Printer Friendly

Differential diagnosis of sleep disorders based on EEG analysis.

INTRODUCTION

Sleep is a quotidian cycle of repeatable sequence of sleep stages over several hours [1, 2]. A healthy sleep cycle is associated with good performance of normal activities during the awake state, such as concentration, learning and memory recollection [3]. About 12.7% of the US population suffer from some form of chronic sleep disorder that may lead to adverse health conditions, such as rise in risk of hypertension and obesity, decrease in the efficacy of immune system, kidney disease, anemia, diabetes, Parkinson's and heart problems [4].

There are two basic states of sleep, non-rapid eye movement (NREM) and rapid-eye movement (REM) sleep, with NREM constituting approximately 75% and REM 25% of a healthy night's sleep. NREM sleep state is further divided into three stages with stage 1 corresponding to a state between wakefulness and sleep, stage 2 where awareness of the outside world begins to fade completely and stage 3 being deep sleep. During NREM stages a sleep specific activity called Cyclic Alternating Pattern (CAP) occurs. CAP is a quasi-periodic EEG activity characterized by cyclic sequences of cerebral activation (phase A; about 8-15 seconds) followed by periods of deactivation (phase B; about 15-20 seconds) [5-7]. Based on occurrence of high voltage slow waves (EEG synchrony) or low-amplitude fast rhythms (EEG desynchrony) [7], CAP-A phase is divided into three subtypes: A1, A2 and A3.

Common types of sleep disorders are parasomnias (periodic limb movement disorder-PLMD, rapid eye movement behavior disorder-REMBD), nocturnal frontal lobe epilepsy-NFLE, and insomnia. In PLMD, involuntarily movement of limbs is observed at periodic intervals of 20-40 seconds in duration [8]. In REMBD, abnormal behavior is observed during the REM stage of sleep, where movement of limbs (kicking, grabbing, or jumping etc.) occurs while the patient unconsciously acts out his dreams. In NFLE, epileptic seizures emanate from the frontal lobes of the brain during sleep, caused by tumors, head trauma or has a genetic base that causes abrasions in the frontal lobes [7]. Insomnia, also called sleeplessness, is when the patient is either unable to fall asleep or has difficulty in sleeping long enough [10].

These sleep disorders are diagnosed by physicians after visual inspection of the patients' sleep patterns from their recorded EEG. This requires an extensive amount of experience and observation time, and is thus a costly and lengthy process [11]. A quantitative and accurate evaluation of sleep EEG can play an important role in the fast and accurate diagnosis of sleep disorders and therefore help provide the necessary treatment with reduced time and cost.

MATERIALS AND METHODS

Data:

In this study, we used polysomnographic (PSG) recordings from the CAP sleep database--PhysioNet [5, 6] recorded at the Sleep Disorders Center of the Ospedale Maggiore of Parma, Italy. The database includes EEG, EMG, EOG, EKG and respiration waveforms from subjects with clinical history of various sleep disorders, as well as from healthy individuals. Recordings from 22 subjects with REMBD, 10 with PLMBD, 37 with NFLE, and 9 with insomnia were included in our analysis. Data from 15 healthy subjects were also included. The data are accompanied by a hypnogram, a temporal profile of the sleep stages scored by an expert neurologist every 30 seconds of recording. We analyzed sleep records based on extraction of features from: a) EEG activity during the different sleep stages, b) CAP activity and c) EEG during CAP activity.

Feature Extraction:

Measures were estimated per nonoverlapping 30 second EEG data segments over the entirety of each subject's recording in order to correspond with the subject's hypnogram. Channel C4-A1 was selected for analysis in most of the subjects. For some subjects where C4-A1 channel was not present, C3-A2 channel was used instead. The EEG was down-sampled to a common sampling frequency of 100Hz. Advanced signal processing techniques were employed to extract quantitative biomarkers relevant to the different sleep stages and CAP phases. Measures used in the analysis of EEG were frequency (median frequency and spectral entropy) and signal complexity (Higuchi fractal dimension) dependent.

Median frequency [1] is defined as the frequency below which 50% of the total EEG spectral power is located and is given by

[mf.sub.50] = mm{f | [[summation].sup.f.sub.i=0Hz] P(i) > 0.5 * [P.sub.50hz]} (1)

where [P.sub.50Hz] is the total power of the EEG signal within the frequency range of 0 to 50 Hz. From the median frequency values we estimated the median period mp=1/ [mf.sub.50].

