Real-time waveform analysis using artificial neural networks

• US 5,751,911 A
• Filed: 11/12/1996
• Issued: 05/12/1998
• Est. Priority Date: 10/18/1991
• Status: Expired due to Term

• Current Assignee: Julian M. Goldman
• Original Assignee: Julian M. Goldman
• Inventors: Goldman, Julian M.
• Primary Examiner(s): Davis, George B.
• Application Number:
  US08/747,460
• Time in Patent Office: 546 Days
• Field of Search: 395/22, 395/23, 395/20, 395/21
• US Class Current: 706/20
• CPC Class Codes: A61B 5/0803 Recording apparatus specially adapted therefor
• CPC Class Codes: A61B 5/0836 Measuring rate of CO2 production
• CPC Class Codes: A61B 5/7264 Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems neural networks per se G06N3/00; expert systems per se G06N5/00
• CPC Class Codes: A61B 5/7267 involving training the classification device
• CPC Class Codes: G16H 50/20 for computer-aided diagnosis, e.g. based on medical expert systems

Abstract

A real-time waveform analysis system utilizes neural networks to perform various stages of the analysis. The signal containing the waveform is first stored in a buffer and the buffer contents transmitted to a first and second neural network which have been previously trained to recognize the start point and the end point of the waveform respectively. A third neural network receives the signal occurring between the start and end points and classifies that waveform as comprising either an incomplete waveform, a normal waveform or one of a variety of predetermined characteristic classifications. Ambiguities in the output of the third neural network are arbitrated by a fourth neural network which may be given additional information which serves to resolve these ambiguities. In accordance with the preferred embodiment, the present invention is applied to a system analyzing respiratory waveforms of a patient undergoing anesthesia and the classifications of the waveform correspond to normal or various categories of abnormal features functioning in the respiratory signal. The system performs the analysis rapidly enough to be used in real-time systems and can be operated with relatively low cost hardware and with minimal software development required.

Claims

1. A method of analyzing a waveform in a signal using a neural network having input nodes and output nodes, said method comprising the steps of:

transmitting a first portion of said waveform to said neural network, such that N consecutive samples of said waveform are transmitted to N input nodes of said neural network;

said neural network being trained to recognize the starting point and ending point of said waveform by producing outputs which are a representation of an index number of the portion of the sample of the waveform corresponding to the starting and ending points; and

classifying said waveform as one characteristic waveform type of a plurality of characteristic waveform types, or as an incomplete waveform, in response to the index number.

2. A method of analyzing a waveform in a signal, said signal having a plurality of characteristic waveform types, said method comprising the steps of:

providing a pretrained neural network having N input nodes and R output nodes, said neural network being trained to produce output signals in said output nodes which represent a plurality of classifications of said characteristic waveform types, one of said classifications indicating a partial waveform signifying that a complete waveform has not been received by said neural network; and

transmitting said waveform to said N input nodes of said pretrained neural network, wherein said pretrained neural network produces an output which classifies said waveform as either a partial waveform or as one of said characteristic waveform types.

3. The method of claim 2 further comprising the steps of:

identifying the starting and end points of said waveform before the step of transmitting; and

normalizing said waveform to comprise a predetermined duration and amplitude, wherein said starting and ending points of said waveform comprise N samples.

4. The method of claim 3 further comprising the step of normalizing the lead and lag portions of said waveform preceding said starting point and following said end point respectively.

5. A method of analyzing a waveform having a characteristic shape in a signal using a neural network, said method comprising the steps of:

using a classification means to classify said waveform based on its shape, said classification means assigning numerical values to the likelihood that said waveform belongs to one of a plurality of predetermined classes;

when said numerical values do not unambiguously indicate a single class, transmitting said numerical values to input nodes of a neural network;

transmitting additional information relating to said waveform to additional input nodes of said neural network; and

said neural network being previously trained to classify said waveform by producing output numerical values representing the likelihood that said waveform belongs to one of a plurality of classes, said neural network being trained with inputs containing said additional information, said additional information serving to resolve said ambiguities between various classifications of waveform shapes by correlating an atypical shape of a particular waveform classification with particular values for said additional information.

