SLEEP 2011 Abstract Supplement

B. Clinical Sleep Science XIV. Instrumentation and Methodology
automated scoring method based on probability analysis from a single
EEG derivation.
Methods: Polysomnograms (F, C and O) spanning a whole night were
recorded from ten healthy subjects and rated by two human scorers every
30 seconds. Discrete Fourier Transform (DFT) was performed on every
second of all the derivations. On every DFT frequency bin (f), probability
distribution (P(s,i,f)) was calculated on each stage (s) of each subject
(i). The contribution rate (C(i,f)) was then calculated on every derivation
of all the subjects. Each pair of probability distribution and contribution
rate of the subject (P(s,i,f) x C(i,f)) was chosen as a classifier, and
sleep stages were estimated every second. The most frequent estimation
within every 30-seconds was selected and compared with the human
scorers’ ratings. We also applied two-dimensional and cluster analyses
on the same signals, and compared the efficacy of the classifiers.
Results: Among the three analyses, we obtained the best results from
the probability analysis. In the probability analysis, the best classifier
worked at more than 80% accuracy on all the polysomnogram derivations.
Although the detection rate of NREM stage exceeded 95% on
most derivations, the rate on Wake and REM stages remained in a lower
range (50-85%). The accuracy and the detection rate of the classifiers
were influenced by many parameters, such as epoch length, frequency
ranges and the contribution function.
Conclusion: Our multi-dimension probability evaluations on human
sleep EEG worked as well as other automated scoring systems. This
simple classifier based on statistical data does not simulate complex decisions
by human scorers, but it can eliminate arbitrary errors. Since this
method does not depend on specific frequencies or wave forms, it can
be easily applied to other biological signals. Using our classifier parameters,
it may also be possible to estimate the EEG features that affect
human scorers’ decisions.
0949
CHARACTERIZATION OF SLEEP MICROSTATES
Carrubba S 1 , McCarty DE 1 , Chesson AL 1 , Frilot C 2 , Marino AA 1
1
Dept. of Neurology, LSU Health Sciences Center-Shreveport,
Shreveport, LA, USA, 2 School of Allied Health Sciences, LSU Health
Sciences Center, Shreveport, LA, USA
Introduction: Nonlinear analysis of EEG signals using the RQA method
quantifies deterministic (nonrandom) changes in brain states of arbitrary
length. RQA has been used to study cognitive processing in normality
and disease, but not to characterize sleep microstates. We sought evidence
indicating RQA’s usefulness for this purpose.
Methods: Polysomnograms scored based on the ASSM manual were
analyzed using published computational procedures (implementing
LabView code available). Employing the RQA variable %R, which
measures amount of law-governed activity whether or not visually perceivable,
we quantified EEGs from frontal, central, and occipital electrodes,
resulting in approximately 25,000 sequential values in successive
1-s intervals (time series). Additionally, phasic events (EEG arousals,
K complexes, spindles, delta bursts) were analyzed by calculating %R
for 100-ms intervals using a step of 2 ms (500 values of %R per sec).
Additional calculations were performed using the RQA variable %D
(independent measure of determinism). Spectral power analysis was
employed as a control procedure.
Results: In each case (6 patients x 6 electrodes), the %R time series consisted
of 3 relative maxima with periods of 200-400 min; as expected,
for each patient, the patterns for all derivations were essentially identical.
The patterns were not detected using power analyses. Visible phasic
events consistently resulted in localized increases in %R; similar changes
occurred in the absence of visually-scored phasic EEG events. Averaged
over 30-s epochs, increased %R correlated strongly with increased
sleep-state depth as assessed by standard scoring (80-90% agreement);
the extent of the agreement was increased when %D was included in the
RQA sleep-stage classification.
Conclusion: Nonlinear EEG analysis reliably permitted quantification
of brain states lasting 0.1-1 s and reproduced the scoring of gold-standard-scored
PSGs, suggesting the possibility of an integrated approach
to the study of brain dynamical changes that includes both background
signals and superimposed phasic changes.