Spectral entropy (SE) measures the irregularity of the EEG signal in the frequency domain. SE is estimated by Shannon's entropy [12] as

SE = [[summation].sup.N.sub.i=1] [p.sub.i]log(pi)/log(N) (2)

where [p.sub.i] are the normalized spectral magnitudes P(i) at frequencies i and N is the number of discrete frequencies in the spectrum.

Higuchi fractal dimension (HFD) measures the complexity and self-similarity of a signal [1, 12] in the time domain. A self-similar signal would have HFD equal to 2, while a simple signal with no self-similarity (e.g. a simple sinus wave) would have HFD equal to 1. Susmakova and Krakovska used the Higuchi fractal dimension to distinguish individual stages of sleep, specifically stage 3 sleep from other sleep stages [14].

Figure 1 shows the characteristic trends of mp, SE and Higuchi fractal dimension (HFD) from the beginning to the end of a sleep record from a healthy subject along with his hypnogram.

Sleep stage-based features were extracted using the values of the above measures. Normalization to compensate for across subject variability was performed as follows. The mean ([MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]) and standard deviation ([MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]) of each measure (?) values were computed for each sleep stage [s.sub.n] separately, where n = 0 stands for the awake state, n = i for sleep stages (i = 1,2,3 and n = 4 for REM). The normalized features were then estimated as [12]

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (3)

Thus, 12 features (candidate quantitative biomarkers), four from each of the three measures ([MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] with n = 1,2,3,4) were obtained.

In addition to the above 12 sleep stage-based features, features based on the cyclic alternating pattern (CAP) in the EEG were also extracted. These CAP-features were categorized according to the various CAP subtypes ([CAP.sub.A(1)] to [CAP.sub.A(3)]), the total number of CAP cycles during sleep and time intervals related to CAP activity. The CAP features were defined as in [15]: CAP rate (the ratio of CAP time to NREM sleep time); [CAP.sub.A(i)/CAP] (the percentage of CAP A(i) subtype in the total CAP time, for i = 1,2,3); [CAP.sub.A(i)/TST] (the percentage of CAP A(i) in the total sleep time TST); [CAP.sub.A(i)/NREM] (the percentage of CAP A(i) in the total NREM time) and [dCAP.sub.A(i)] (average duration of CAP subtypes, i.e. total duration of CAP A(i) subtype divided by the number of CAP A(i) cycles, for i = 1,2,3). Similarly, for CAP A phases: [CAP.sub.A/CAP] (the percentage of CAP A phase in the total CAP time); [CAP.sub.A/TST] (the percentage of CAP A phase in the total sleep time); [CAP.sub.A/NREM] (the percentage of CAP A phase in the total NREM time); and [dCAP.sub.A] (average duration of CAP A phases). A total of 17 traditional CAP based features were thus obtained.

Finally, the median period, Higuchi fractal dimension and the spectral entropy were also estimated for the CAP phases. Each measure was estimated for each CAP A phase and its ensuing CAP B phase, the ratios of the measure values were then estimated and averaged across all CAP cycles. These are denoted as median period of CAP phase ([CAP.sub.mp]), Higuchi fractal dimension of CAP phase ([CAP.sub.HFD]) and spectral entropy of CAP phase ([CAP.sub.SE]). So a total of 32 features were estimated from the EEG recordings: sleep stage-based (12 features), CAP-based measures (17 features) and CAP/EEG based measures (3 features).

Statistical Analysis:

Statistical testing was performed using Analysis of Variance (ANOVA) [16] and Multivariate Analysis of Variance (MANOVA) to test for statistically significant difference between the 5 groups of subjects on the basis of combined features. In order to rank and combine features, feature selection was done to select the top ten best features with the minimum Redundancy Maximum Relevance (mRMR) procedure [17]. mRMR is a sequential process that gives the set of the most relevant features that can be used for differentiating between groups of subjects.

Figure 2 shows the steps involved in our analysis, from data selection, feature extraction and statistical analysis to differentiating between groups.

RESULTS

With five subject groups, ten pairs of comparisons can be performed (e.g. normal vs PLMBD, normal vs. NFLE, etc). When ANOVA test was performed for the sleep stage-based and CAP-based features individually, we were able to find statistically significant differences in a maximum of six out of the ten group pairs (e.g. [CAP.sub.A(3)/CAP] feature).

After performing the mRMR procedure we were able to rank the features with respect to different subject groups (see Table 1). When the highest ranked feature of Higuchi fractal dimension of total CAP cycle ([CAP.sub.HFD] in Table 1) was used in ANOVA, we see from Figure 3 that the insomnia group wass statistically significant different from all other groups (p-value< 0.001). Using this feature, four out of the ten pairs of groups are statistically significant.

