteristics of time series segments. First of all, it was necessary to decide how the parts of the time series would be segmented, and then to devise a way to describe the morphology of those segments. Inflection points were natural for segmentation, and morphological characteristics were tried to be summarized in a set of simple and fast calculations. The concept of using specialized neurons in describing the time series is shown in Figure 5. Both the segmentation and the morphological description of the series take place in the time domain, i.e. on the original series without converting it to the frequency domain or applying some other transformation (eg, wavelet). The signal from the time domain will be transformed by a series of mentioned calculations in such a way that a multidimensional signal is obtained from a one-dimensional signal, the dimensions of which will actually speak about the different morphological characteristics of the original signal. We will call the set of values of such a multidimensional representation of the original signal corresponding to exactly one moment in the original series "Perception vector" [image] . The perception vector [image] at time T consists of a set of values m1T ;m2T ; ….mnT ϵ M which summarize the morphology of the signal in the vicinity of the point at time T.

[image]

In terms of machine learning, the perception vector [image] will represent a transformed series at time T which is a de facto set of variables (features) and can be associated with a class K that will be used as a dependent or target variable in supervised learning or statistical modeling or can be considered independently as an observation of the original series at time T and used as a separate observation in unsupervised learning. The transformation that will fill all the dimensions of the vector [image] will simultaneously segment the series and fill in the vector or its components from the set M in such a way that some elements of the set M represent the morphology in the immediate vicinity of M1, and other elements represent the morphology in the further vicinity of M2 of the same segment, [image] and [image] . The transformation for creating set M is described below. In order to "capture" the morphology of the signal and fill all the elements of set M, the transformation of the original series will take place sequentially in the direction of the passage of time or the arrival of observations (recordings or measurements). This means that at different times t1 and t2 two vectors VPt1 and VPt2 will be filled. The transformation will encompass morphological characteristics through a series of cumulative calculations described below, and the above can be implemented in real time using the sliding window method. As can be seen in Figure 7, there are constant shifts of the window in which the time series is observed, so we figuratively say that the window "slides" through time. As will be shown later, the transformation does not have to take place only in one direction (the direction of the passage of time), but an equivalent procedure can be performed in the opposite direction, generating the elements of the set M, creating the subset MRev. The newly created set of morphological characteristics filled in this way can be called the metaset MM which is filled from both directions, i.e. [image] and [image] . In such a case, implementation would not be possible in real time, but in approximately real time using the hopping window method. Figure 8 shows the windows in which the time series is observed, which are not of constant spacing, but each subsequent one starts from some specific point of the time series detected in the previous window (extreme or inflection). Figuratively speaking, we say that the window "jumps" through time. Practically, this concept would be equivalent to the process in which a doctor watching the ECG signal in real time, before making his judgment, waits for the whole QRS complex to be drawn on the screen. After that, the doctor focuses his gaze on the drawn wave (the most prominent inflection - the extreme, i.e. the top of the wave) and based on the shape of the entire wave (left and right surroundings), makes a judgment, i.e. a diagnosis. The described concept is shown in Figure 9. The procedure for creating a perception vector by generating sets from one and the other direction is described below, given that the problem of shape recognition is being considered, where it is important to look at the entire shape, i.e. its left and right sides. For the sake of simplicity, in the rest of the text we will use the label M as a label for the metaset MM of morphological characteristics (both the left and right direction of transformation), i.e. the set of values that make up the vector [image] The transformation for the purpose of filling the elements of the vector [image] can take place on the original signal i.e. the original series or on a series that has previously been preprocessed (transformed) or resampled by some other method, eg minmax transformation, z-scaling or resampling (interpolation). The reason why such a thing would be necessary is the comparability of series from different sources. For example, if signals from different measuring devices are to be compared or processed with the same algorithm. When the signals are made in different scales, i.e. amplitudes, the min-max transformation can be applied

[image]

or z-scaling (reduction to standard deviation units)

[image]

where μ is the mean value and b is the standard deviation of the signal amplitude. The min-max transformation allows an arbitrary scale, but it has the disadvantage that we need to know the minimum and maximum values of the series we are measuring, which is often impossible in real-time implementations. In such situations, we can periodically calculate the standard deviation of the signal amplitude and express the original series in standard deviation units via the specified z-scaling transformation. This kind of transformation is necessary for the problem of variable amplitudes of the QRS complex. If the series we want to process are discretized with different frequencies, then before transformation into perception vectors it is necessary to resampling, considering that the values of individual vectors [image] directly depend on the number of observations that are included in a certain transformation cycle, which will be explained below. This research included data from the same reference base, and therefore it was not necessary to carry out such transformations, except when calculating the increment of the function, i.e. the angular deflection (angle), which will be explained below, and the normalized value is also stored in the independent variable normalized as a component of the vector [image] .

