Apparatus for classifying physiological events

The present invention relates to a device for classification physiological events

Physiological events evoke physiological signal or filters even signals constitute the basis of which they can be classified. The Classifying physiological events or signals is particularly in implantable medical devices such as pacemakers or implantable defibrillators useful to events requiring treatment by those who are not in need of treatment, or events for which different treatments are shown, each other to differ. Based on the classification, the implantable medical device able to independently install any necessary initiate treatment.

Existing devices for the classification of physiological events, especially of intracardiac events in implantable medical Devices are essentially based on a filtering of the waveform and in the provision of a threshold value or more Thresholds combined with time analysis of the over- / undershooting the threshold or thresholds.

To the previous devices acceptable sensitivity to the Signals of physiological events and an acceptable distinctness To achieve of events, it is necessary, during the cardiac cycle, in the event falls, the recording of further physiological Signals suspend. However, this exposure includes reliable Recognize various important event classes of intracardiac events and their effective treatment, so eg. an abnormal Relationship between the two ventricles.

It is therefore an object of the present invention to provide an improved Apparatus for classifying physiological events, particularly intracardiac events, to provide that the said Disadvantages helps overcome.

This problem is physiological by a device for classification Events dissolved Claim 1. The dependent claims contain further embodiments of the invention.

An inventive apparatus for classifying physiologic Events has a signal input for inputting a on a physiological event based, or it representative or representing the physiological signal on and a physiological Signal based on its waveform classifying classification unit. In the A device according to the invention comprises the classification unit:<ul><li>a transformation unit for performing such a transform of the physiological signal is adapted that it as an output signal a number of the physiological signal representing and process based on the transformation values ​​outputs; and</li></ul>   one for receiving the values ​​associated with the transform unit probabilistic neural network, the number one containing physiological events representing event classes (Which are typically implemented in the form of so-called. Nodes) represented in turn by a respective set of comparison values , said probabilistic neural networks to configured is one based on the comparison of the values ​​with the comparison values Assignment of the directory represented by the values ​​of physiological Signal make one of the event classes.

Classifying done so by associating the physiological Event representative or representing the physiological signal to an event class. As physiological signal in this case is also a before Entering the classification unit recycled, for example. Normalized, filtered, Adjusted or similar reshaped physiological signal viewed will.

The combination of transformation unit and probabilistic neural Network forms an effective classification unit for classifying Signals intracardiac events or other events or other Signals of biological origin.

The transformation unit is a development of the invention one on a wavelet transform based transformation unit, ie it is to perform the transformation on the basis of wavelets and the values ​​to be output using the wavelets defining transformation rule configured.

The wavelet transform is easy to perform and allows Signals to represent a relatively small number of values ​​(coefficients). Simultaneously obese in the wavelet transform enough information receive via the signal to provide a reliable classification in probabilistic neural network to ensure. In addition, provides the Wavelet transformation, the possibility of the transformation within the for calculating the transformation existing mathematical limits to adjust the effective recognition of individual event classes.

contain the values ​​obtained by means of the wavelet transformation preferably values ​​that describe a stem wavelet and additionally Scaling values ​​and translation values ​​(coefficients) which, based on a respective stem wavelet the shape of the input signal (physiological Signal) characterize.

The comparison of values ​​and the comparison values ​​in the probabilistic neural network is guaranteed especially true when the Comparison values ​​to a transformation based, in which the same Wavelets and the same transformation rule as in the find transformation unit use.

In one embodiment of the probabilistic neural network includes this of the represented for performing the assignment of the values physiological signal to one of the event classes:<ul><li>at least one detection unit for determining association probabilities the physiological signal to the Event classes on the basis of comparison of the values ​​with the Comparative values ​​of each event class and outputting the determined association probabilities; and</li><li>a to the determination unit for receiving the association probabilities associated selection unit configured to Extracting the highest probability of belonging from the Association probabilities and to associate the physiological Signal to the event class having the highest probability of belonging is trained.</li></ul>

In a development of the inventive device can be used for a Event class two or more sets of the same event class representing comparative values ​​or coefficients exist. If Such event classes are present, the determination unit is particularly for determining a plurality of association probabilities for each of these event classes formed. The selection unit is then such designed such that it for those event classes that two or more sets have representative of the same event class comparison values, Averages of the corresponding association probabilities and forms when extracting the highest probability of belonging to Averages instead of individual values ​​draws on. With this development, the Sensitivity of the device can be reduced to signal noise. The may signal noise particular to classification errors and consequently result in a lower classification accuracy when representing a certain variation in the same physiological event is physiological signals available. With the mentioned Development of the device according to the invention can be such Variations by verschiedne sets of comparative values ​​into account, whereupon the possible with a non-noisy signal classification accuracy largely intact even when such signals remains whose Noise level without considering the variations the classification accuracy would be significantly reduced.