When MANOVA test was performed with increasing number of features (i.e., using the first feature only, then the first and second, etc.), we were able to see statistically significant differences between all five groups (all ten out of the ten group pairs). As the number of features increases, we see that the performance is reduced, most probably due to the small sample size in some of the patient groups.

Group scatter plots using the first two canonical variables obtained from Canonical Analysis for the combinations of first five features from Table 1 are shown in Figure 4. Five clusters with statistically distinct specific centers (p-value < 0.001--not fully visible in the 2D representation). The features set contained nonlinear features extracted from CAP activity of the EEG signal (i.e. Higuchi of total CAP cycle-[CAP.sub.HFD]) as well as non-linear features from the sleep-stages (i.e. Higuchi dimension of REM stage, and sleep stage 1--[nHFD.sub.s4] and [nHFD.sub.S1]) along with the more traditional CAP features (i.e. CAP rate and [CAP.sub.A(3)/CAP)].

DISCUSSION AND CONCLUSIONS

Our single EEG channel-based quantitative approach investigated the possibility to differentiate between four sleep disorders as well as against the healthy condition, based on features relevant to specific characteristics of the EEG, in particular frequency content (median frequency and spectral entropy) and signal complexity (Higuchi fractal dimension), and to specific sleep-related patterns (sleep stages, CAP). In total thirty-two features (candidate quantitative biomarkers) were obtained from PSG records of 93 subjects. To differentiate between the five groups of subjects, ANOVA and MANOVA tests were performed. Univariate analysis showed that only some of the groups were statistically significant different from the other groups. Multivariate analysis that included sleep and CAP-based biomarkers revealed that all five groups of subjects can be differentiated using the combination of five features, that is, [CAP.sub.HFD], [CAP.sub.rate], [CAP.sub.A(3)/CAP], [nHFD.sub.4] and [CAP.sub/A/CAP]. In conclusion, combination of sleep stage and CAP features properly extracted from a single scalp EEG channel can be employed to construct novel quantitative and very effective biomarkers for non-invasive differential diagnosis of sleep disorders. Further analysis with incorporation of machine learning techniques could improve the classification of subjects to a specific sleep condition (disorder or healthy).

REFERENCES

[1] B. Koley, D. Dey, An ensemble system for automatic sleep stage classification using single channel EEG signal. Computational Biology and Medicine, 1. 2012, 1186-95.

[2] W.J. Randerath, B.M. Sanner, V.K. Somers, Sleep Apnea Current Diagnosis and Treatment. Rochester, MN, Karger Science and medical Publishers, 35, 2006.

[3] D.P. White, Sleep apnea, Proceedings of the American Thoracic Society, 3, 2006, 124-128.

[4] T. Young, P.E. Peppard, D.G. Gottlier, Epidemiology of obstructive sleep apnea, a population health perspective, American Journal Respiratory Critical Care Medicine, 165, 2002, 1217-1239.

[5] M.G. Terzano, L. Parrino, A. Sherieri, R. Chervin, S. Chokroverty, C. Guilleminault, M. Hirshkowitz, M. Mahowald, H. Moldofsky, A. Rosa, R. Thomas, A. Walters. Atlas, rules, and recording techniques for the scoring of cyclic alternating pattern (CAP) in human sleep. Sleep Medicine, 2, 2001, 537-553.

[6] A.L. Goldberger, L.A.N. Amaral, L .Glass, J.M. Hausdorff, P.Ch. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.K. Peng, H.E. Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals. Circulation 101(23):e215-e220 [Circulation Electronic Pages; http://circ.ahajournals.org/cgi/content/full/101/23/ e215]; 13, 2000.

[7] L. Parrino, F.D. Paolis, G. Milioli, G. Gioi, A. Grassi, S. Riccardi, E. Colizzi, and M.G. Terzano, "Distinctive polysomnographic traits in nocturnal frontal lobe epilepsy," Epilepsia, 53, 2012, 1178-1184.

[8] L. Parrino, M. Boselli, G. P. Buccino, M. C. Spaggiari, G. Di Giovanni, and M. G. Terzano, "The cyclic alternating pattern plays a gate-control on periodic limb movements during non-rapid eye movement sleep," Journal of Clinical Neurophysiology. Off. Publication American Electroencephalography Society, 13, 1996, 314-323.

[9] B. F. Boeve, "REM sleep behavior disorder: Updated review of the core features, the REM sleep behavior disorder-neurodegenerative disease association, evolving concepts, controversies, and future directions," Annals of the New York Academy of Science, 1184, 2010, 15-54.