Considering the way in which human perception might function, and based on experiential knowledge (linguistic considerations and knowledge from neuroscience), signal inflection points were taken as dividers of individual segments. The methods of differential calculus make it possible to find extrema and inflection points by applying the first and second derivatives, i.e. by measuring the angle of the tangent to the curve at the moment T

[image]

that is, when Δx tends to zero

[image]

Since these are digital or digitized signals, the inflection points cannot be calculated by deriving a continuous function, but they are approximated as follows:

[image]

where x can be normalized to the range 0-1, and ΔX is 1, so there is no denominator. This is exactly how the first component of the vector [image] was calculated, which we will refer to below as angle (angle for the word angle). Alternatively, when we know that the sampling frequency during discretization of the original signal was constant, the derivative at time T can be considered as the angular deflection of the secant of the curve that passes through the points T − 1 and T + 1, given that such secant and tangent at point T are parallel:

[image]

The encounter of the transformation process with the extremum of the function will entail the logic of resetting the associated cumulatives, thus reflecting the morphological characteristics of the signal. In addition to the extremes themselves, finer changes in the trend, i.e. changes in the detailed trend, were taken as dividers for more detailed morphological characteristics (cumulative). In this way, it is possible to store the morphological characteristics of the signal for the near and far "neighborhood". As an identifier of these changes in the trend, the system of limits, i.e. the range between the values of the angular deflection, described in the REF II model was adopted, with the fact that, in addition to the gradation described by the REF II model, a more general component of the trend was also introduced, as shown in table 1 below.

[image]

Table 1: Value ranges for trend and detailed trend (REF) labels

The value of the limits g1-g6 was chosen empirically. The limit values applied in the subject invention are shown in Table 1. Different limit values correspond to different types of problems that are attempted to be solved by the described method. These ranges were chosen in such a way that they well encompass the differences in the steep parts of the QRS complex of the ECG signals of different pathologies. It is possible that the specified limits are not optimal. Table 2 shows the limits for angular deflection applied in the research.

[image]

Table 2

Thanks to the approximation of the first derivative, we can find the local extremes, that is, the minima and maxima of the curve. Furthermore, by the second derivative we can find the inflection points.

[image]

This fills in another component of the vector [image], which we will call the stationary point or according to the English name stationary_point.

The area under the curve was also taken from the REF II model, and the other morphological components described below were additionally included in the vector [image]. The listed components are the original contribution to the subject invention. The basic idea of calculating the area comes from the calculation of a definite integral, or the trapezoid rule. For the general case of a function over n points (x1; f(x1)), (x2; f(x2)), ...., (xn; f(xn)), where [image] are arranged in ascending order, the approximate value of the integral [image ] is determined by

[image]

As in the case of a signal discretized with a uniform frequency of 1, then the area between two points (T and T − 1) can be simply approximated by the formula

[image]

This fills in the third component [image], which we will call surface. An additional morphological characteristic that we will include is the concavity of the curve. On a discrete signal, we can find out the concavity of the curve at point T by observing the value of the point before and after the moment T and the relationship between the value at that point and the neighboring values. For this purpose, we will introduce the term midpoint, which will denote a fictitious mean value between the points that are previous and following in relation to point T. We will calculate the midpoint value as an interpolation between the previous and next point, that is, the value that the specified point would take in the event that the curve actually was the direction passing through the points T − 1 and T − 2. As we will calculate the transformation in one pass through the values of the time series, all calculations will be performed for the previous point (T −1 calculated on the basis of the adjacent points T and T − 2).

[image]

Furthermore, we will define the difference of the previous value in relation to the calculated mean, Δ mean as

[image]

and the difference between the current (real) value of the series and the previously calculated mean, Δ real as

[image]

We will put the calculated values of the middle in the ratio, and we will call the obtained value the concavity coefficient

[image]

Based on the relationship between the mean and the value at the previous moment (t−1), we will determine the concavity of the curve at the moment t−1:

[image]

The calculated coefficient forms a component of the vector [image], which we will call coefficient_concavity, and as in the case of the trend, we will discretize this variable by dividing it into classes as shown in table 2. This fills in another component of the vector [image], which indicates the detailed concavity, and we will call ju according to the English translation concavity_detail. In the case of a linear curve, i.e. no concavity, the concavity and detailed concavity variables will take the value 0. As with the limits for trends, the limits for concavities are chosen arbitrarily based on empirical conclusions. Also in the case of these limits, it is possible to design a model for their optimization.

The basis of the method of creating a vector of perception [image] is the effort to describe a one-dimensional curve or its shape using multidimensional vectors that will be unique for each point of the curve (time series), yet similar to other vectors that represent those points of the curve whose surroundings are similar in shape to each other alike. Such an effort was presented, that is, the procedure for creating a vector [image] was proposed. The procedure described below, to the best of the author's knowledge, has not been previously described in the literature and represents an original contribution. The idea to include the morphology of the signal "in the neighborhood" of the observed point to which the vector [image] is associated implies the so-called dynamic cumulatives that actually try to capture the exact morphological characteristics of the curve sections with simple and quick mathematical calculations. These cumulatives are dynamic because the duration of their cumulation is not fixed but is determined by the trend and the detailed trend on the segment of the curve being processed. Thus, we distinguish between cumulatives that are calculated for the duration of the trend and those that are calculated for the duration of the detailed trend. The discretization of the angular deflection variable into these two trend variables was important precisely because of these cumulatives, ie to achieve "resistance" to small changes in the trend and enable adequate accumulation of the variables. The general process of accumulating values that can be applied to any continuous variable from a transformation is described below.

