DE4307545A1 - Device and method for determining the location and/or the extent of ischemias and/or infarcts in the heart of a patient - Google Patents

Abstract

A device for determining the location and/or the extent of ischemias and/or infarcts comprises a multichannel measuring system (2, 4) for measuring electric and/or magnetic field quantities (variables) which are generated by a cardiac activity during at least a portion of a cardiac cycle at a number of measuring points. Connected to the multichannel measuring system (2, 4) is a classifier (18) which is capable of learning and whose inputs can be fed input values which correspond to the measured values measured during the at least one portion of the cardiac cycle. The classifier (18) which is capable of learning is connected to a visual means (22) of representation which serves to display the location and/or the extent of the ischemia and/or of the infarct as a function of localisation data output at the outputs (20) of the classifier (18) which is capable of learning. Likewise claimed is a method for operating the device for locating ischemias and/or infarcts. <IMAGE>

Description

A device and a method for determining the location and / or the expansion of ischemia and / or infarcts in the Heart of a living being is known from the article by R. S. Mac Leod / M. J. Gardner / B. M. Horacek with the title: "High-resolution ECG mapping procedures for examinations during coronary dilation ", appeared in Biomedical Technik 35 (supplementary volume 1990), pages 236 to 237. There becomes the electrical caused by cardiac activity Potential with the help of elec trodes measured and in a reconstruction the Potentials on the heart surface (epicardial potentials) calculated. The distribution and the time course of the epicardial potentials allow conclusions to be drawn about the location and the expansion of damaged ischemic heart weave. However, the model of epicardial potential leaves important electrophysiological properties, such as. B. the endocardial potential as well as the spread of excitation in the heart, except eight.

From a measured with a biomagnetic measuring system Magnetocardiogram (MKG) can be with the model of the Infarct dipoles, as described in the article by M. Saarinen / T. Katile / J. Nenonen / M. Seppaenen / P. Siltenen with the title: "Simulated and measured magnetocardiograms in localized inyocardial abnormalities "appeared in" Biomagnetism: Applications and Theory, edited by H. Weinberg and T. Katila, Pergamon Press, New York, 1985, pages 59 to 63, generally reconstruct a single place. statement About ischemia occurring simultaneously in several places  and about infarcts with speckled geometry are with a reasonable computing effort is no longer possible.

When determining the location and / or the extent of Ischemia or infarction gives rise to the problem that one multiple source models, such as the single dipole, the electric Do not reproduce the physiology of the heart with sufficient accuracy and thus only valid statements about the place and give the expansion. Desirable more complex rakes however, models are so time-consuming that they are one Obstacles to application in clinical practice.

From the article by U. R. Abeyratne / Y. Kinouchi / H. Oki / J. Okada / F. Shichÿo / K. Matsumoto with the title: "Artificial Neural Networks for Source Localization in the Human Brain ", published in Brain Topography, Vol.4, No.1, 1991, pp. 3 to 21, is known for the localization of single di Poland has an adaptable classifier in the form of an ent to use a speech-trained neural network. The neural networks become those of a multi-channel measurement system fields of an electrophysiology measured at a point in time activity. From this the neural network forms six output variables that correspond to the parameters of a dipole speak. However, the neural network uses sources localization only at a single time measured information. Ischemia or infarctions can be based on this localization process not be vote.

The invention is based on the object, a device and specify a procedure by means of which the location and / or the extent of ischemia and / or infarcts from electrical and / or magnetocardiographic measurements can be determined.

The first-mentioned task is performed by a device with the Merk paint the claim 1 solved. The second task  is by a method having the features of claim 15 solved.

The use of a learnable and accordingly trai nated classifier allows a complex heart model to use that the electrophysiological properties accurately reproduced by ischemia and / or infarction. Of the a classifier capable of learning can be trained with the model the, whereby the computing effort required for training only must be operated once and be correspondingly high may. On the other hand, the classifica once trained enables the detection and localization of the damaged heart webes from the measured field sizes in a short time. With in other words: despite a good picture of reality The complex model can be built in an acceptable time Detect and localize pathological tissue.

An advantageous embodiment is characterized by that the number of inputs the total of the abge the measured values correspond to a cardiac cycle. All in the Existing information can thus be measured without Data compression can be evaluated.

Another embodiment is characterized in that the cardiac cycle is divided into up to 200 sampling times is. This can be used to localize ischemia and / or infarcts sufficient information from the time the course of the measured values can be obtained.

Another embodiment is characterized in that the distance between the sampling times in the QRS complex of the heart cycle is less than in the rest of the cardiac cycle. This scan adapted to the cardiac cycle produces less redundant information as an unmatched scan at equal intervals.  

