WO2001085018A2 - Procede et dispositif de classification de valeurs de serie, support d'enregistrement lisible par ordinateur, et element de programme informatique - Google Patents

Procede et dispositif de classification de valeurs de serie, support d'enregistrement lisible par ordinateur, et element de programme informatique Download PDF

Info

Publication number
WO2001085018A2
WO2001085018A2 PCT/DE2001/001745 DE0101745W WO0185018A2 WO 2001085018 A2 WO2001085018 A2 WO 2001085018A2 DE 0101745 W DE0101745 W DE 0101745W WO 0185018 A2 WO0185018 A2 WO 0185018A2
Authority
WO
WIPO (PCT)
Prior art keywords
time series
series values
sequence
values
type
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/DE2001/001745
Other languages
German (de)
English (en)
Other versions
WO2001085018A3 (fr
Inventor
Gustavo Deco
Bernd SCHÜRMANN
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Siemens AG
Siemens Corp
Original Assignee
Siemens AG
Siemens Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Siemens AG, Siemens Corp filed Critical Siemens AG
Publication of WO2001085018A2 publication Critical patent/WO2001085018A2/fr
Anticipated expiration legal-status Critical
Publication of WO2001085018A3 publication Critical patent/WO2001085018A3/fr
Ceased legal-status Critical Current

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A61B5/346—Analysis of electrocardiograms
    • A61B5/349—Detecting specific parameters of the electrocardiograph cycle
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition
    • G06V30/19—Recognition using electronic means
    • G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
    • G06V30/194—References adjustable by an adaptive method, e.g. learning