[image]

Table 3: Value ranges for concavity labels and detailed concavities

[image]

Table 4: Limits for concavity applied in the research

The procedure for calculating the cumulative component of the x vector [image] for the duration of the trend is calculated as follows. If the trend at point T is equal to the trend at point T-1, then the trend duration is increased by one and the cumulative is increased by the value of the variable x. If the trend at point T is not equal to the trend at point T-1 then the trend duration is set to 1 , and cumulative to the value of the variable x. If the detailed trend at point T is equal to the detailed trend at point T-1, then the duration of the detailed trend is increased by one and the detailed cumulative is increased by the value of the variable x. If the detailed trend at point T is not equal detailed trend at point T-1 then the duration of the detailed trend is set to 1, and the detailed cumulative to the value of the variable x.

In addition to accumulating for the duration of the trend, we will also fill the vector [image] with components that speak about the morphology for the duration of the degree of concavity.

The procedure for calculating the cumulative x component of the vector [image] for the duration of the degree of concavity is calculated as follows. If the degree of concavity at point T is equal to the degree of concavity at point T-1, then the duration of the degree of concavity is increased by one and the cumulative value is increased by the value of the variable x. If the degree of concavity at point T is not equal to the degree of concavity at point T-1 then sets the duration of the degree of concavity to 1 and the cumulative to the value of the variable x. If the detailed degree of concavity at point T is equal to the detailed degree of concavity at point T-1, then the duration of the detailed degree of concavity is increased by one and the detailed cumulative is increased by the value of the variable x If the detailed degree of concavity at point T is not equal to the detailed degree of concavity at point T-1, then the duration of the detailed degree of concavity is set to 1, and the detailed cumulative to the value of the variable x.

The following variables and cumulatives were calculated in this way: duration of the trend (duration_trend), cumulative surface (cum_surface), cumulative concavity for the duration of the trend (cum_concavity_trend), cumulative change of the surface (cum_change_surface_trend), cumulative angular deflection (cum_angle), duration of the detailed trend ( duration_ref), cumulative surface change for the duration of the detailed trend (cum_change_surface_ref), cumulative angular deflection for the duration of the detailed trend (cum_angle_detail), cumulative concavity coefficient for the duration of the concavity degree (cum_concavity), cumulative concavity coefficient for the duration of the detailed concavity degree (cum_concavity_detail) ), cumulative concavity coefficient for the duration of the trend (cum_concavity_trend). Detailed cumulatives can also be made for the variables of cumulative area, cumulative concavity coefficient, etc.

In addition to the cumulatives described above, another non-linear method was devised for creating features or components of the vector [image], which aims to describe the morphology of the curve in the vicinity of a particular point of the series. Given that the specified calculation for the processed value x directly depends on the duration of the calculation (the number of points of the series included in the calculation) and on the shape of the curve (the calculation is performed on the basis of morphological characteristics), the obtained values are called "temporal morphological determinants" (abbreviated VMD) . In contrast to dynamic cumulatives that provide certain information about what shape is in the immediate environment, the goal of temporal morphological determinants is to provide information about where a shape is located and additionally to give the shape's deviation (morphology) further in the past (future). less affects the value of the VMD variable at the observed moment.

The procedure for calculating the VMD value of the component x of the vector [image] for the duration of the trend is calculated as follows. If the trend at point T is equal to the trend at point T-1, then the duration of the VMD counter is increased by one and the VMD component is increased by the product of the variable x and the VDM counter. If the trend at point T is not equal to the trend at point T-1, then the VMD counter is set to 1, and the VMD component is set to the value of variable x. If the detailed trend at point T is equal to the detailed trend at point T-1, then the value of the detailed VMD counter is increased by one and the detailed VMD component is increased by the product of the variable x and the detailed VDM counter. If the detailed trend at point T is not equal to the detailed trend at point T-1 then the value of the detailed VMD counter is set to 1 and the detailed VMD component is set to the value of variable x.

