Waveform analysis apparatus and method using neural network techniques

• US 5,092,343 A
• Filed: 11/17/1989
• Issued: 03/03/1992
• Est. Priority Date: 02/17/1988
• Status: Expired due to Term

• Current Assignee: A. Robert Spitzer
• Original Assignee: Wayne State University
• Inventors: Hassoun, Mohamad, Spitzer, Robert
• Primary Examiner(s): Kamm, William E.
• Application Number:
  US07/438,581
• Time in Patent Office: 837 Days
• Field of Search: 128/696, 128/702, 128/731, 128/733, 364/413.02, 364/413.05, 364/413.06
• US Class Current: 600/515
• CPC Class Codes: A61B 5/389 Electromyography [EMG]
• CPC Class Codes: A61B 5/7239 using differentiation including higher order derivatives
• CPC Class Codes: A61B 5/7267 involving training the classification device
• CPC Class Codes: G06F 2218/12 Classification; Matching
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/10 Interfaces, programming languages or software development kits, e.g. for simulating neural networks
• CPC Class Codes: G16H 15/00 ICT specially adapted for medical reports, e.g. generation or transmission thereof
• CPC Class Codes: G16H 40/63 for local operation
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems
• CPC Class Codes: Y10S 128/925 Neural network

Abstract

A waveform analysis assembly (10) includes a sensor (12) for detecting physiological electrical and mechanical signals produced by the body. An extraction neural network (22, 22'"'"') will learn a repetitive waveform of the electrical signal, store the waveform in memory (18), extract the waveform from the electrical signal, store the location times of occurrences of the waveform, and subtract the waveform from the electrical signal. Each significantly different waveform in the electrical signal is learned and extracted. A single or multilayer layer neural network (22, 22'"'"') accomplishes the learning and extraction with either multiple passes over the electrical signal or accomplishes the learning and extraction of all waveforms in a single pass over the electrical signal. A reducer (20) receives the stored waveforms and times and reduces them into features characterizing the waveforms. A classifier neural network (36) analyzes the features by classifying them through nonliner mapping techniques within the network representing diseased states and produces results of diseased states based on learned features of the normal and patient groups.

Claims

1. A wave analysis assembly for discriminating electrical signals and for producing output indicative of disease characteristics, said assembly comprising;

means (12) for producing an electrical signal, control means (20) for identifying the location of individual unknown waveforms occurring about a peak within said electrical signal and producing a set comprised of the individual unknown waveforms, extraction neural network (22, 22'"'"') having a distributed network of synaptic weights for receiving said set and for learning the unknown waveform within said electrical signal through a mapping transformation of said individual waveforms of the set in the distributed network of synaptic weights to develop and store an idealized representation of the unknown waveform producing a learned waveform and for extracting said learned waveform from said electrical signal, and output means (42) for outputting information based on said extracted waveform.

2. An assembly as set forth in claim 1 further characterized by said network (22, 22'"'"') including weight matrix means for storing an idealized representation of the repetitive waveform by synaptic weights and updating the synaptic weights by an error signal.

3. An assembly as set forth in claim 2 further characterized by said neural network (22, 22'"'"') including learning means for receiving one of said unknown waveforms and an output waveform produced by sensing said unknown waveform through said weight matrix, and comparing said unknown waveform and said output waveform to produce said error signal.

4. An assembly as set forth in claim 3 further characterized by said neural network (22'"'"') comprising a multilayer network for receiving the learning sets of all waveforms and for learning and extracting the waveform in a first pass over the electrical signal.

5. An assembly as set forth in claim 1 further characterized by including reducing means (34) for analyzing said extracted waveform to produce features characterizing said waveform.

6. An assembly as set forth in claim 5 further characterized by including classifier network means (36) containing standard features and combinations thereof associated with states for comparing said features to said standard features and combination to produce an output representative of states based on the best match of said features with said information.

7. An assembly as set forth in claim 6 further characterized by said extraction network means (22, 22'"'"') including buffer means for storing said extracted waveform of said electrical signal.

8. An assembly as set forth in claim 7 further characterized by said extraction network means (22, 22'"'"') including learning means for learning said repetitive waveform and a weight matrix means (26) for storing said learned waveform.

9. An assembly as set forth in claim 8 further characterized by said extraction network means (22) including subtracting means (32) for subtracting said learned waveform from said electrical signal.

10. An assembly as set forth in claim 9 further characterized by said neural network (22) comprising a single layer network for receiving one learning set for a waveform and learning the waveform in a first pass and extracting the waveform in a second pass over the electrical signal.