Definitions

  • the invention relates to a driving and a device for classifying time series values, a computer-readable storage medium and a computer program element.
  • [1] describes generating a feature vector for each pixel of a texture image in the context of a texture analysis with neural networks, which is then classified using a neural classifier.
  • a neural network with a feed-forward structure is used as a neural classifier, hereinafter referred to as a feed-forward neural network, which in [1] eight
  • a disadvantage of such a feed-forward neural network is, in particular, that a very large number of training data is required in the context of the classification of time series values into time series values of a system of a first type or into time series values of a system of a second type.
  • Dynamic of a system is further understood to mean that in the event that a system can be described with a deterministic description of the future system behavior based on the description of the system behavior in the past.
  • Time series values are, for example, samples of analog signals, for example of biomedical signals such as an electroencephalogram signal (EEG signal), an electrocardiogram signal (EKG signal), electromyogram signal, generally all analog bioelectrical or biomagnetic signals, such as, for example can be found in an overview in [2].
  • biomedical signals such as an electroencephalogram signal (EEG signal), an electrocardiogram signal (EKG signal), electromyogram signal, generally all analog bioelectrical or biomagnetic signals, such as, for example can be found in an overview in [2].
  • a so-called electrocardiogram device is also known from [2], with which EKG signals can be recorded.
  • BTT Backpropagation Through Time learning process for a recurrent neural network.
  • the neural network is trained with time series values as training data, the influence of temporally preceding time series values decreasing exponentially in a current adaptation step of the weights of the recurrent neural network.
  • the invention is therefore based on the problem of classifying time series values of a system into time series values of a system of a first type or into time series values of a system of a second type, the system of the first system being a system which has a predetermined dynamic range and the system of the second type is a system that has no dynamics.
  • a method for classifying time series values into time series values of a system of a first type or into time series values of a system into time series values of a second type, wherein a system of a first type is a system which has a predetermined dynamic range, and wherein one type of a second system is a system that has no dynamics, has the following method steps.
  • the time series values that describe a system are fed to a competitor artificial neural network.
  • a sequence of initial recurrence values is generated by the recurrent artificial neural network based on the time series values.
  • a sequence of recurrence training errors is determined from the sequence of the recurrence output values.
  • the consequence of recurring training mistakes is used as the basis for one Classification of the time series values is used, the time series values being classified into time series values of a system of the first type if the sequence of recurrence training errors fulfills a predetermined criterion and classifies the time series values into time series values of a system of the second type if the sequence of recurrence training errors does not meet the specified criterion.
  • a device for classifying time series values into time series values of a system of a first type or into time series values of a system of a second type, wherein a system of the first type is a system which has a predetermined dynamic range, and wherein a system of the second Type is a system that has no dynamics, has a processor that is set up in such a way that the following method steps can be carried out:
  • time series values that describe a system are fed to a recurrent artificial neural network, a sequence of output recurrence values is generated from the recurrent artificial neural network based on the time series values,
  • a sequence of recurrence training errors is determined from the sequence of recurrence starting values, • from the sequence of recurrence training errors, the time series values are classified into time series values of a system of the first type if the sequence of recurrence training errors is a predetermined one Criterion fulfilled, and
  • time series values are classified into time series values of a system of the second type if the sequence of recurrence training errors does not meet the specified criterion.
  • a computer program for classifying time series values into time series values of a system of a first type or into time series values of a system of a second type is in a computer-readable storage medium stored, wherein a system of the first type is a system which has a predetermined dynamic range, and wherein a system of the second type is a system which has no dynamic range, which, when executed by a processor, has the following method steps:
  • time series values that describe a system are fed to a recurrent artificial neural network
  • a sequence of output recurrence values is generated by the recurrent artificial neural network on the basis of the time series values
  • a sequence of recurrence training errors is determined from the sequence of the recurrence output values
  • the time series values are classified into time series values of a system of the first type if the sequence of recurrence training errors fulfills a predetermined criterion, and
  • time series values are classified into time series values of a system of the second type if the sequence of recurrence training errors does not meet the specified criterion.
  • a sequence of recurrence output values is generated from the recurrent artificial neural network on the basis of the time series values, a sequence of recurrence training errors is determined from the sequence of the recurrence output values, • from the sequence of recurrence training errors, the time series values are classified into time series values of a system of the first type if the sequence of recurrence training errors fulfills a predefined criterion, and • the time series values are classified into time series values
  • Systems of the second type are classified if the sequence of recurrence training errors does not meet the specified criterion.
  • This procedure clearly means that a recurrent artificial neural network is used for the first time to determine whether time series values describing a system that has a given dynamic range or whether the system described by the time series values is pure noise without any statistical correlation between the individual time series values.
  • the very different learning behavior in particular the strong reduction of the training error in the course of training a competitor artificial neural network with time series values of a system that is based on dynamics, is used to classify the system, which is described by the time series values.
  • the invention is also suitable for use in the context of financial market analysis, since it can also be determined in this area of application if there is dynamics within time series values that a forecast of the future behavior of the financial market becomes possible.
  • the invention is preferably used in the analysis of an electrocardiogram signal in such a way that it is possible, based on the initial value, to classify whether the patient whose time series values have been measured is an ischemic patient or not. This is possible because the sequence of electrocardiogram signals in an ischemic patient is based on a predetermined dynamic range, which is not the case in a healthy patient.
  • the invention is also particularly suitable for use in the classification of electroencephalogram signals with regard to the classification as to whether the time series values describe an electroencephalogram signal which indicates a brain tumor or not.
  • the sampled values can also be sampled by a so-called Licox signal, which enables the time series values to be classified as to whether sudden cardiac death is highly likely or not.
  • FIGS. 1 a and 1 b show a flow chart in which the individual method steps of an exemplary embodiment of the invention are shown;
  • Figure 2 is a sketch showing the recording of electrocardiogram signals and their further processing to classify the electrocardiogram signals
  • FIGS. 3a and 3b show a sketch of a usual period of an electrocardiogram signal (FIG. 3a) and a sequence of such periods (FIG. 3b);
  • FIG. 4 shows a sketch of a recurrent neural network which is used according to an exemplary embodiment of the invention
  • FIGS. 5a and 5b show a diagram of a sequence of recurrence training errors and a sequence of forward training errors for time series values
