Time series data analyzer, and a computer-readable recording medium recording a time series data analysis program

• US 8,296,108 B2
• Filed: 04/02/2010
• Issued: 10/23/2012
• Est. Priority Date: 04/02/2010
• Status: Expired due to Fees

• Current Assignee: Yugen Kaisha Suwa Torasuto
• Original Assignee: Yugen Kaisha Suwa Torasuto
• Inventors: Tanaka, Yukio
• Primary Examiner(s): Patel, Shambhavi
• Application Number:
  US12/753,457
• Publication Number:
  US 20110246144A1
• Time in Patent Office: 935 Days
• Field of Search: 600/27, 600/300, 600/301, 600/323, 600/428, 600/431, 600/483, 600/500, 600/504, 600/508, 600/509, 600/544, 607/40, 607/45, 702/183, 702/186, 702/189, 702/191, 706/20, 706/52, 703/2, 703/11
• US Class Current: 703/2
• CPC Class Codes: A61B 5/316 Modalities, i.e. specific diagnostic methods
• CPC Class Codes: A61B 5/374 Detecting the frequency distribution of signals, e.g. detecting delta, theta, alpha, beta or gamma waves

Abstract

A time series data analyzer includes a segment condition input section, an analysis condition input section, and an optimum condition deriving section for analyzing all segments based on the segment conditions and analysis conditions inputted in the respective input sections, under all analysis conditions by a maximum entropy method and a nonlinear least squares method. The time series data analyzer derives the optimum segment length and the optimum lag value in correspondence to selected results, and an analysis execution section executes analysis by the maximum entropy method by setting the optimum analysis conditions derived as described above. The trending of the spectrum of electroencephalogram data is used as an indicator of the state of the subject based on the findings that the spectrum of electroencephalogram data is an exponential spectrum and the gradient changes depending on the state of the subject.

Claims

1. A time series data analyzer for analyzing the segments obtained from time series data, comprising:

a non-transitory memory device storing a program; and

a processor that executes the program, which causes the time series data analyzer to operate as;

a segment condition input section, into which a shortest segment length, a longest segment length and a total number of obtained segments including the shortest segment, the longest segment and other segments are input as input items, the other segments being segments with different lengths ranging from the shortest segment length to the longest segment length or located at each time step between the shortest segment and the longest segment,

an analysis condition input section, into which a minimum lag value, a maximum lag value, and intervals for setting a series of lag values between the minimum lag value and the maximum lag value are inputted as input items,

an optimum analysis condition deriving section, in which each segment is analyzed based on segment conditions inputted in the segment condition input section and analysis conditions inputted in the analysis condition input section from all analysis conditions using a maximum entropy method and a nonlinear least squares method, for selecting one appropriate segment and one appropriate analysis condition from all analysis results as a selected result and deriving an optimum segment length and an optimum lag value as optimum analysis conditions in correspondence to the selected result, and

an analysis execution section, in which the optimum analysis conditions derived by the optimum analysis condition deriving section are set as analysis conditions, for executing an analysis by the maximum entropy method.

2. A time series data analyzer according to claim 1, wherein the optimum analysis condition deriving section comprises:

a setting processing means for obtaining multiple segments different in length from the time series data based on the conditions inputted in the segment condition input section, to set the segments as sample segments and for reading the analysis conditions inputted in the analysis condition input section, to set the analysis conditions;

a first processing means for calculating power spectral densities for each of the sample segments set by the setting processing means, under all analysis conditions using the maximum entropy method;

a second processing means for extracting dominant spectral peaks for the respective power spectral densities obtained by the first processing means, and calculating respective quantities of trigonometric polynomial expressions for the sample segments from the dominant spectral peaks by the nonlinear least squares method; and

a third processing means for determining validity of the trigonometric polynomial expressions obtained by the second processing means for the sample segment data and consistency of the trigonometric polynomial expressions for the power spectral densities, and selecting one sample segment corresponding to a trigonometric polynomial expression having a specified validity and consistency, for deriving a segment length and a lag value corresponding to the sample segment as optimum segment length and the optimum lag value.