Depending on the requirements for the accuracy of the Detection of pathological tissue is the number of Measuring points between approximately 20 and 256. One for normal Adequate accuracy will be applied in a further advantageous embodiment achieved in that the The number of measuring points is 30 to 64.

In a further advantageous embodiment, everyone Output assigned to an area of the heart model. Through the A summary into larger areas results in a favorable structure for the classifier.

Another advantageous embodiment stands out from that the classifier is trained so that the Values at the outputs are a measure of the expansion of the Ischemia and / or the infarction around the centers of the Represents corresponding areas.

In a further advantageous embodiment, the location and / or the expansion is shown graphically.

In a further advantageous embodiment, there is Heart model from an arrangement corresponding to the physiology Model heart cells, which electrophysiological Cardiac parameters are assigned, with three in practice Parameters, namely excitation propagation speed, Refractory period and form of action potential are sufficient for Modeling pathological tissue.

The invention is described below with reference to two figures explained. It shows:

Fig. 1 shows the structure of a device for determining the location and / or the extent of ischemia and / or infarction and

Fig. 2 shows a variant of the device of FIG. 1 with a data compression level.

In Fig. 1, 2 as a multichannel measuring system symbolizes a Magnetokar diogram device (MKG device), as is known from the basic structure of US Pat. No. 5,152,288. The MKG device disclosed therein comprises a 12-channel measuring arrangement with which a magnetic field generated by an electrophysiological activity can be measured contactlessly and simultaneously at 12 measuring points. Since ischemic or infarcted areas and not just a single point in the heart are to be localized, at least about 20 measuring channels are required. Such a device is described in the Schneider et al. with the title U. Multichannel Biomagnetic System for Study of Electrical Activity in the Brain and Heart "published in Radiology, Vol. 176, No. 3, Sept. 1990, pp. 825-830. Depending on the achievable in the evaluation Accuracy increases the number of measuring channels up to 256. A good compromise between the expenditure on equipment and the accuracy is around 30 to 60 measuring channels.

Potentials generated by the cardiac activity on the body surface can also be measured with an electrocardiography device 4 (EKG device) as a multi-channel measuring system. However, the EKG device 4 must include an increased number of measuring channels compared to a standard EKG device with 12 channels. The number of measuring channels required is in the same range as for the MKG device 2 . The MKG device 2 and / or the EKG device 4 are connected to a signal processing circuit 6 . The signal processing circuit 6 serves to adapt and process the measured values of the MKG device 2 and the EKG device 4 . The Signalverar processing circuit 6 comprises in each channel an amplifier, a bandpass filter, a scanner 7 and a subsequent digitization stage. If necessary, the measured values are also averaged over several cardiac cycles.

An evaluation unit 8 is connected to the outputs of the signal processing circuit 6 . The evaluation unit 8 comprises a shift register 10 , the inputs 12 of which are connected to the signal processing circuit 6 . The shift register 10 is used for the intermediate storage of the serially sampled measured values from at least one cardiac cycle. The number of shift register inputs 12 is identical to the number of measuring points or measuring channels. The length of the shift register 10 is sufficient to be able to temporarily store the sample values of at least part of a cardiac cycle, but preferably a complete cardiac cycle. At the end of the cardiac cycle, all sampled measured values are simultaneously available for output from outputs 14 of the shift register 10 . The number of shift register outputs 14 thus corresponds to the total number of measured values that are recorded during a cardiac cycle.

The shift register outputs 14 are connected to inputs 16 of an adaptive classifier 18 designed as a neural network, which also belongs to the evaluation unit 8 .

A special hardware on which a suitable neural Network is executable is in the article by Ramacher, U .; Raab, W .; Anlauf, J .; Hachmann, U .; Wesseling, M. with the Title "Synapse X: A general purpose neurocomputer", he appeared in proc. II Int. Conf. on Micro Electr. for neural Networks, 1991.

Since not all measured values are correlated with one another in time, a partially networked neural network 18 is set here. At the outputs 20 of the neural network 18, the values are output that are a measure of the location and extent of ischemia and / or infarcts. Depending on the desired resolution in the localization, the neural network comprises 10 to 200 outputs. The outputs 20 of the neural network 18 are connected to visual display means 22 , which comprises a graphics unit 24 . The graphics unit 24 prepares the values output at the outputs 20 in order to be able to give the viewer a good idea in a graphical representation of the position and / or extension of the damaged area on a screen 26 . At the same time, the measured values can be recorded in a log 28 .

An area of the heart model is assigned to each output 20 . The localization data output at the outputs 20 are a measure of the extent of a damaged area. For example, a "zero" at an exit 20 means that only healthy tissue is present in the corresponding area, while a "one" means that the entire area is damaged. Intermediate values indicate that corresponding smaller areas are damaged.

