Pattern recognition of spectral entropy features for detection of alcoholic and control visual ERP’s in multichannel EEGs

T. K. Padma Shri, N. Sriraam

Research output: Contribution to journalArticle

2 Citations (Scopus)

Abstract

This paper presents a novel ranking method to select spectral entropy (SE) features that discriminate alcoholic and control visual event-related potentials (ERP’S) in gamma sub-band (30–55 Hz) derived from a 64-channel electroencephalogram (EEG) recording. The ranking is based on a t test statistic that rejects the null hypothesis that the group means of SE values in alcoholics and controls are identical. The SE features with high ranks are indicative of maximal separation between their group means. Various sizes of top ranked feature subsets are evaluated by applying principal component analysis (PCA) and k-nearest neighbor (k-NN) classification. Even though ranking does not influence the performance of classifier significantly with the selection of all 61 active channels, the classification efficiency is directly proportional to the number of principal components (pc). The effect of ranking and PCA on classification is predominantly observed with reduced feature subsets of (N = 25, 15) top ranked features. Results indicate that for N = 25, proposed ranking method improves the k-NN classification accuracy from 91 to 93.87% as the number of pcs increases from 5 to 25. With same number of pcs, the k-NN classifier responds with accuracies of 84.42–91.54% with non-ranked features. Similarly for N = 15 and number of pcs varying from 5 to 15, ranking enhances k-NN detection accuracies from 88.9 to 93.08% as compared to 86.75–91.96% without ranking. This shows that the detection accuracy is increased by 6.5 and 2.8%, respectively, for N = 25, whereas it enhances by 2.2 and 1%, respectively, for N = 15 in comparison with non-ranked features. In the proposed t test ranking method for feature selection, the pcs of only top ranked feature candidates take part in classification process and hence provide better generalization.

Original languageEnglish
Pages (from-to)147-158
Number of pages12
JournalBrain Informatics
Volume4
Issue number2
DOIs
Publication statusPublished - 01-06-2017

Fingerprint

Enterprise resource planning
Entropy
Electroencephalography
Pattern recognition
Principal Component Analysis
Principal component analysis
Classifiers
Alcoholics
Set theory
Evoked Potentials
Feature extraction
Statistics

All Science Journal Classification (ASJC) codes

  • Neurology
  • Computer Science Applications
  • Cognitive Neuroscience

Cite this

@article{b49ff10f17f64a57a2a224ee79cab48c,
title = "Pattern recognition of spectral entropy features for detection of alcoholic and control visual ERP’s in multichannel EEGs",
abstract = "This paper presents a novel ranking method to select spectral entropy (SE) features that discriminate alcoholic and control visual event-related potentials (ERP’S) in gamma sub-band (30–55 Hz) derived from a 64-channel electroencephalogram (EEG) recording. The ranking is based on a t test statistic that rejects the null hypothesis that the group means of SE values in alcoholics and controls are identical. The SE features with high ranks are indicative of maximal separation between their group means. Various sizes of top ranked feature subsets are evaluated by applying principal component analysis (PCA) and k-nearest neighbor (k-NN) classification. Even though ranking does not influence the performance of classifier significantly with the selection of all 61 active channels, the classification efficiency is directly proportional to the number of principal components (pc). The effect of ranking and PCA on classification is predominantly observed with reduced feature subsets of (N = 25, 15) top ranked features. Results indicate that for N = 25, proposed ranking method improves the k-NN classification accuracy from 91 to 93.87{\%} as the number of pcs increases from 5 to 25. With same number of pcs, the k-NN classifier responds with accuracies of 84.42–91.54{\%} with non-ranked features. Similarly for N = 15 and number of pcs varying from 5 to 15, ranking enhances k-NN detection accuracies from 88.9 to 93.08{\%} as compared to 86.75–91.96{\%} without ranking. This shows that the detection accuracy is increased by 6.5 and 2.8{\%}, respectively, for N = 25, whereas it enhances by 2.2 and 1{\%}, respectively, for N = 15 in comparison with non-ranked features. In the proposed t test ranking method for feature selection, the pcs of only top ranked feature candidates take part in classification process and hence provide better generalization.",
author = "{Padma Shri}, {T. K.} and N. Sriraam",
year = "2017",
month = "6",
day = "1",
doi = "10.1007/s40708-017-0061-y",
language = "English",
volume = "4",
pages = "147--158",
journal = "Brain Informatics",
issn = "2198-4018",
publisher = "Springer Berlin",
number = "2",

}

Pattern recognition of spectral entropy features for detection of alcoholic and control visual ERP’s in multichannel EEGs. / Padma Shri, T. K.; Sriraam, N.