3. A time series data analyzer according to claim 2, wherein

the second processing means extracts a number of the dominant spectral peaks, and respective peak frequencies and peak powers of the dominant spectral peaks from the power spectral densities calculated by the first processing means, to calculate the respective quantities of the trigonometric polynomial expressions by calculation of the nonlinear least squares method for minimizing sum of squares of residuals, with the number of the dominant spectral peaks being the number of terms and the inverse numbers of the respective peak frequencies as the initial values of the periods of respective trigonometric terms, and

wherein the third processing means includes:

a first selection function for comparing standard deviations of residuals in the respective quantities of the trigonometric polynomial expressions concerning the respective sample segments, with a set value, to reject the trigonometric polynomial expressions, the standard deviations of which not being lower than a set value and retaining other trigonometric polynomial expressions as selection candidates,

a second selection function for comparing periods of respective terms of polynomials decided by the nonlinear least squares method with initial values of periods set from inverse numbers of peak frequencies of power spectral densities, to reject the trigonometric polynomial expressions, for which differences obtained as results of the comparison are not lower than a set value, and to retain the other trigonometric polynomial expressions as selection candidates, and

a third selection function for comparing powers obtained from the amplitudes of the respective terms of the polynomials decided by the nonlinear least squares method, with the powers of peaks corresponding to power spectral densities pair by pair, to reject the trigonometric polynomial expressions having the amplitudes, the differences of which obtained as results of the comparison are not lower than a set value, and to retain the other trigonometric polynomial expressions as selection candidates,

wherein the processings by the first, second and third selection functions are performed one after another for selecting the trigonometric polynomial expression(s) high in validity and consistency.

4. A time series data analyzer according to claim 3, wherein the third processing means has a fourth selection function for selecting a trigonometric polynomial expression that is smallest in the standard deviation of residuals in the case where multiple trigonometric polynomial expressions remain as selection candidates after performing the processings by the first, second and third selection functions one after another.

5. A time series data analyzer according to claim 3, wherein the third processing means has a fifth selection function for selecting the longest sample segment as the optimum sample segment in the case where multiple sample segments having the selected trigonometric polynomial expression remain after performing the processings by the first, second and third selection functions to ensure that the sample segment selected by the fifth selection function can be selected as the optimum segment and that the lag value used for calculating the power spectral densities in correspondence to the selected trigonometric polynomial expression can be selected as the optimum lag value for the sample segment.

6. A time series data analyzer according to claim 4, wherein the third processing means has a fifth selection function for selecting the longest sample segment as the optimum sample segment in the case where multiple sample segments having the selected trigonometric polynomial expression remain after performing the processings by the first, second, third and fourth selection functions, to ensure that the sample segment selected by the fifth selection function can be selected as the optimum segment and that the lag value used for calculating the power spectral densities in correspondence to the selected trigonometric polynomial expression can be selected as the optimum lag value for the sample segment.

7. A time series data analyzer according to claim 1, wherein the analysis execution section comprises:

a segment preparation means for obtaining segments from the time series data based on the optimum segment length derived by the optimum analysis condition deriving section and set in an analysis condition setting means;

a power spectral density calculation means for calculating the power spectral densities for the prepared segments by the maximum entropy method using the optimum lag value derived by the optimum analysis condition deriving section and set in the analysis condition setting means;

a peak extraction means for extracting the number of dominant peaks of the calculated power spectral densities, and center frequencies and peak powers of the respective peaks; and

a characteristic quantity extraction means for extracting characteristic quantities from an extracted peak string.

8. A time series data analyzer according to claim 7, wherein the peak extraction means extracts dominant peaks by a first processing of obtaining a first local peak in each range having as width at least three times the average interval between peaks obtained from all the peaks of power spectral densities, and fusing the obtained local peak with any adjacent peak within each interval shorter than the average interval thereby forming a first fused peak, and by a second processing of fusing all the peaks other than the first local peak and the first fused peak with the nearest second local peak or second fused peak respectively, wherein in the case where multiple peaks are fused, the peak powers prevailing before and after the fusion are conserved, and a center frequency of each fused peak is obtained in proportion to the peak powers of the multiple peaks existing before the fusion.

9. A time series data analyzer according to claims 7, wherein the characteristic quantity extraction means comprises:

a trend line calculation means for obtaining a gradient and a y-intercept of a trend line of an exponential spectrum by the linear least squares method for a set of logarithmic peak power values and peak frequencies of the multiple dominant peaks extracted by the peak extraction means;

a relative power calculation means for obtaining relative power by dividing the peak power values of all the peaks obtained by the power spectral density calculation means, by the power values shown by the trend line, for normalization, and

a characteristic quantity calculation means for calculating characteristic quantities of the relative peak power distribution from the peaks discretely arranged on the frequency axis and the relative peak powers thereof.