[10] T. Roth. "Insomnia: Definition, prevalence, etiology, and consequences". Journal of clinical sleep medicine: JCSM: official publication of the American Academy of Sleep Medicine, 3,2007, S710.

[11] N. V. Thakor, "Quantitative EEG Analysis Methods and Clinical Applications". Artech House Publishers-2009.

[12] J. Fell, J. Roschke, K. Mann, C. Schaffner," Discrimination of sleep stages: a comparison between spectral and nonlinear EEG measures," Electroencephalography and Clinical Neurophysiology, 98, 1996, 401-410.

[13] C. Gomez, A. Mediavilla, R. Hornero, D. Abasolo, and A. Fernandez, "Use of the Higuchi's fractal dimension for the analysis of MEG recordings from Alzheimer's disease patients," Medical Engineering and Physics, 31, 2009, 306-313.

[14] K. Susmakova, A. Krakovska, "Discrimination ability of individual measures used in sleep stages classification," Artificial Intelligence in Medicine, 44, 2008, 261-277.

[15] L. Parrino, F.D. Paolis, G. Milioli, G. Gioi, A. Grassi, S. Riccardi, E. Colizzi, and M.G. Terzano, "Distinctive polysomnographic traits in nocturnal frontal lobe epilepsy," Epilepsia, 53, 2012, 1178-1184

[16] H. J. Keselman, Carl J. Huberty, Lisa M. Lix, Stephen Olejnik, Robert A. Cribbie, arbara Donahue, Rhonda K. Kowalchuk, Laureen L. Lowman, Martha D. Petoskey, Joanne C. Keselman and Joel R. Levin, "Statistical Practices of Educational Researchers: An Analysis of Their ANOVA, MANOVA, and ANCOVA Analyses, " Review Of Educational Research, 1998, 350-386.

[17] H.C. Peng, F. Long and C. Ding, "Feature selection based on mutual information: criteria of max-dependency, max-relevance, and minredundancy". IEEE Transactions on Pattern Analysis and Machine Intelligence, 27, 2005, 1226-1238. doi:10.1109/TPAMI.2005.159. PMID 16119262.

Sai Mohan Rudrashetty, Ashmit Pyakurel, Bharat Karumuri, Rui Liu, Ioannis Vlachos, Leonidas Iasemidis

Biomedical Engineering, Louisiana Tech University, Ruston, LA

Table 1. Effect of increasing the number of features (up to top ten
features) on the number of statistically significant different
group pairs.

Number   Top 10 CAP and sleep           MANOVA P values
         stage-based features

1.       [CAP.sub.HFD]                  5.825 x [10.sup.-9]

2.       [CAP.sub.rate]                 1.74 x [10.sup.-11]

3.       [CAP.sub.A(3)/CAP]             7.9 x [10.sup.-16]

4.       [MATHEMATICAL EXPRESSION      2.0308 x [10.sup.-18]
         NOT REPRODUCIBLE IN ASCII]

5.       [CAP.sub.A]/CAP                1.8 x [10.sup.-18]

6.       [MATHEMATICAL EXPRESSION       2.13 x [10.sup.-17]
         NOT REPRODUCIBLE IN ASCII]

7.       [CAP.sub.A/TST]               2.107 x [10.sup.-16]

8.       [dCAP.sub.A(2)]                4.34 x [10.sup.-18]

9.       [nSE.sub.s4]                   4.79 x [10.sup.-17]

10.      [CAP.sub.SE]                   2.87 x [10.sup.-18]

Number   No. of statistically
         significant different
                 pairs

1.                 4

2.                 6

3.                 8

4.                 9

5.                10

6.                10

7.                 9

8.                 9

9.                 9

10.                9
COPYRIGHT 2015 Mississippi Academy of Sciences
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 2015 Gale, Cengage Learning. All rights reserved.

Article Details
Printer friendly Cite/link Email Feedback
Author:Rudrashetty, Sai Mohan; Pyakurel, Ashmit; Karumuri, Bharat; Liu, Rui; Vlachos, Ioannis; Iasemidis, L
Publication:Journal of the Mississippi Academy of Sciences
Article Type:Report
Date:Apr 1, 2015
Words:2756
Previous Article:Towards the design of a distractive and mobility-enabling back support device.
Next Article:Spatiotemporal dynamics of interictal spikes.
Topics:

Terms of use | Privacy policy | Copyright © 2020 Farlex, Inc. | Feedback | For webmasters