  • Electrocardiogram signals describe a healthy patient ( Figure 5a) as well as a diagram with a sequence of recurrence training errors and a sequence of forward training errors for time series values, the electrocardiogram signals of an ischemic
  • Electrocardiogram signals 203 are fed to an electrocardiogram device 204 via the electrodes 201 and the cable 202.
  • the electrocardiogram device 204 has the following components: a preprocessing unit 205, A memory 206,
  • a processor 207 A processor 207,
  • a signal generator 219 which are each coupled via a computer bus 208.
  • the electrocardiogram device 204 has a display unit 220 on which the electrocardiogram signals 203 and / or an alarm signal 224 explained in the following are shown.
  • the electrocardiogram signal 203 is fed to the preprocessing unit 205 in the electrocardiogram device 204.
  • the electrocardiogram signal 203 is subjected to an analog / digital conversion and noise filtering.
  • the sampled signal generated in this manner has a plurality of time series values that are stored in memory 206.
  • Logical components are further symbolically represented in processor 207. This is to be understood in such a way that the logic components are implemented by the processor 207 by means of a
  • the time series values are read out from the memory 206 and fed to a recurrent artificial neural network 209.
  • time series values 211 are fed to a forward artificial neural network 210.
  • the recurrent artificial neural network 209 forms recurrence output values 212, which are of one unit to form recurrence training errors 213.
  • Forward output values 215 are generated by the forward neural network 210 and are supplied to a unit for forming forward training errors 216.
  • Forward training errors 217 are formed by the forward training error forming unit 216.
  • the forward training errors 217 and the recurrence training errors 214 are fed to a subtractor 218.
  • Subtractor 218 formed difference signal 221 is fed to a classifier 222.
  • the classifier 222 If the difference signal 221 is greater than a predetermined value, the classifier 222 generates an alarm generation signal 223 and supplies it to a signal generator 219.
  • An alarm signal 224 is generated by the signal generator 219 and displayed on the screen 220 to a doctor, that is to say generally to a user of the electrocardiogram device 204.
  • 3a shows a typical period 300 of an electrocardiogram signal 203.
  • the period 300 of the electrocardiogram signal 203 essentially has five so-called spikes:
  • a P-wave 301 which corresponds to the electrical depolarization of the atrium of the heart
  • a Q-wave 302 which corresponds to the beginning of ventricular excitation
  • An R wave 303 which corresponds to the maximum ventricular excitation of the heart
  • An S-wave 304 which corresponds to the excitation of the basal parts of the ventricles of the heart
  • a sequence 310 of periods 300 of the electrocardiogram signal 203 result in the electrocardiogram signal 203.
  • Such a sequence 310 of periods 300 of an EKG signal 203 are shown in FIG. 3b.
  • a distance Ti is referred to below as a time series value 211, so that the sequence of distances (TI, T2, ..., Ti, ..., Tn) is referred to as a sequence of time series values.
  • the time series values 211 are stored in the memory 206.
  • step 100 the electrocardiogram signal 203 is recorded.
  • step 101 the time series values 211 (Ti) are determined.
  • the determined time series values 211 are stored in the memory 206 in a further step (step 102).
  • the stored time series values are read out (step 103) and fed to a neural network 400 shown in FIG. 4 (step 104).
  • X (t) is used to denote a time series value 211 at a time t.
  • the recurrent artificial neural network 400 has d input neurons 401.
  • Outputs of input neurons 401 are connected to inputs of hidden neurons 403 via weighted connections 402.
  • Outputs of the hidden neurons 403 are connected to an output neuron 405 via weighted connections 404.
  • An output of the output neuron 405 is connected to an input neuron 401 via a feedback 406.
  • recurrence output values are formed (step 105), which in turn are fed back via feedback 406 to the input neuron of the artificial neural network 400.
  • the formed recurrence output value 212 is further fed to the recurrence training error formation unit 213 (step 106).
  • recurrence output values are supplied to the respective input neuron 401 on the one hand, and are "decoupled” and are supplied to the recurrence training error formation unit to form a recurrence training error.
  • a recurrence output value 212 and thus also a recurrence training error 214 are formed in a further step (step 107).
  • the individual weights of the recurrent artificial neural network are updated in accordance with the variant of the BTT procedure described in [3] according to the following rules:
  • the weights of the weighted connections 404 between the hidden neurons 403 and the output neurons 405 are designated,
  • W-jk denotes the weights of the weighted connections 402 between the input neurons 401 and the hidden neurons 403,
  • a predetermined weighting value is designated by g a parameter is designated by ⁇ j which describes the exponential weighting of the evaluation function
  • Wkj or w ⁇ j can designate why the index k can run across different dimensions.
  • the index k is used at ⁇ ⁇ j to denote that neuron k which is linked by the linkage, i.e. Coupling j is coupled forward.
  • step 108 the time series values are fed to the forward artificial neural network 210.
  • the forward artificial neural network has the same structure as the recurrent artificial neural network 400, only without the feedback 406.
  • forward output values are formed by the forward artificial neural network 210 and the forward output values are supplied to the unit for forming forward training errors (step 110).
  • recurrence training errors are formed, which are fed to the subtractor 218 together with the recurrence training errors as a respective sequence of recurrence training errors or a sequence of forward training errors (step 112).
  • step 113 the subtractor 218 makes a difference between a recurrence training error in each case as a result of recurrence training errors and the temporally corresponding forward training error as a result of a forward training error, the is the difference between the respective training errors at one point in time.
  • the difference values formed are fed to the classifier 222 in a further step (step 114).
  • the classifier 222 checks whether the difference values are greater than a predetermined value (step 115).
  • the time series values are classified as time series values of a system which has no dynamics (step 116).
  • the time series values are classified as time series values of an ischemic patient, i.e. as time series values of one System that has a predetermined dynamic (step 117).
  • the classifier 222 If the time series values are classified as time series values of a system with a predetermined dynamic range (step 117), the classifier 222 generates an alarm generation signal 223 (step 118) and supplies it to the signal generator 219 (step 119).
  • the alarm signal 224 is then generated by the signal generator 219 (step 120).
  • the alarm signal 224 can either be displayed visually on the screen 220 or can also be made available to a doctor as an audio signal.
  • 5a and 5b show result diagrams 500, 510 in the following test.
  • a 12-channel electrocardiogram signal from two different patients was recorded for 24 hours. The first patient was female, 65 years old and healthy. A second patient was 66 years old, female and was myocardial.
  • the electrocardiogram signal was sampled at 500 Hz sampling frequency and encoded with 12 bit resolution. To train the neural networks, 45,000
  • the invention can also be used to check a simulation model of a technical system with regard to its quality. This use is based on the following knowledge.
  • the sequence of error signals is based on a predetermined dynamic, which can be determined using the training behavior of a competitor artificial neural network in the manner described above, the sequence of error signals in the simulation model still has statistical correlations, so that it can be concluded that this Simulation model is faulty.
  • the invention can be used in any field in which it is necessary to differentiate time series values of a system based on dynamics from time series values of a system based on no dynamics.
  • a comparison with time series values which are fed to a forward-oriented artificial neural network does not necessarily have to be used as a comparison criterion.
  • the training behavior of the recurrent neural network alone allows conclusions to be drawn about a dynamic which is the basis of the system described by the time series values or not. This clearly means that if the recurrence training error from the sequence of recurrence training errors is greatly reduced, it can be concluded that the system described by the time series values is based on a predetermined dynamic.