In one variant, a data compression stage 29 is connected upstream of the neural network 18 in FIG . The number of inputs 16 can thus be reduced, which in turn can have a favorable effect on the training of the neural network. The data compression level also allows the number of patterns required for training to be significantly reduced because the number of parameters to be determined during training, namely the number of neurons and weights, is smaller. The data compression stage 29 works z. B. by the method of main component generation, the measured values of all channels present at a time being considered as an n-dimensional vector, n being the number of measuring channels.

In addition to compression by filtering and compression with regard to the time course of the measured values, a comm pression regarding the number of time to be processed rows or measuring channels possible. That is a good thing methods of main component decomposition mentioned above or a neural compression network, as in the German patent application P 42 07 595.5 is.  

The neural network 18 is only capable accordingly when it has been trained with a large number of field patterns which are characteristic of the location and extent of the ischemia and / or the infarction. A heart model can be used for this, as described in the article by R. Killmann, P .; Wach F .; Dienstl, entitled: "Three-dimensional computer model of the entire human heart for simulation of reentry and tachycardia: gap phenomeon and Wolff-Parkison-White syndrome" Basic Research in Cardiology 86, 1991, pp. 485-501. In the heart model, the heart is divided into volume cells that are related according to the cardiac physiology. Electrophysiological parameters are assigned to each cell. Here the parameters of excitation propagation speed, refractory time and action potential form are assigned to the cells. They can be freely chosen within the framework of the physiologically sensible. The cells are approximately 2.5 mm wide. Starting from an excitation at the sinus node, the individual volume cells are then activated during the excitation propagation in accordance with the electrophysiological parameters assigned to them. The excitation propagation is accompanied by an electric and magnetic field, which can then be calculated at the corresponding model locations. Noise can be superimposed on the calculated values for training purposes.

To describe the ischemic and / or infarcted The heart model is divided into 10 to 200 areas divides, assuming a damaged area in the center of the range with a corresponding off stretch is arranged.

The weight factors assigned to the individual neurons of the neural network 18 are determined in training. At the beginning of the training, all weights are assigned random values. With some network types, it is possible to introduce knowledge about the modeled process, here the determination of location and extent from the measured values, by pre-assigning the weights before the training. This may have a favorable effect on the time required for training or on the stability of the training process, especially when a large number of weights have to be occupied. The pre-assignment can e.g. B. via K-means clustering or K-nearest neighbor clustering.

For training, the electrophysiological parameters are now varied in a random and random order in accordance with the model ischemia and / or the model infarction in the heart model. The training can also take place after the permutation has been carried out, ie all training values are offered to the network 18 in a random order in one epoch, or the batch learning method, ie all training values are offered to the network 18 in a given order in one epoch . The section from the training phase is defined here as the epoch, in which the neural network 18 is offered all learning patterns once for training. The advantage of stochastic learning is that the network 18 does not learn the order of the training data. An additional advantage of permutation is that the network sees all training data equally often - in every epoch. With permutation, training also converges faster than with batch learning.

In Fig. 1 and 2, the training phase blocks by the multifunction 30, 32 and 34 symbolizes. The function block 30 represents the heart model indicated above with the division into volume elements. The function block 32 varies the location and the extent of the model ischemia or / or model infarction. In the arithmetic unit 34 , a magnetic and / or electric field generated by the set heart model is now calculated at the model measuring points and fed to the inputs 16 of the neural network 18 via the shift register. In the training phase, the outputs 20 are supplied with the location and the extent of the model ischemia and / or the model infarction on which the computing values are based. The weights associated with the hidden neurons are set therefrom via an internal feedback in the neural network 18 .

The training phase of the neural network can take up to several hours or days of computing time, depending on the complexity of the heart model. However, once the neural network 18 has been trained, ie the weights have been determined, the neural network 18 can output the corresponding output data in a short time on the basis of the input data.

Claims (27)

1. Device for determining the location and / or the extent of ischemia and / or infarcts in the heart of a living being with

• a multichannel measuring system ( 2 , 4 ) for measuring electrical and / or magnetic field quantities generated by cardiac activity during at least part of a cardiac cycle at a number of measuring points,
• - A with the multichannel measuring system ( 2 , 4 ) connected trained learnable classifier ( 18 ), the inputs ( 16 ) of which input values can be supplied which correspond to the measured values measured during the at least part of the cardiac cycle, with the learnable classifier ( 18 ) by Feeding of training data corresponding to the field sizes at the measuring locations to which the inputs ( 16 ) are trained, which training data are calculated with the aid of a cardiac model which shows the course of excitation of the heart during the cardiac cycle from varied locations and varied extents of model ischemia and / or replicates model infarcts, and wherein at the outputs of the classifier ( 18 ) capable of learning, localization data can be output which correspond to the location and / or the extent of the ischemia and / or the infarction, and
• - A visual representation means ( 21 ) connected to the classifier ( 18 ) capable of learning, for displaying the location and / or the extent of the ischemia and / or the infarct depending on the localization data.