In the apparatus according to the invention, the transformation unit is a Adjustment unit for centering the physiological signal in a Time window of predetermined window width and outputting the centered be preceded by physiological signal to the transformation unit to a single input format for the existing physiological signals to achieve. The aforementioned variations in the same se physiological signals can thereby eg. in the form of offsets in the of centering adjustment unit made occur. These offsets would if they were not included, the sensitivity of the device against signal noise increase and thus the classification accuracy to reduce. In order to consider the offsets, it is therefore advantageous in those event classes that two or more Sets of the same event class representative comparison values comprise sets of comparative values ​​different offsets in the Centering of the centered physiological signal correspond.

The inventive apparatus, both in the form of hardware as can also be implemented in form of software, is suitable in particular for use in an implantable medical device, such as in a pacemaker or defibrillator.

Further features, properties and advantages of the device according to the invention result from the following description of an embodiment with reference to the accompanying drawings.<dl tsize="6"><dt>Fig. 1</dt><dd>shows an embodiment for the inventive device for the classification of physiological events.</dd><dt>FIG. 2</dt><dd>shows a pair of test signals, based on which the operation of the A device according to the invention will be explained.</dd><dt>Fig. 3</dt><dd>shows the coefficients obtained by wavelet transform of the two test signals shown in Fig. 2.</dd><dt>Fig. 4</dt><dd>shows the effect that an offset in the centering of the two Test signals has a time window on the classification result.</dd><dt>Fig. 5</dt><dd>shows the effect that an offset in the centering of the two Test signals in a time slot on the classification result has, when offset by two or more sets of the same Event class representing comparison coefficient consideration find.</dd><dt>Fig. 6</dt><dd>shows the error rate in the classification for different when Centering occurs in the time frame offsets depending on the Signal noise.</dd></dl>

Fig. 1 shows an embodiment for the inventive device. The core of the Erfindungemäßen device forms a classification unit 1, a transformation unit 3 as well as a probabilistic neural Network 5 comprises the reception of the transformation unit 3 Coefficients (ie values), which in the transformation unit 3 incoming and possibly processed physiological signal representing, connected is. The transformation unit 3 is a signal processing unit 20 upstream of which in the present embodiment an anti-aliasing filter 22, a broadband analog-to-digital converter 24, hereinafter referred to A / D converter, called a detecting stage 26 for detecting a physiological event, and a combined adjusting / standardizing 28 comprises, in from an incoming physiological input signal A this dance sequence are run through, and the one for outputting processed physiologic signal to the transformation unit connected is.

Below is the breakdown on which the signal processing unit 20 Signal processing of the physiological input signal A, the in present embodiment is an intracardiac electrogram (IEGM), take as example. by means of a cardiac pacemaker is described briefly. However, it should be noted that the present invention with the classifiable physiological signals not on intracardiac EGMs are limited.

In the anti-alias filter 22 filtering the IEGM by means of an anti-aliasing low-pass filter and a suitable amplifying and / or scaling the IEGM. As for systems of sampled data (sampled data systems) known is suppressed, the filter signal components at frequencies above can half the sampling occur and the subsequent signal processing steps be impressed. To ensure the accuracy and to obtain morphology of the waveform of the IEGMs, takes place in addition no further filtering.

The filtered IEGM is of anti-alias filter 22 to the A / D converter 24 passed, the a conventional analog-digital converter with a stepwise linear relationship between the input signal and the Output is. The sampling rate and the resolution of the output signal are adapted to the requirements of the classification process. in the Generally they are at 1024 Hz or below or at 8 bits or more. Depending on the requirements, various A / D converter architectures be used, including the so-called. "A bit designs" (One-bit design). In special cases where input signals with present a wide dynamic range, the use of non-linear (Companding, ie the signal be compressed and subsequently re- expanding) A / D converter structures may be advantageous. The converted IEGM is from the A / D converter 24 as an output signal to the detection stage 26 and forwarded to the adjusting / standardizing 28th

In the detection stage 26 the detection of an event takes place on the basis of an event to rate adaptive event threshold consideration. The Result of the detection indicates a type of the waveform of the input signal at. When the detection stage 26 detects an event, it is a Trigger signal (trigger signal) to the adjusting / standardizing 28 from which a Adjust and / or normalizing the physiological signal triggers.