6. The method of claim 5 wherein said waveform is a signal derived from a capnogram waveform and said classes produced by said classification means represent abnormalities in said capnogram waveform and said additional information transmitted to said neural network includes minimum CO₂, maximum CO₂ and respiratory rate.

7. A method for analyzing a waveform in a signal using neural networks having input nodes and output nodes, said method comprising:

transmitting a first portion of said waveform to first and second neural networks, such that N consecutive samples of said waveform are transmitted to N input nodes of said first and second neural networks;

said first and second neural networks being trained to recognize the starting point and ending point of said waveform by producing outputs which are a representation of an index number of the portion of the sample of the waveform corresponding to the starting and ending points;

providing a third neural network having N input nodes and R output nodes, said third neural network being trained to produce output signals at said output nodes which represent a plurality of classifications of respiratory waveform types, and one of said classifications indicating a partial waveforms signifying that a complete waveform has not been received by said third neural network;

transmitting said waveform to said N input nodes of said third neural network, wherein said third neural network produces an output which classifies said waveform as either a partial waveform or as one class of said respiratory waveform types;

when said numerical values do not unambiguously indicate a single class, transmitting said waveform to input nodes of a fourth neural network;

transmitting additional information relating to said waveform to additional input nodes of said fourth neural network; and

said fourth neural network being previously trained to classify said waveform by producing output numerical values representing the likelihood that said waveform belongs to one of a plurality of classes, said fourth neural network being trained with inputs containing said additional information, said additional information serving to resolve said ambiguities between various classifications of said respiratory waveform types, by correlating an atypical waveform shape of a particular classification with particular values for additional information.

8. A system for analyzing a waveform in a signal using a neural network having input nodes and output nodes, said system comprising:

a neural network trained to recognize the starting point and ending point of said waveform by producing outputs which are a representation of an index number of the portion of the sample of the waveform corresponding to the starting and ending points;

means for transmitting a first portion of said signal to said neural network, such that N consecutive samples of said waveform are transmitted to N input nodes of said neural network; and

means for classifying said waveform as one characteristic waveform type of a plurality of characteristic waveform types, or as an incomplete waveform in response to the N consecutive samples.

9. System for analyzing a waveform in a signal, said signal having a plurality of characteristic waveform types, said system comprising:

a pretrained neural network having N input nodes and R output nodes, said neural network being trained to produce output signals in said output nodes which represent a plurality of classifications of characteristic waveform types, one of said classifications indicating a partial waveform signifying that a complete waveform has not been received by said neural network; and

means for transmitting each of said neural network samples to one of said N input nodes of said pretrained neural network, wherein said pretrained neural network produces an output which classifies said waveform as either a partial waveform or as one of said characteristic waveform types.

10. The system of claim 9, further comprising:

means for identifying the starting and end points of said waveform; and

means for normalizing said waveform to comprise a predetermined duration and amplitude, wherein said starting and ending points of said waveform comprise N samples.

11. The system of claim 10 further comprising means for normalizing the lead and lag portions of said waveform preceding said starting point and following said end point respectively.

12. A system for analyzing a waveform having a characteristic shape in a signal using a neural network, said system comprising:

a classification means for classifying said waveform based on its shape, said classification means assigning numerical values to the likelihood that said waveform belongs to one of a plurality of predetermined classes;

means for transmitting said numerical values to input nodes of a neural network when said numerical values do not unambiguously indicate a single class;

means for providing additional information relating to said waveform to additional input nodes of said neural network; and

wherein said neural network being previously trained to classify said waveform by producing output numerical values representing a likelihood that said waveform belongs to one of a plurality of classes, said neural network being trained with inputs containing said additional information, said additional information serving to resolve said ambiguities between various classifications of waveform shapes by correlating an atypical shape of a particular waveform classification with particular values for said additional information.

13. The method of claim 12 wherein said waveform is a signal derived from a capnogram waveform, said classes produced by said classification system represent abnormalities in said capnogram waveform, and said additional information transmitted to said neural network includes minimum CO₂, maximum CO₂ and respiratory rate.