If we observe an extreme of the curve and assume that the trend before that extreme lasted for some time, thanks to the linear increase of vmd_counts and its multiplication with the target value x, we develop nonlinearity. More precisely, we can observe that thanks to this nonlinearity (multiplication), morphological changes closer to the observed point have a greater influence on the temporal morphological determinant than those "further in the past" (i.e. left or right). The following temporal morphological determinants (VMD) were created in the manner shown: VMD of surface changes for the duration of the trend (cum_vmd_surface_trend), VMD of the first derivative for the duration of the trend (cum_vmd_trend), VMD of the concavity coefficient for the duration of the trend (cum_vmd_concavity_trend), VMD of surface changes for duration of the detailed trend (cum_vmd_surface_ref), VMD of the first derivative for the duration of the trend (cum_vmd_ref), VMD of the coefficient of concavity for the duration of the trend (cum_vmd_concavity_ref). In backward transformation, the same features are generated with the rev prefix. As with dynamic cumulatives, the possibility of creating new time-morphological determinants is limited only by the creativity of the researcher. The entire described transformation was also performed backwards, and the list of created features (variables or vector components [image] ) is shown in the table below.

[image]

[image]

[image]

Table 5: Transformation variables applied in the research

Figure 4 shows several sections of the curve with illustrated morphological differences that are attempted to be captured by vectors as a result of the transformation. Differences in slope and surface (a), concavity (b), dynamic cumulatives (c) and VMD vector components (d) are highlighted. In order to illustrate the idea for individual components of the vector, several pairs of curves are shown in Figure 3. The first pair of curves (in Figure a) have obvious differences in slope and area under the curve. The second pair of curves (in figure b) shows pronounced differences in concavity. The third pair (in figure c) shows two sigmoidal curves which, after processing with the described transformation, differ greatly in the components of dynamic cumulatives. The fourth pair of curves (marked d in the figure) shows two similar segments that have a similar "belly", but located on different parts of the curves. This is covered by the VMD components of the vector [image] .

METHOD TESTING

For the purpose of conducting this research, a well-known existing data set - MITBIH Arrhythmia Database (hereinafter MIT-BIH AD) was used. The mentioned database is widely accepted and most works in the field report results on this database, which makes it a good "benchmark" for comparing various approaches and algorithms. The MIT-BIH AD base was created through the collaboration of MIT (Massachusetts Institute of Technology) and Beth-Israel Hospital (BIH) in Boston, USA. The database was developed as a standard test for evaluating algorithms for computerized ECG processing. The database contains 48 half-hour ECG records excluded from 47 patients. The signals are digitized with a frequency of 360 Hz. The recordings were selected in such a way that 23 contain normal sinus rhythm (NSR) and representative arrhythmias, while the other 25 contain rarer but clinically significant pathological ECG waves. The base is annotated, which means that the signals are also associated with the markings of individual ECG artifacts. These annotations were created by two or more cardiologists independently labeling the signals and then comparing and finally agreeing the labels by consensus. Each signal contains data from two ECG leads. In 45 cases the first lead is a modified arm lead II (MLII) and the second lead is usually a modified lead V1 (sometimes V2, V5 and in one case V4). In the other three cases, the first lead is V5 and the second lead is V2 (two records) or MLII (record 114 has reversed leads). The shapes of individual waves are different in the first and second drain. In general, it can be observed that normal sinus rhythm has clearer waves in the first lead, and some pathological conditions in the second lead.

TEST DESCRIPTION

Most algorithms for QRS detection presented in the literature do not require a learning phase. This means that the authors designed an algorithm that was tested on every single record from the MIT-BIH AD database with the same algorithm settings. With such approaches, there is no problem with selecting samples for learning and testing. The approach presented here requires a learning phase, and with such algorithms, the most realistic test is the so-called subject-based test. In such a test, each record from the database is extracted from the learning sample and serves for the test method, while the other signals are used for learning (training) and, if necessary, model validation (regularization and prevention of overtraining). Such an approach is the closest to the potential clinical application of the method, because during the test the algorithm is exposed to the ECG signal of a patient "whom it has never seen before" (the data was not even used for learning a series of model validations). This effectively means that 48 separate models (one model for each track) were built and tested. The models were built (learned) on a sample that contained 5% of irrelevant extremes and all QRS complexes from 47 records and were then tested on the entire data set of the 48th record. In this way, the degree of generalization of the model, ie its application to new data, is checked in the best possible way. With algorithms that do not require learning, there is a possibility that the parameters of the algorithm are adjusted for all records from the test database, and although the algorithm does not require learning, it is not possible to reliably assess the generalization possibilities for application to new data. The Fast Random Forest implementation was applied, and the built model had 1000 trees in the forest, each of which was built by considering 15 random variables. Due to extremely unbalanced data, the experiments showed that the best results were achieved by applying the meta classifier with the following cost matrix, shown in Table 6.

[image]

Table 6: Cost matrix of individual classifications

The value of 3.2 for the FN error cost was chosen for two reasons. The first reason is that after subsampling the nonsignificant wave class, the ratio was still 1.6 to 1 in favor of nonsignificant, and the second reason is that tall and sharp T waves were observed in some records versus low-amplitude QRS complexes in others (eg, record 108). Therefore, the cost of the FN error is twice the ratio of the distributions within the beat class.