11. A wave analysis assembly for discriminating physiological signals produced within the body and for producing output indicative of disease characteristics, said assembly comprising;

means for supplying a plurality of independent features characteristic of a waveform contained within a physiological signal, a classifier neural network (36) having a nonlinear distributed network of learned features and combinations thereof associated with a plurality of disease and normal patient states for transforming the features characterizing the physiological signal through the distributed network to produce an output representative of any of said plurality diseased and normal states, and output means for receiving the output and indicating the diseased and normal states.

12. An assembly as set forth in claim 11 further characterized by said classifier neural network (36) including reference memory means (40) comprised of a weight matrix for storing the learned features and combinations for diagnosis and for nonlinearly mapping said features within said weight matrix to produce said output.

13. An assembly as set forth in claim 12 further characterized by including sensing means (12) for sampling a physiological signal produced within the body and for producing an electrical signal.

14. An assembly as set forth in claim 13 further characterized by including reducing means (36) for analyzing said electrical signal and for producing features characterizing said electrical signal.

15. An assembly as set forth in claim 14 further characterized by including extraction network means (22, 22'"'"') for receiving said electrical signal and identifying and learning a waveform comprising said electrical signal and for extracting said learned waveform, and for storing said extracted waveform and associated times of occurrence of substantially similar waveforms within said electrical signal.

16. An assembly as set forth in claim 15 further characterized by including preprocessing means (18) for filtering and differentiating said electrical signal to isolate spikes associated with waveforms.

17. An assembly as set forth in claim 16 further characterized by including control means (20) for controlling said extraction network means (22) and for identifying waveforms in said electrical signal.

18. A wave analysis assembly for discriminating electrical signals having a waveform, said assembly comprising;

input means (12) for sampling an electrical signal, control means (20) for identifying the location of individual unknown waveforms occurring about a peak within said electrical signal and producing a set comprised of the individual unknown waveforms, extraction neural network (22, 22'"'"') having a distributed network of synaptic weights for receiving the set and for learning the unknown waveform by a mapping transformation of said individual waveforms of the set in the distributed network of synaptic weights to produce a learned waveform and for extracting the learned waveform from the electrical signal, and output means (42) for outputting information of said extracted waveform, said extraction neural network including learning means for learning a template waveform inputted by the user for the extraction.

19. An assembly as set forth in claim 18 further characterized by said extraction neural network (16) including extracting means (28) for extracting waveforms from said electrical signal substantially similar to said learned waveform.

20. A method of discriminating electrical signals and producing an output, said method comprising the steps of:

sensing the signals and producing an electrical signal, identifying the location of individual unknown waveforms occurring about a peak within the electrical signal, producing a set comprised of the individual unknown waveforms, learning the unknown waveform within the electrical signal by a mapping transformation of the individual waveforms of the set in a distributed network of synaptic weights to produce a learned waveform, extracting the learned waveform from the electrical signal, and outputting information based on the extracted waveform.

21. A method as set forth in claim 20 further characterized by storing an idealized representation of the repetitive waveform by synaptic weights.

22. A method as set forth in claim 21 further characterized by receiving an unknown waveform and sending same through the weights to produce an output waveform.

23. A method as set forth in claim 22 further characterized by comparing the unknown waveform to the output waveform to produce an error signal.

24. A method as set forth in claim 23 further characterized by updating the synaptic weights by the error signal.

25. A method as set forth in claim 24 further including filtering and differentiating the electrical signal to isolate spikes associated with occurrences of the repetitive waveform.

26. A method as set forth in claim 25 further including reducing the waveform into features characterizing the waveform.

27. A method as set forth in claim 26 further including analyzing the features with respect to information comprising standard features and combinations thereof associated with disease states.

28. A method as set forth in claim 27 further including outputting states data based on the best match of the features with the information.

29. A method as set forth in claim 28 further including storing the extracted waveform and associated times of occurrence of the waveform within the electrical signal.

30. A method as set forth in claim 29 further including subtracting the learned waveform from the electrical signal.

31. A method for discriminating physiological signals produced within the body and for producing output indicative of disease characteristics, the method including the steps of;

producing a plurality of independent features characteristic of a waveform contained within a physiological signal, learning and storing known features and combinations thereof associated with a plurality of disease and normal patient states in a distributed network, transforming the features through the distributed network to produce an output signal representative of any of said plurality of diseased states.

32. A method as set forth in claim 31 further characterized by storing the known features and combinations thereof in a weight matrix, and nonlinearly mapping the features within the weight matrix to produce the output.

33. A method as set forth in claim 32 further characterized by acquiring additional features and combinations thereof associated with disease and normal patient states, and training the distributed network by updating the weight matrix with the additional features and combinations thereof associated with disease and normal patient states.