2. Device according to claim 1, characterized records that the at least part of the Cardiac cycle includes the QRS complex and the ST segment. 3. Apparatus according to claim 1 or 2, characterized in that the number of inputs ( 16 ) corresponds to the number of total measured values of a cardiac cycle. 4. Apparatus according to claim 1 or 2, characterized in that a data compression stage ( 29 ) is arranged in front of the learnable classifier ( 18 ) and that the number of inputs ( 16 ) is less than the total number of measured values sampled. 5. The method according to claim 1 to 4, characterized in that a scanner ( 7 ) between the multi-channel measuring system ( 2 , 4 ) and the neural network ( 18 ) is arranged, which simultaneously measures the measured values measured during the part of the cardiac cycle in all channels scans at up to 200 sampling times. 6. Apparatus according to claim 5, characterized in that the scanner ( 7 ) is suitable to choose the distance between the sampling times in the QRS complex of the cardiac cycle less than in the rest of the cardiac cycle. 7. Device according to one of claims 1 to 6, characterized characterized in that the number of measuring points is approximately 20 to 256. 8. Apparatus according to claim 7, characterized records that the number of measuring points is approximately 30 to 64. 9. Device according to one of claims 1 to 8, characterized in that a region of the heart model is assigned to each output ( 20 ). 10. Apparatus according to claim 9, characterized in that the adaptable classifier ( 18 ) is trained so that the localization data represent a measure of the extent of the ischemia and / or the infarcts around the centers of the areas corresponding to the exits ( 20 ). 11. Device according to one of claims 1 to 10, characterized in that the visual representation means ( 22 ) is suitable for graphically representing the location and / or the extent. 12. Device according to one of claims 1 to 11, there characterized by that the heart model from an arrangement corresponding to the physiology of model heart cells, which are electrophysiological Heart parameters are assigned. 13. Apparatus according to claim 12, characterized ge indicates that the electrophysiological Parameters the rate of excitation propagation, the Refractory period and the action potential form include. 14. Device according to one of claims 1 to 13, characterized in that the learnable classifier ( 18 ) is formed as a neural network. 15. A method for determining the location and / or extent of ischemia and / or infarction in the heart of a living being, comprising the steps of:

• - Measuring electrical and / or magnetic field quantities generated by a cardiac activity during at least part of a cardiac cycle at a number of measuring points with a multi-channel measuring system ( 2 , 4 ),
• - Feeding input values, which correspond to the measured values measured during at least part of a cardiac cycle, to inputs of a trained, learnable classifier ( 18 ), the learnable classifier ( 18 ) by supplying training data that match the field sizes to the Measuring locations correspond to which the entrances are trained, which training data have been calculated with the aid of a heart model which simulates the course of excitation of the heart during the cardiac cycle from varied locations and varying extents of model ischemia and / or model infarcts, and at the outputs of the learnable classifier ( 18 ) localization data are output which correspond to the location and / or the extent of the ischemia and / or the infarction, and
• - Displaying the location and / or the extent of the ischemia and / or the infarction by means of visual representation ( 22 ) depending on the localization data.

16. The method according to claim 15, characterized ge indicates that the at least part of the cardiac cycle includes the QRS complex and the ST segment. 17. The method according to claim 15 or 16, characterized in that the number of inputs ( 16 ) corresponds to the number of total sampled measured values of a cardiac cycle. 18. The method according to claim 15 or 16, characterized characterized in that the measured values of a Undergo data compression. 19. The method according to claim 15 or 18, characterized characterized that the cardiac cycle in up is divided into 200 sampling times. 20. The method according to claim 19, characterized ge indicates that the distance between the samples times in the QRS complex of the cardiac cycle is less than in the rest of the cardiac cycle. 21. The method according to any one of claims 15 to 20, there characterized in that the number the measuring points are 20 to 256.   22. The method according to claim 21, characterized ge indicates that the number of measuring points is 30 is up to 64. 23. The method according to any one of claims 15 to 22, characterized in that the heart model is divided into areas and that each area is assigned an output ( 20 ). 24. The method according to claim 23, characterized in that the adaptable classifier ( 18 ) is trained so that the localization data represents a measure of the extent of the ischemia and / or the infarcts around the centers of the areas corresponding to the exits ( 20 ) . 25. The method according to any one of claims 15 to 24, there characterized by that the place and / or the extent can be graphically represented. 26. The method according to any one of claims 15 to 25, there characterized by that the heart model from an arrangement corresponding to the physiology of model heart cells, which are electrophysiological Heart parameters are assigned. 27. The method according to claim 26, characterized ge indicates that the electrophysiological Parameters the rate of excitation propagation, the Refractory period and the action potential form include.