Below, Table 7 shows the results of QRS detection over filtered data - 360 ms

[image]

[image]

[image]

Table 7

The results show the excellent ability of the method in distinguishing important from non-essential waves of the ECG signal, which is comparable to state-of-the-art methods published in the literature. Errors are noticeable in records that are difficult to recognize even according to the literature and that contain a lot of noise or small QRS complexes (like record 108). Of particular note are the 30 FP errors from record 113. An examination of the errors revealed that they are T waves that are outside the threshold of 360 ms. As the sensitivity of the classifier is increased by lowering the probability threshold for accepting a QRS complex to 0.4, it can be expected to recognize a large number of T waves as QRS complexes. These errors would be eliminated by increasing the threshold (eg to 375 ms), but then a fair comparison with works from the literature would not be possible. A comparison with the most successful approach according to the available literature is given in Table 8.

[image]

Table 8: Comparison of the proposed method and approaches from the literature

When we talk about the classification of ECG waveforms, it is very important to highlight the difference between the testing methodologies used when publishing the results. The testing methodology, i.e. the way of selecting samples for learning and testing significantly affects the test results. A common approach in recent works that present research in the field of classification of arrhythmias, ie ECG waveforms, is class-based testing. In this type of research, the entire ECG signal database is divided into a learning data set and a testing data set. A common method is cross-testing with several folds, which means that the performance of the algorithm is measured on different test samples and then average performance measures are calculated. The problem with this approach is method bias due to the already mentioned variation in ECG waveforms that exists between patients, which is not adequately covered by the test in this way.

The AAMI standard describes procedures and metrics for testing algorithms for recognizing ECG signal pathologies. Over the past few years, papers have appeared that follow the mentioned recommendations, which enables a relevant comparison of different approaches. There are surprisingly few papers reporting results according to this standard in the literature. In this part, the results of testing will be presented in accordance with the recommendations of the standards and practices in the aforementioned published research. The specified standard recommends the omission of ECG records that contain a controlled rhythm (pacemaker), specifically records 102, 104, 107 and 217. Furthermore, the set of other records is divided into two subsets, each of which contains 22 records. The division of records into subsets for learning and testing is shown in the table. As this approach to testing ensures that all records that are included in the test data set are not in the learning set, this method is also understood in the literature under the phrase "subject-based testing". Some of the published works use this testing methodology with the additional exception of including a small portion of the test signal (eg the first 5 minutes) in the learning process to adapt the algorithm to the patient in question. Although the results in this case are significantly better, it should be noted that in this case it is not a fully automatic classifier, because in real clinical application the intervention of an expert would be required, who would manually annotate part of the signal for each patient during recording. The results presented here do not include such adjustments but are a fully automatic classifier.

[image]

Table 9: Division of the MIT-BIH AD database into a learning subset and a testing subset

The division is common in the mentioned works and was done in such a way that in both subsets there is an approximately equal number of observations of each class. The original 16 beat classes are grouped according to the AAMI standard into 5 classes, namely "N" (any designation not belonging to the "S", "V", "F" or "Q" classes), "S" (supraventricular ectopic beat ), "V" (ventricular ectopic beat), "F" (coupled beat) and "Q" (unknown beat). The original 16 classes are shown in Table 12. The mapping of MIT-BIH AD classes to AAMI classes is shown in Table 11. Algorithm learning is performed on the first subset, and testing on the second. In this way, the algorithm is exposed to half of the test database in such a way that none of the records on which it is tested is included in learning, and this kind of test methodology is a good indicator of the possibility of generalization of the algorithm and the possibility of potential clinical application. As some waves belonging to different classes are very similar in shape, but differ in dynamic characteristics (eg N and A waves), it is necessary to include certain dynamic characteristics as components in the vectors [image] .

[image]

Table 10: Original MIT-BIH AD classes (entire database without excluding the pacemaker)

[image]

Table 11: Mapping of MIT-BIH AD classes to AAMI classes

Given that the subject invention is focused on the analysis of morphological characteristics, an arbitrary minimum of such components is included. We find numerous other dynamic characteristics in the literature, and their inclusion can be the subject of future research and potentially improve recognition results. The dynamic characteristics that are included as features are shown in table 12. The time of the observed point expressed in milliseconds is marked as T0, the time of the previous point as T−1, and so on.

[image]

Table 12: Dynamic features included in the vector

In order to "capture" the variability of the rhythm, that is, the relative position of the observed wave in relation to its immediate neighboring waves, the local average of the duration of the interval between the previous ten QRS complexes was observed. Additionally, only one wave in the future is included, which leaves the possibility of near-real-time implementation using the pop-up method described earlier. The ratios of the previous and next intervals and the average of the last ten waves were put in an additional mutual ratio, and due to the asymmetry of the distribution of the respective variable, a logarithmic transformation was performed.

In order to recognize the waves, two models were built, one based on the records from the first ECG lead and the other based on the records from the second lead. Both models have the same structure, i.e. they are forests with 500 trees, each of which is built by considering 15 random variables.