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Cardiology (AREA)
  • Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Multimedia (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Pathology (AREA)
  • Databases & Information Systems (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Theoretical Computer Science (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Veterinary Medicine (AREA)
  • Image Analysis (AREA)
  • Debugging And Monitoring (AREA)
  • Computer And Data Communications (AREA)

Abstract

Selon la présente invention, des valeurs de série qui décrivent un système, alimentent un réseau neuronal artificiel récurrent et ledit réseau neuronal artificiel récurrent est testé. L'apparition du signal d'erreur durant la mise en oeuvre du procédé d'essai peut être évitée afin de permettre au système, décrit par les valeurs de série, d'être classifié.
PCT/DE2001/001745 2000-05-09 2001-05-07 Procede et dispositif de classification de valeurs de serie, support d'enregistrement lisible par ordinateur, et element de programme informatique Ceased WO2001085018A2 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
DE10022481 2000-05-09
DE10022481.4 2000-05-09

Publications (2)

Publication Number Publication Date
WO2001085018A2 true WO2001085018A2 (fr) 2001-11-15
WO2001085018A3 WO2001085018A3 (fr) 2003-02-27

Family

ID=7641257

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/DE2001/001745 Ceased WO2001085018A2 (fr) 2000-05-09 2001-05-07 Procede et dispositif de classification de valeurs de serie, support d'enregistrement lisible par ordinateur, et element de programme informatique

Country Status (1)

Country Link
WO (1) WO2001085018A2 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2004030534A1 (fr) * 2002-10-02 2004-04-15 Medicale Intelligence Inc. Procede et appareil pour la detection de tendance dans un signal de surveillance d'electrocardiogramme