When the adjusting / standardizing 28 a trigger signal from the detection stage 26 receives, so the underlying IEGM is in an event window a predetermined window width, which is generally 64 sampling steps is recognized and centered in the window. The window is expected to Event Type adjusted. It also carried out a determination of the time interval the last detected event to the current event and a Normalization of the waveform on a standardized peak-to-peak amplitude based on a normalization factor to a normalized Event signal to obtain. is the time interval and the normalization factor the adjusting / standardizing 28 to the probabilistic neural network 5 further, whereas they centered the window and normalized event signal passes on transformation unit. 3

The transforming unit 3 receives a wavelet transform of the centered and the normalized signal before event that a number of the signal has coefficients representing the result. Wavelet transformation is a well known method for compact Show any signals. The transformation of a signal by means done of reference wavelets and a calculation method specifies how the Reference wavelets are offset against the signal. The details of the Transformation can within the calculation by the environment given mathematical limits are chosen such that they themselves can use certain signal classes very effective. In the present Embodiment, the implantable for use in a medical device is intended, allows the wavelet transform, an event window with a window size of 64 Scanning steps (64 coefficients DWT) with less than 16 wavelet transform coefficients display and simultaneously enough to obtain information of the signal to a reliable event classification in probabilistic neural network 5 to ensure.

For performing the wavelet transformation comprises the transformation unit 3 a wavelet store 6, in which the stored reference wavelets are, and a computing unit 4 to the adjusting / standardizing 28 Receiving the centered and normalized in the window event signal and with is the wavelet store 6 connected to receive the reference wavelets. In the computing unit 4 is calculating the coefficients, that is actual wavelet transform instead.

There are a number calculation methods of calculating Wavelet transform are suitable. Likewise, a variety exists appropriate Reference wavelets. For the calculation of the wavelet transformation in the computing unit 4, the rate used reference wavelets eg. in Terms are selected on the achievable computing power. When you select the calculation method and the reference wavelets however preferably to make sure that when calculating the wavelet transform in the computing unit 4 is the same calculation methods and the same set of reference wavelets are used as the Calculating the comparison coefficients (see below) for the application arrived.

The computing unit 4 outputs the result of the wavelet transform, which is the Wavelet transform coefficients as a set of coefficients to the probabilistic neural network 5 (hereinafter abbreviated PNN) from.

The PNN 5 includes a PNN structure 8, an input layer 7 (input layer), and an output layer or summation unit 9 (the output layer). The PNN structure 8 has a number of internal nodes and is connected to the Input layer 7 (input layer), which has a number of input nodes, and the summation unit 9, comprising a number output node, connected.

The internal nodes of the PNN structure 8 each contain a certain Coefficient vector, the comparison coefficients as a set of comparative values, a specific comparison time interval and a certain Comparison normalization factor contains and a certain class of Events characterized. In the present embodiment includes the PNN structure 8 for each class exactly one node, but it is also possible a class more internal nodes easily with each assign different coefficient vectors, so that a class of a node cluster is represented. The coefficient vectors usually in advance from a plurality of signals to the for Event class typical waveform extracted.

The object of the input layer 7 of PNN 5 it is, from the transformation unit 3, the coefficients and of the adjusting / standardizing 28 the time interval and to receive the normalization factor of the current IEGM and evenly spread over the internal nodes of the PNN structure. 8

In the internal nodes, the respective coefficient vectors a Signal vector received from the transformation unit 3 Coefficient, as well as the time interval and the normalization factor of the current IEGM is formed, compared by the difference signal vector and coefficient vector is formed. In addition, the determined vector differences associated likelihood values, wherein the probability value is the greater, the smaller the difference vector fails. Determining the probability values, a Gaussian transfer function with selectable standard deviation sigma (the distance of the Turning points of the curve from the curve center indicates) include. The Arbitrary standard deviation allows the boundaries, ie, the maximum permissible deviation from the respective Standardsignaiform a class define.

The PNN structure 8 is to pass on the signal vector and the for Signal vector determined probability values ​​to the summation unit 9.

The summation unit 9 has exactly one output node for each Event class, for their recognition apparatus of the invention is configured. The output node receives the respective for Event class determined probability value of the signal vector. If an event class associated with a plurality of internal nodes, receives the corresponding output node all probability values ​​of these internal Node and calculates the average value of the corresponding probability values. In both cases, the probability value of an output node the summation unit 9, the probability is that the the IEGM or the triggering event by the output node represented class belongs. The class which in the summation unit 9 having the highest probability value, which is the IEGM associated triggering event, provided this probability value a exceeds classification threshold. If it exceeds the threshold classification not, the event is classified as unknown and possibly to used to trigger an adjustment of the PNN structure for detecting a new event class results. The summation unit 9 are finally the Signal vector and the event class, which it has been assigned, as Result from the classification.