10. A time series data analyzer according to claim 7, wherein the time series data is electroencephalogram data, and the characteristic quantities extracted by the characteristic quantity extraction means include three frequencies at which an integrated value obtained by integrating the relative peak powers from a low frequency side in the frequency band of interest of an electroencephalogram spectrum ranging from 1 to 30 Hz or 0.5 to 30 Hz corresponds to 25%, 50% or 75% of a total relative peak power of the frequency band of interest.

11. A non-transitory computer-readable recording medium recording a time series data analysis program, for analyzing the segments obtained from time series data, the time series data analysis program causing a computer to perform steps comprising:

a step of inputting a shortest segment length, a longest segment length, and a total number of obtained segments including the shortest segment, the longest segment and other segments as input items, the other segments having different lengths ranging from the shortest segment length to the longest segment length or located at each time step between the shortest segment and the longest segment;

a step of inputting a minimum lag value, a maximum lag value, and intervals for setting the series of lag values between the minimum lag value and the maximum lag value, as input items;

a step of analyzing each of the segments based on inputted segment conditions and inputted analysis conditions from all analysis conditions using a maximum entropy method and a nonlinear least squares method, for selecting one appropriate segment and one appropriate analysis condition from all analysis results as a selected result and deriving optimum analysis conditions consisting of an optimum segment length and an optimum lag value in correspondence to the selected result; and

a step of setting the derived optimum analysis conditions as analysis conditions for executing the analysis by the maximum entropy method.

12. A non-transitory computer-readable recording medium recording a time series data analyzing program, according to claim 11, wherein the step of deriving analysis comprises:

a step of obtaining multiple segments different in length from the time series data based on the inputted segment conditions and setting them as sample segments;

a step of setting analysis conditions;

a first processing step of calculating power spectral densities for each of the set sample segments under all the set analysis conditions using the maximum entropy;

a second processing step of extracting dominant spectral peaks for respective power spectral densities obtained by the first processing step, and calculating respective quantities of trigonometric polynomial expressions for the sample segments from the dominant spectral peaks by the nonlinear least squares method; and

a third processing step of determining validity of the trigonometric polynomial expressions obtained by the second processing step for the sample segment data and consistency of the trigonometric polynomial expressions for the power spectral densities and selecting one sample segment corresponding to the trigonometric polynomial expression having a certain validity and consistency, and for deriving the segment length and the lag value corresponding to the selected sample segment, as the optimum segment length and the optimum lag value.

13. A non-transitory computer-readable recording medium recording a time series data analysis program, according to claim 12, wherein

wherein the second processing step includes:

a step of extracting a number of the dominant spectral peaks, and peak frequencies and peak powers of the dominant spectral peaks from the power spectral densities calculated by the first processing step, and

a step of calculating the respective quantities of the trigonometric polynomial expressions by calculation of the nonlinear least squares method for minimizing a sum of squares of residuals, with the number of the dominant spectral peaks as the number of terms, and the inverse numbers of the respective peak frequencies as the initial values of the periods of respective trigonometric terms, and

wherein the third processing step includes;

a first selection step of comparing standard deviations of residuals in the respective quantities of the trigonometric polynomial expressions concerning the respective sample segments, with a set value, to reject the trigonometric polynomial expressions, the standard deviations of which are not lower than the set value and to retain the other trigonometric polynomial expressions as selection candidates,

a second selection step of comparing the periods of the respective terms of the polynomials decided by the nonlinear least squares method, with the initial values of the periods set from inverse numbers of the peak frequencies of power spectral densities, to reject the trigonometric polynomial expressions, for which differences obtained as results of the comparison are not lower than a set value, and to retain the other trigonometric polynomial expressions as selection candidates, and

a third selection step of comparing powers obtained from amplitudes of the respective terms of the polynomials decided by the nonlinear least squares method, with the powers of the peaks corresponding to power spectral densities pair by pair, to reject the trigonometric polynomial expressions having the amplitudes, differences of which obtained as results of the comparison are not lower than a set value, and to retain the other trigonometric polynomial expressions as selection candidates,

wherein the processings by the first, second and third selection steps are performed one after another for selecting the trigonometric polynomial expression(s) high in validity and consistency.