[image]

Table 13: Confusion matrix for recognizing ECG waveforms based on the first lead and all attributes

The overall accuracy of the model built on the basis of the test set of the first drain is 95.812%. We can notice that the worst classification is "F" waves. These results coincide with the findings of other authors whose approaches also significantly err in recognizing the "F" beat [2].

[image]

Table 14: Confusion matrix for the recognition of ECG waveforms based on the second lead

The model based on the second lead has a significantly lower overall accuracy of 90.6795%, but achieves a significantly better classification of "F" waves. This is to be expected because the second lead (which gives a perspective on the heart from another angle) shows such beats with a higher amplitude, so the classifier can distinguish them more easily. Considering only the classification results obtained on the first drain and comparing them with the current state of the art, i.e. published works in the literature, regardless of the impossibility of recognizing the "F" class, and considering the very good results in recognizing the other classes, we can state that the method belongs to the current "state of the art" methods. It is also possible to combine conclusions from two leads using conditional probabilities (Bayesian product), which is already described in the literature. For the purpose of comparison with some of the most successful algorithms available in the literature, the confusion matrices of the research presented in [2], [3] and [1] are presented.

[image]

Table 15: Confusion matrix for recognizing the shape of ECG waves ref. [2]

[image]

Table 16: Confusion matrix for recognizing the shape of ECG waves ref. [3]

[image]

Table 17: Confusion matrix for recognizing ECG waveforms ref. [1]

[image]

Table 18: Comparison of the performance of different approaches

Table 18 shows a comparison of three variants of the proposed method and some of the most successful approaches that have published results according to the AAMI standard. The proposed method marked as "a" refers to the classification based on the first ECG lead that includes all features. The variant marked with "b" implies a classification also based on drain 1, but with the use of the most relevant 30 attributes, while the variant "c" indicates a classification based on the second drain. The success rates where a particular approach is the most successful (among the three proposed) are marked in bold, while the higher success rates reported in the literature are indicated in italics. There are other studies in the literature that use the AAMI testing standard, but with the inclusion of the first 5 minutes of test ECG recordings in the learning phase. Such tests are not a completely correct indicator of potential clinical application, considering that it would take 5 minutes for each patient to mark the signals, which is not always practical in clinical situations and is time-consuming. Therefore, a comparison with such research was not even made.

Below is a brief description of the system for the detection and recognition of waveforms in time series, especially the physiological ECG signal, which is based on the application of signal transformation, as well as on a computer program and on a computer-readable medium.

A system for detecting and distinguishing waveforms in time series, especially physiological signals, which contains an input circuit arranged to receive, record and store waveforms in time series, a device arranged to remove noise from the original signal, which further contains:

- a processor arranged in such a way as to extract the morphological features of the signal by transforming the original one-dimensional signal into a series of characteristic VPT perception vectors that summarize the morphology of the signal in the vicinity of the point at time T, which VPT vectors consist of a set of features m1T ;m2T ;…;mnT , at where a part of the mentioned features describes the geometric characteristics of the signal segment, a part is dynamic cumulative, and a part is a temporal morphological determinant of VMD, whereby the transformation of the series is carried out forward and backward in relation to the time sequence;

- model I of the classifier stored on the computer in the form of a program code that uses the result of the transformation for the classification or detection of interesting parts of the signal in new signals;

- model II of the classifier stored on the computer in the form of a program code that uses the result of the transformation and the result of the detection of interesting shapes for classification, that is, the recognition of waveforms, which is used to recognize the waveforms of new signals; and

- a knowledge base that contains a set of extended features created by extracting dynamic and morphological characteristics of signals for the purpose of processing with machine unsupervised learning algorithms and finding the most similar waves for a new wave.

The subject invention also relates to a computer program for the detection and recognition of waveforms in time series, especially physiological signals, which is adapted so that when executed on a computer, it causes the computer to perform a computer-implemented process of detection and recognition of waveforms in time series. According to the subject invention, the waves in the time series are ECG waves. The present invention also relates to a computer-readable medium on which the said computer program is stored.

The subject invention, i.e. a computer-implemented procedure in the form of a program code, which makes the procedure suitable for various applications such as software on a computer, a mobile phone or a portable EKG device (holter). Scientific and professional literature mentions systems in which applications for receiving ECG data via a wireless connection (e.g. Bluetooth) are implemented on mobile phones, and applications on the mobile phone process the signal and send critical parts of the signal or the entire signal to a central database via the mobile network. However, systems described in the literature do not display signal processing with such accuracy as the present invention. In addition to the implementation of software on a mobile phone, the subject invention can also be implemented as embedded software in a specialized ECG device (e.g. holter). In this case, the specialized ECG device, in addition to the function of recording ECG signals, also has the function of signal processing and detection of QRS segments and classification of ECG waveforms. Such a device can also have circuitry and software for communicating with other systems and sending signals and notifications to a central database.