0950
DIGITIZED FEATURES OF PHASIC ACTIVITY OF SURFACE
ANTERIOR TIBIALIS ELECTROMYOGRAPHY (EMG):
VALIDATION BY CONSENSUS PANEL
Fairley J 1,2 , Georgoulas G 2 , Mehta N 2 , Gray A 2 , Trotti L 1 , Wilson A 1 ,
Greer S 1 , Hollars S 1 , Rye DB 1 , Bliwise DL 1
1
Neurology, Emory University School of Medicine, Atlanta, GA, USA,
2
Electrical Engineering, Georgia Tech, Atlanta, GA, USA
Introduction: Phasic EMG activity has been shown to occur at high
rates in synucleinopathic conditions and may also reflect medication effects.
Its quantification using visual analyses is time-intensive. We describe
here an automated approach to such measurement.
Methods: Five scorers evaluated independently 16,200 seconds of
low-noise AT EMG from a PSG containing a moderate amount of phasic
activity, encompassing NREM and REM sleep. Overall X rate of
phasic activity across scorers (% of 1-sec intervals with activity) was
9.3% (range 7.67-12.7%). Signal processing features included: high freq
spectral power (HF), Spectral Edge, Skew, Variance, Kurtosis, Simple
Entropy, Mobility, 75th % Amplitude, Complexity, Absolute Amplitude,
Curve Length (CL), Energy, Zero Crossing (ZC), Non-Linear Energy
(NLE) and Spectral Entropy (SENTRO). Signals were digitized at 200
Hz and were acquired with Embla N7000 PSG.
Results: Consensus agreement among 5 scorers for presence or absence
of phasic activity within 1-sec segments was very high (% agreement =
93.52%; Kappa = .818, p < .0001). All examined digitized features differentiated
phasic present vs absent (nearly all p’s < .0001) seconds. We
also compared expert consensus agreement (binary) versus continuous,
Gaussian distributions of each feature using binary classification models
derived from MATLAB. Data showed wide divergence for features,
with highest True Positive/True Negative Rates for SENTRO, NLE, CL
and Variance (all > 88%), lowest for HF (65%), and ZC (71%) and other
features intermediate. Among consensually agreed upon seconds with
phasic activity (N = 1090), coefficients of variation were highest for
SENTRO and NLE, approximating a 100-fold difference from the poorest
functioning features.
Conclusion: Experts can agree on the presence or absence of phasic
EMG events in sleep at rates far exceeding chance. Parametric analyses
suggest nearly all digitized features make this discrimination well, but
classification optimization may favor novel non-linear approaches.
Support (If Any): NS-050595; RR-025009 (ACTSI); NS-055015-03S1
0951
TOP-DOWN INTEGRATED SLEEP-STATE BIOMARKERS IN
PEDIATRIC BRAIN DISORDERS
Kothare SV 1 , Wood C 2 , Mietus J 3 , Thomas RJ 2
1
Neurology, Children’s Hospital, Harvard Medical School, Boston,
MA, USA, 2 Medicine, Beth Israel Deaconess Medical Center, Boston,
MA, USA, 3 Interdisciplinary Medicine & Biotechnology, Beth Israel
Deaconess Medical Center, Boston, MA, USA
Introduction: During non-rapid eye movement (NREM) sleep, the
scalp electroencephalogram (EEG) in the 0.5 to 4 Hz (delta) range reflects
dynamic changes within cortical activity. The sleep spectrogram
is an EEG-independent, electrocardiogram (ECG) - derived method to
map the coupling of heart rate variability and respiration-driven ECG-
QRS amplitude fluctuations. High frequency coupling (HFC), a proposed
cardiopulmonary (CPC) spectrogram biomarker of “effective”
sleep, correlates with slow wave power across the night, providing a
new metric (Cortico-CPC, a correlation metric) of cortical modulation
A325
SLEEP, Volume 34, Abstract Supplement, 2011