As stated above, 8 contains a node of a structure PNN given coefficient vector, which represented by the node Event class characterized. Since this coefficient vectors a Event class advance from a representative set of signals on the corresponding event class decline, have been extracted, is not to be expected that such a coefficient vector to a single classifying input signal accurately reflects. The same method was the Classifying, in the inventive device on the basis of similarity between the returning signal to the input signal vector in which the plug-in end in the input signal information is encoded, and the coefficient vectors the knot.

The plug end of the input signal information is after conditioning of Input signal in the signal conditioning unit 20 in the transformation unit 3 encoded by the wavelet transform. Wavelet transformation but is not invariant with respect to a time Shift of the input signal in the signal window, ie the result of the Transformation will change when the maximum of the amplitude by one or several sampling steps in the signal window forward or is redeployed. As a result, the values ​​of the transformation unit of issued coefficients essential for the classification of are important fluctuate. The extent of the fluctuation will depend on the Centering accuracy of the input signal in the signal window. Although the direct impact of these fluctuations on the classification process in is generally low, significant classification errors can occur when noisy input signals are present.

Subsequently, a device of the invention by means of which is carried out Classification on the basis of figures 2-6, the intermediate steps in the representing the end of the classification, with reference to two exemplary verschiedner Test input signals explained. The test signals are used on the one hand a so-called sinusoidal. Haversine signal and on the other hand a triangle signal. there in particular the concepts for reducing the influence of Factors that worsen the classification accuracy, explained.

In Figure 2 are the Haversine signal and the triangular signal, the erfindüngsgemäßen with the Device are to be distinguished, as a function of Time shown. Both signals, the signal duration is 40 ms, the peak amplitude 1 V and the sampling rate 1024 per second. The basic form of two signals similar to conventional waveforms intracardiac EGMs.

Figure 3 shows the lower-order coefficients, which for the Haversine signal and the triangle signal in the transformation unit 3 by means of the wavelet transformation were recovered. It took a Daubechies-4-transformation as a wavelet transform using, each at a Haversine signal or triangular signal is applied, the above in a has been centered Signal window with a width of 64 sampling steps. indeed are the similarities between the respective coefficients large, but are determine adequate Distinguish between the coefficients. Other Wavelet transforms give similar results as the Daubechies-4-transformation.

Fig. Figure 4 shows both the Haversine signal and the triangular signal, the Impact that different offsets (measured in sampling steps) when Centering of the signal in the signal window in a node of the PNN structure 8 with a Haversine signal representing coefficient vector (ie, the node is configured to represent an event class to a Haversine signal leads) leave.

Given a perfect centering of the signal in the signal window, ie The case of an offset with the value zero, are the nodes of the Haversine signal represents, for Harversine signal as an input signal a Output value 1 and for the triangular signal as an input signal a Output value of about 0.25. On the other hand, an offset to a Sampling in the centering of the signal or the Haversine Triangular signal in the signal window exists (offset +1 or -1), thus reducing the output value for the Haversine signal as an input signal to approximately 0.55 and for the triangular signal as an input signal to about 0.18.

The difference between the two in Fig. 4 output values ​​shown forms the basis of the event classification using the probabilistic neural network 5. It guarantees a large difference (as in Offset 0) low susceptibility of the classification process towards Signal noise. The reduction of the difference due to the offset of +1 or -1 (measured in sampling steps) to about half of the value at an offset of zero means that in the case of the offset of +1 or -1 Classification errors already occur in noise levels, the half that are, as the noise level, in which classification error at an offset occur from zero. This may in some applications to cause significant reduction in the performance of the device.

As mentioned, it is possible in the PNN structure 8 of an event class more internal nodes, each with slightly different coefficient vectors to you, so that an event class of a node cluster represented. The node cluster can now be configured such that an event class two or more nodes are associated, each contain a coefficient vector comprising a (preferably small) Offset corresponds to the expected input signal. The coefficient vectors are usually in advance from a plurality of signals to the for Event class typical waveform extracted, said for the signals different nodes of a node cluster (easy) to today time lag , ie have an offset. The size of the time offset can be in With regard to the expected in the input signal offset, that is to expected accuracy in the centering of the input signal Signal window, are selected.

As a result, such a node cluster provides a widened portion, in which a large difference between its initial value at a Haversine signal as an input signal and its output value at a Triangle signal as an input signal is present (Fig. 5). Thereby, a high Classification accuracy even at noise levels that otherwise the classification accuracy would significantly affect, be maintained. The behavior of the error rate in the classification in a Haversine signal 6 representing the node cluster is shown in Fig. For various in Centering occurs in the time frame offsets depending on the Signal noise shown.