APPLICATION OF THE METHOD IN UNSUPERVISED LEARNING

The previously described way of learning algorithms, where in the data set for each observation we have information about the class to which it belongs, belongs to the so-called supervised learning. In a situation where we do not have information about the class to which an observation belongs, the task is to determine which group it belongs to, i.e. which of the known observations are most similar to it. We also call the mentioned problem grouping or clustering. Only a conceptual model of such a process is presented below. The basis of unsupervised machine learning algorithms is the calculation of the distance, that is, the difference between observations. The simplest example of distance calculation is the Euclidean distance. If two observations a and b are described by the variables (a1; a2; a3) and (b1; b2; b3), then the distance between them (Δ) can be calculated as

[image]

When calculating this distance, it is necessary to carry out the previously mentioned normalization so that all variables are on the same scale (minmax transformation, z-scaling, etc.). Euclidean distance does not take into account the variability of data within individual variables, and therefore there are more "advanced" methods of distance calculation. One of the more advanced distance calculation methods that takes data dispersion into account is the Mahalanobis distance. Scatter capture was realized using the data covariance matrix. In fact we can say that the Mahalanobis distance is in a way a multi-dimensional z-scaling. With a given list X consisting of N observations where each observation can be K dimensional (vector length) and a vector μx (consisting of individual means μ1,……K), the covariance is a K * K matrix

[image]

With the inverse of the matrix Σ, we can calculate the Mahalanobis distance analogously to the Euclidean distance by excluding the space covariance from the calculation using the matrix Σ −1

[image]

Mahalanobis distance is a number that is actually a measure of similarity between two observations of a multidimensional space. There are many other methods of distance calculation, but their explanation is beyond the scope of this paper.

Claims (17)