14. A non-transitory computer-readable recording medium recording a time series data analysis program according to claim 13,

wherein the third processing step has a fourth selection step of selecting the trigonometric polynomial expression that is smallest in the standard deviation of residuals in the case where multiple trigonometric polynomial expressions remain as selection candidates after performing the processings by the first, second and third selection steps one after another.

15. A non-transitory computer-readable recording medium recording a time series data analysis program, according to claim 13,

wherein the third processing step has a fifth selection step of selecting a longest sample segment as an optimum sample segment in the case where multiple sample segments having the selected trigonometric polynomial expression remain as selection candidates after performing the processings by the first, second and third selection steps, to ensure that the sample segment selected by the fifth selection step can be selected as an optimum segment and that a lag value used for calculating the power spectral densities in correspondence to the selected trigonometric polynomial expression can be selected as the optimum lag value for the sample segment.

16. A non-transitory computer-readable recording medium recording a time series data analysis program, according to claim 14,

wherein the third processing step has a fifth selection step of selecting a longest sample segment as an optimum sample segment in the case where multiple sample segments having the selected trigonometric polynomial expression remain as selection candidates after performing the processings by the first, second, third and fourth selection steps, to ensure that the sample segment selected by the fifth selection step can be selected as an optimum segment and that a lag value used for calculating the power spectral densities in correspondence to the selected trigonometric polynomial expression can be selected as the optimum lag value for the sample segment.

17. A non-transitory computer-readable recording medium recording a time series data analysis program, according to any one of claims 11, wherein the analysis execution step comprises:

a segment preparation step of obtaining segments from the time series data based on the optimum segment length derived by the optimum analysis condition deriving step and set in an analysis condition setting step,

a power spectral density calculation step of calculating the power spectral densities for the prepared segments by the maximum entropy method using the optimum lag value derived by the optimum analysis condition deriving step and set in the analysis condition setting step,

a peak extraction step of extracting a number of dominant peaks of the calculated power spectral densities, and center frequencies and peak powers of the respective peaks, and

a characteristic quantity extraction step of extracting characteristic quantities from an extracted peak string.

18. A non-transitory computer-readable recording medium recording a time series data analysis program, according to claim 17, wherein the peak extraction step comprises:

a first processing step comprising

a step of obtaining an average interval between peaks from all the peaks of power spectral densities,

a step of obtaining a first local peak in each range having a width at least three times the obtained average interval and

a step of fusing the obtained the first local peak with any adjacent peak within each interval shorter than the average interval for forming a first fused peak, and

a second processing step of fusing all the peaks other than the first local peak and the first fused peak with a nearest second local peak or a second fused peak respectively, wherein

in the first processing step and the second processing step, in the case where multiple peaks are fused, the peak powers prevailing before and after the fusion are conserved, and a center frequency of each fused peak is obtained in proportion to the peak powers of the multiple peaks existing before the fusion.

19. A non-transitory computer-readable recording medium recording a time series data analysis program, according to claim 17, wherein the characteristic quantity extraction step comprises:

a trend line calculation step of obtaining a gradient and a y-intercept of a trend line of an exponential spectrum by the linear least squares method for a set of logarithmic peak power values and peak frequencies of the multiple dominant peaks extracted by the peak extraction step,

a relative power calculation step of obtaining relative power by dividing the peak power values of all the peaks obtained by the power spectral density calculation step, by the power values shown by the trend line, for normalization, and

a characteristic quantity calculation step of calculating characteristic quantities of the relative peak power distribution from the peaks discretely arranged on the frequency axis and the relative peak powers thereof.

20. A non-transitory computer-readable recording medium recording a time series data analysis program, according to claim 17, wherein the time series data is electroencephalogram data, and the characteristic quantities extracted by the characteristic quantity extraction step include three frequencies at which an integrated value obtained by integrating the relative peak powers from a low frequency side in a frequency band of interest of an electroencephalogram spectrum ranging from 1 to 30 Hz or 0.5 to 30 Hz corresponds to 25%, 50% or 75% of a total relative peak power of the frequency band of interest.