In: Brain Informatics, Vol. 4, No. 2, 01.06.2017, p. 147-158.

Research output: Contribution to journalArticle

TY - JOUR

T1 - Pattern recognition of spectral entropy features for detection of alcoholic and control visual ERP’s in multichannel EEGs

AU - Padma Shri, T. K.

AU - Sriraam, N.

PY - 2017/6/1

Y1 - 2017/6/1

N2 - This paper presents a novel ranking method to select spectral entropy (SE) features that discriminate alcoholic and control visual event-related potentials (ERP’S) in gamma sub-band (30–55 Hz) derived from a 64-channel electroencephalogram (EEG) recording. The ranking is based on a t test statistic that rejects the null hypothesis that the group means of SE values in alcoholics and controls are identical. The SE features with high ranks are indicative of maximal separation between their group means. Various sizes of top ranked feature subsets are evaluated by applying principal component analysis (PCA) and k-nearest neighbor (k-NN) classification. Even though ranking does not influence the performance of classifier significantly with the selection of all 61 active channels, the classification efficiency is directly proportional to the number of principal components (pc). The effect of ranking and PCA on classification is predominantly observed with reduced feature subsets of (N = 25, 15) top ranked features. Results indicate that for N = 25, proposed ranking method improves the k-NN classification accuracy from 91 to 93.87% as the number of pcs increases from 5 to 25. With same number of pcs, the k-NN classifier responds with accuracies of 84.42–91.54% with non-ranked features. Similarly for N = 15 and number of pcs varying from 5 to 15, ranking enhances k-NN detection accuracies from 88.9 to 93.08% as compared to 86.75–91.96% without ranking. This shows that the detection accuracy is increased by 6.5 and 2.8%, respectively, for N = 25, whereas it enhances by 2.2 and 1%, respectively, for N = 15 in comparison with non-ranked features. In the proposed t test ranking method for feature selection, the pcs of only top ranked feature candidates take part in classification process and hence provide better generalization.

AB - This paper presents a novel ranking method to select spectral entropy (SE) features that discriminate alcoholic and control visual event-related potentials (ERP’S) in gamma sub-band (30–55 Hz) derived from a 64-channel electroencephalogram (EEG) recording. The ranking is based on a t test statistic that rejects the null hypothesis that the group means of SE values in alcoholics and controls are identical. The SE features with high ranks are indicative of maximal separation between their group means. Various sizes of top ranked feature subsets are evaluated by applying principal component analysis (PCA) and k-nearest neighbor (k-NN) classification. Even though ranking does not influence the performance of classifier significantly with the selection of all 61 active channels, the classification efficiency is directly proportional to the number of principal components (pc). The effect of ranking and PCA on classification is predominantly observed with reduced feature subsets of (N = 25, 15) top ranked features. Results indicate that for N = 25, proposed ranking method improves the k-NN classification accuracy from 91 to 93.87% as the number of pcs increases from 5 to 25. With same number of pcs, the k-NN classifier responds with accuracies of 84.42–91.54% with non-ranked features. Similarly for N = 15 and number of pcs varying from 5 to 15, ranking enhances k-NN detection accuracies from 88.9 to 93.08% as compared to 86.75–91.96% without ranking. This shows that the detection accuracy is increased by 6.5 and 2.8%, respectively, for N = 25, whereas it enhances by 2.2 and 1%, respectively, for N = 15 in comparison with non-ranked features. In the proposed t test ranking method for feature selection, the pcs of only top ranked feature candidates take part in classification process and hence provide better generalization.

UR - http://www.scopus.com/inward/record.url?scp=85048559148&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85048559148&partnerID=8YFLogxK

U2 - 10.1007/s40708-017-0061-y

DO - 10.1007/s40708-017-0061-y

M3 - Article

AN - SCOPUS:85048559148

VL - 4

SP - 147

EP - 158

JO - Brain Informatics

JF - Brain Informatics

SN - 2198-4018

IS - 2

ER -