1. A computer-implemented procedure for detecting and recognizing waveforms in time series, especially the ECG physiological signal, which includes the process of preprocessing the original signal for the purpose of removing noise, extracting morphological features of the signal with reduced noise, detecting interesting parts of the signal, extracting dynamic features, recognizing and classifying shapes waves and grouping/clustering of similar segments of the signal, indicated by the fact that the extraction of morphological features of the signal takes place by transforming the original one-dimensional signal into a series of characteristic vectors of VPT perception that summarize the morphology of the signal in the vicinity of the point at time T, which VPT vectors consist of a set of features m1T; m2T;...;mnT, where a part of the mentioned features describes the geometric characteristics of the signal segment, a part is dynamic cumulative, and a part is the time-morphological determinant VMD, where the transformation of the series is carried out forward and backward in relation to the time sequence. 2. The method according to claim 1, characterized in that the result of the signal transformation is used as an input to the classifier for the purpose of building model I for the detection of interesting shapes of signal segments, and the use of model I for the detection of interesting signal segments for the purpose of detecting such segments in new signals, and storing model I for detection of interesting segments of the signal in the form of program code. 3. The method according to claims 1 and 2, characterized by the fact that the transformed signal segments are grouped by shape into groups/clusters with similar signal segments, whereby new associated information is generated about the grouping of segments with the original signal, which are used in the process of unsupervised machine learning to find similar shapes of signal segments for a new signal segment. 4. The procedure according to claims 1, 2 and 3, characterized by the fact that the signal is simultaneously subjected to the classification of the shape of signal segments using model I, after which the dynamic features are extracted and transformed into a series of characteristic vectors of VPT perception, whereby the result of these two simultaneous steps are vectors with extended features for the purpose of building a shape recognition model II, and using the signal shape recognition model II for the purpose of shape recognition in new signals, and storing the signal shape recognition model II in the form of program code. 5. The method according to claim 4, characterized in that the transformed signal forms are grouped by shape into groups/clusters with similar signal forms, whereby new associated information about the grouping of forms is generated in addition to the original signal, which in the process of unsupervised machine learning is used to find similar signal forms for a new form of signal. 6. The method according to claim 1, characterized in that the dynamic cumulative components of the x vector [image] 6. calculate separately for the duration of the trend of the signal and for the duration of the concavity of the signal. 7. The method according to claim 6, characterized in that the calculation of the cumulative component of the x vector [image] 7. for the duration of the trend, it takes place in the following way: if the trend at point T is equal to the trend at point T-1, then the duration of the trend is increased by one and the cumulative value is increased by the value of the variable x, if the trend at point T is not equal to the trend at point T-1 then the duration of the trend is set to 1 and the cumulative to the value of variable x, if the detailed trend at point T is equal to the detailed trend at point T-1, then the duration of the detailed trend is increased by one and the detailed cumulative is increased by the value variable x, and if the detailed trend at point T is not equal to the detailed trend at point T-1, then the duration of the detailed trend is set to 1, and the detailed cumulative to the value of the variable x. 8. The method according to claim 6, characterized in that the calculation of the cumulative component of the x vector [image] 8. during the duration of the concavity, it takes place as follows: if the degree of concavity at point T is equal to the degree of concavity at point T-1, then the duration of the degree of concavity is increased by one and cumulatively increases by the value of the variable x, if the degree of concavity at point T is not equal to the degree of concavity at point T-1 then the duration of the degree of concavity is set to 1 and the cumulative to the value of variable x, if the detailed degree of concavity at point T is equal to the detailed degree of concavity at point T-1 then the duration of the detailed degree of concavity is increased by one and the detailed cumulative is increased by the value of variable x, and if the detailed degree of concavity at point T is not equal to the detailed degree of concavity at point T-1 then the duration of detailed degree of concavity is set to 1 and the detailed cumulative is set to the value of variable x. 9. The method according to claim 1, characterized in that the calculation of the VMD value of the component x of the vector [image] 9. for the duration of the trend it takes place as follows: if the trend at point T is equal to the trend at point T-1, then the duration of the VMD counter is increased by one and the VMD component is increased by the product of the variable x and the VDM counter, if the trend at point T is not equal to the trend at point T-1 then the VMD counter is set to 1 and the VMD component is set to the value of variable x, if the detailed trend at point T is equal to the detailed trend at point T-1 then the value of the detailed VMD counter is increased by one and the detailed VMD component is incremented by the product of the variable x and the detailed VDM counter, and if the detailed trend at point T is not equal to the detailed trend at point T-1 then the value of the detailed VMD counter is set to 1 and the detailed VMD component is set to the value of the variable x. 10. The method according to claim 1, indicated by the fact that the geometric characteristics of the segment of the signal include: angular deflection and backward angular deflection, normalized original value, area under the last segment and area under the last backward segment, approximations of the first derivative and approximation of the first derivative backward, the last change of the surface and last area change backward, trend and backward trend, detailed trend and detailed backward trend, trend duration and backward trend duration, detailed trend duration and detailed backward trend duration, concavity coefficient and backward concavity coefficient, concavity label and backward concavity label, detailed label concavities and detail backward concavity label, type of stationary point and duration of degree of concavity and duration of degree of backward concavity. 11. The method according to claims 1 to 6, characterized by the fact that the dynamic cumulative signal segment consists of: cumulative area for the duration of the trend, cumulative area for the duration of the backward trend, cumulative change of the area during the trend, cumulative change of the area during the backward trend, cumulative change area during the detailed trend, cumulative area change during the detailed trend back, angle cumulative during the trend, angle cumulative during the trend back, angle cumulative during the detailed trend, angle cumulative during the detailed trend back, concavity cumulative for duration of concavity, cumulative concavity for the duration of backward concavity, cumulative concavity for the duration of concavity in detail, cumulative concavity for the duration of concavity in detail backward, cumulative concavity for the duration of the trend and cumulative concavity for the duration of the trend backward ag. 12. The method according to claim 1 and 9, characterized in that the temporal morphological determinants of the signal segment (VMD) consist of: first derivatives for the duration of the trend, VMD of the first derivatives for the duration of the backward trend, VMD of the first derivatives for the duration of the detailed trend, VMD of the first derivatives for the duration of the detailed trend backward, VMD concavities for the duration of the trend, VMD concavities for the duration of the trend backward, VMD concavities for the duration of the detailed trend, VMD concavities for the duration of the detailed trend backward, VMD surfaces for the duration of the trend, VMD surfaces for the duration of the backward trend, the VMD area for the duration of the detailed trend and the VMD area for the duration of the detailed trend backward. 13. Procedure according to the previous requirements, characterized by the fact that the Random Forest algorithm is applied for signal classification. 14. The method according to the previous claims, characterized in that the interesting parts of the signal are the QRS complex in the ECG physiological signal. 15. A system for detecting and distinguishing waveforms in time series, especially physiological signals, which contains an input circuit that is arranged for receiving, recording and storing waves in time series, a device arranged for removing noise from the original signal, characterized by the fact that it further contains: a processor which is organized in such a way as to extract the morphological features of the signal by transforming the original one-dimensional signal into a series of characteristic VPT perception vectors that summarize the morphology of the signal in the vicinity of the point at time T, which VPT vectors consist of a set of features m1T ;m2T ;…;mnT , where part of the mentioned features describes the geometric characteristics of the signal segment, part of which is dynamic cumulative, and part of which is the time-morphological determinant of VMD, whereby the transformation of the series is carried out forward and backward in relation to the time sequence; model I classifier stored on the computer in the form of program code that model I uses the result of the transformation for classification or detection of interesting signal segments in new signals; model II classifier stored on the computer in the form of program code, which model II uses the result of transformation and the result of detection of interesting shapes for classification, i.e. recognition of waveforms, which is used for recognition of waveforms of new signals; and a knowledge base that contains a set of extended features created by extracting dynamic and morphological characteristics of the signal for the purpose of processing with machine unsupervised learning algorithms and finding the most similar signal segments for a new signal segment or the most similar waveforms for a new wave. 16. A computer program for the detection and recognition of waveforms in time series, in particular a physiological signal, characterized in that it is adapted to, when executed on a computer, cause the computer to perform the procedure according to claims 1 to 14. 17. Computer-readable medium, characterized in that it contains a stored computer program from claim 16.