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH1063634A (ja) * 1996-04-05 1998-03-06 Nec Corp 時系列予測・分類のための方法及び装置
EP1027663A1 (fr) * 1997-11-04 2000-08-16 Siemens Aktiengesellschaft Procede et dispositif pour la classification d'une premiere serie chronologique et au moins d'une deuxieme serie chronologique

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
None

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2004030534A1 (fr) * 2002-10-02 2004-04-15 Medicale Intelligence Inc. Procede et appareil pour la detection de tendance dans un signal de surveillance d'electrocardiogramme

Also Published As

Publication number Publication date
WO2001085018A3 (fr) 2003-02-27

Similar Documents

Publication Publication Date Title
DE69324052T2 (de) Neuronalnetzwerk-Lernsystem
DE69331254T2 (de) Erkennungssystem zum Erkennen von Personen
DE69710502T2 (de) Gerät zur gewebestimulation
DE112018002822T5 (de) Klassifizieren neuronaler netze
EP0910023A2 (fr) Méthode et dispositif pour le modélisation neuromimétique d'un système dynamique avec un comportement non-linéaire stochastique
WO2021069338A1 (fr) Génération d'images irm du foie sans amélioration du contraste
DE4039648A1 (de) Messwertverarbeitungssystem fuer ein biologisches objekt
EP1456798A2 (fr) Evaluation d'images du cerveau obtenues par tomographie par resonance magnetique fonctionnelle
EP4348573A1 (fr) Apprentissage machine dans le domaine de la radiologie avec agent de contraste
DE102006012499A1 (de) Verfahren zum Durchführen von Regression auf Bildbasis mit Boosten
EP1519301A2 (fr) Dispositif de classification d'événements physiologiques
WO2000073980A1 (fr) Reconnaissance de formes par examen de caracteristiques complementaires apres traitement partiel
DE102023104927A1 (de) Verfahren und Vorrichtung zur Erkennung einer Anomalie im Verhalten eines dynamischen Systems
DE112020004025T5 (de) Gegnerisches und koopoeratives Nachahmungslernen zur dynamischen Behandlung
WO2023237742A1 (fr) Procédé et dispositif pour fournir des images irm relatives à au moins une partie du corps d'un patient avec une administration réduite de produit de contraste
EP0890153B1 (fr) Procede de determination de poids aptes a etre elimines, d'un reseau neuronal, au moyen d'un ordinateur
WO2001085018A2 (fr) Procede et dispositif de classification de valeurs de serie, support d'enregistrement lisible par ordinateur, et element de programme informatique
EP1934895A2 (fr) Procede d'apprentissage assiste par ordinateur d'un reseau neuronal, et reseau neuronal correspondant
DE102019217524A1 (de) Verfahren und Bildverarbeitungsvorrichtung zum Segmentieren von Bilddaten und Computerprogrammprodukt
EP4032030A1 (fr) Apprentissage quantifié rapide de modules aptes à l'apprentissage
EP1034464B1 (fr) Dispositif servant a predire une anomalie dans un systeme et a prendre des mesures permettant de corriger cette anomalie
DE112022003672T5 (de) Modellieren des herzens eines subjekts
WO1998034176A1 (fr) Procede pour la transformation d'une logique floue servant a la simulation d'un processus technique en un reseau neuronal
EP1359539A2 (fr) Modèle neurodynamique de traitement d'informations visuelles
DE19608734C1 (de) Verfahren zur Klassifikation einer meßbaren Zeitreihe, die eine vorgebbare Anzahl von Abtastwerten aufweist, insbesondere eines elektrischen Signals, durch einen Rechner und Verwendung des Verfahrens

Legal Events

Date Code Title Description
AK Designated states

Kind code of ref document: A2

Designated state(s): CN JP US

AL Designated countries for regional patents

Kind code of ref document: A2

Designated state(s): AT BE CH CY DE DK ES FI FR GB GR IE IT LU MC NL PT SE TR

121 Ep: the epo has been informed by wipo that ep was designated in this application
DFPE Request for preliminary examination filed prior to expiration of 19th month from priority date (pct application filed before 20040101)
122 Ep: pct application non-entry in european phase
NENP Non-entry into the national phase

Ref country code: JP