Detection of cyclic sleep phenomena using instantaneous heart rate

Electroencephalography and Clinical Neurophysiology, 40 (1976) 169--177

169

© Elsevier Scientific Publishing Company, Amsterdam -- Printed in The Netherlands

DETECTION OF CYCLIC SLEEP PHENOMENA USING INSTANTANEOUS HEART RATE * M.J. LISENBY, P.C. RICHARDSON and A.J. WELCH

Bio-Medical Engineering Program, College of Engineering, The University of Texas, Austin, Texas (U.S.A.) (Accepted for publication: September 8, 1975)

Since the advent of the polygraph and the advancement of the digital computer, many researchers have expended efforts toward developing new automated methods for the quantitative analysis of sleep patterns (Martin et al. 1972; Smith and Karacan 1972; Gaillard and Tissot 1973; Naitoh et al. 1973). Characteristic sleep patterns have been analyzed by such techniques as spectral analysis (Johnson et al. 1969), period analysis (Welch 1971), and baseline cross analysis (Itil et al. 1969) of the electroencephalogram or EEG. Typically, these methods involve seeking out definitive measures of the EEG (usually frequency components) and then developing either an analog, digital, or hybrid algorithm capable of using these measures for classifying patterns into one of several defined categories. In recent years, several researchers in our Bio-Medical Engineering Heart Rate Laboratory have conducted a number of projects dedicated toward development of automated methods for the classification of sleep ~atterns (Aldredge and Welch 1973; Weber e~ al. 1973; Welch and Richardson 1973). However, our primary goal was to develop a sleep pattern detection process using an easily derived physiologic parameter coupled with a rapid, inexpensive algorithm for quantitative analyses. Instead of the typically used EEG, we have chosen beat-by-beat heart rate as our

* This research was sponsored by U.S. Army Medical Research and Development Command, Contract No. DAMD17-74-C-4081.

physiologic parameter. Although the validity of heart rate as a sleep detector has been disputed among researchers, it is well known that heart rate does exhibit concomitant characteristic changes with change in sleep depth (Brooks et al. 1956; Dement and Kleitman 1957; Snyder et al. 1964; Bond et al. 1973). It is also agreed that phasic change is the most pronounced heart rate variable (Brooks et al. 1956; Snyder 1967). We chose beat-by-beat heart rate as our criteria because the ECG is an easily recorded physiologic parameter which is more in keeping with our desire for low-cost and limited data bulk. Regardless of the physiologic parameter chosen, the m e t h o d of analysis must achieve a certain degree of consistency in its results which is independent of the sleep subject being analyzed. In the past, lack of consistency between nights of sleep for a given subject and between subjects has been cited as a major obstacle confronting quantitative methods of analysis (Itil et al. 1969; Lubin et al. 1969; Roessler et al. 1970). Two of our previous studies conducted by Aldredge and Welch (1973) and Welch and Richardson (1973) investigated beat-by-beat heart rate as a sleep detection parameter. The important findings of these studies were as follows: (1) beat-by-beat heart rate did contain sleep pattern information; (2) heart rate patterns were not the same toward the end of the night as they were near the onset of sleep; (3) heart rate patterns differed from sleep cycle to sleep cycle (where a cycle is defined as the approximately 90 min normal r h y t h m

170

of oscillation from light to deep sleep); and (4) that the rapid eye movement (REM) periods appeared to set the trends of the deep-sleep heart rate patterns that followed. It seemed logical that if the REM periods effected the succeeding heart rate patterns, then for heart rate analysis a sleep cycle should be defined as beginning at the onset of a REM period and ending with the onset of the next REM period. (Classically, the sleep cycle begins with a non-REM period and ends with the onset of the next non-REM period.) We also hypothesized that since heart rate patterns differed from cycle to cycle each cycle should be modeled separately. In a third study conducted by Weber et al. (1973) we found that techniques in spectral analysis of beat-by-beat heart rate could be used to detect transient oscillations which occurred concurrently with eye movements during the REM periods. Since we now feel that it is important to model sleep cycles separately, we believed that the m e t h o d devised by Weber et al. (1973) might provide us with a computer-aided process for separating cycles using only heart rate as the criterion. The purpose of this paper is to report the results of our efforts to devise a computerbased spectral analysis of beat-by-beat heart rate for separating the REM and non-REM (NREM) states. If applicable, this method would then serve as c o m p o n e n t procedure in automating the detection of individual sleep stages using only heart rate. Our efforts were directed toward modeling phenomena which were characteristic of most subjects in order to improve consistency among data.

Method Beat-by-beat heart rate was determined by measuring the time in milliseconds between successive R-waves on the electrocardiogram. The variation in these R-to-R intervals from one beat to the next was the basis for what we called the "Heart Beat Domain" (Fig. 1). The independent variable in this domain was

M.J. L I S E N B Y ET AL, HEART BEAT DOMAIN

9

R-R 7iM~_ INTERV~I (rnsec)

i

,,

~'

r

I

L

?e

79

,

o'C 4

I

60

~1

110 6 4

|20

BEATS

Fig. 1. T h e H e a r t Beat D o m a i n . T h e b e a t s or h e a r t b e a t n u m b e r s are equally spaced at integer values of 1, 2, ... 128.

heart beat number in any given minute of data. The dependent variable was the magnitude or duration of the corresponding R-to-R interval. It should be noted that the data of the independent variable were discrete and equally spaced in the Heart Beat Domain. Each 1 min epoch was zero-filled to 128 beats in order to facilitate use of our Discrete Fourier Analysis (FFT) routine (Bergland 1969). One min epochs of R-to-R interval data for a known level of sleep were converted to the Heart Beat Domain and then subjected to averaging techniques and Fourier analysis. The result was a template spectrogram in what we labeled the "Beatquency Domain". Rather than cycles/sec as in the conventional frequency domain, our Beatquency Domain was expressed in terms of normalized amplitude versus cycles/beat. "Normalized" amplitude indicated that spectrum points had been normalized such that all amplitudes lay within the range 0--1. This allowed direct comparison of spectrograms derived from different levels of sleep. In all, we performed Beatquency Domain analysis for each stage of sleep (Stages awake, 1, 2, 3, 4 and REM) on 9 normal subjects and 2 complete nights for each subject. From these analyses we concluded the following: The typical Beatquency Domain spectra for Stages awake, 1, and REM were significantly different from those of Stages 2, 3 and 4 and could be used as criteria for distinguishing between heart rate patterns of these two groups. The data base used in this research was the same as that used in our previous studies

CYCLIC S L E E P P H E N O M E N A A N D H E A R T R A T E

(Aldredge and Welch 1973; Welch and Richardson 1973). Analog recordings were made on 9 normal subjects over 2 complete nights of sleep for a total of 18 nights of recording as described elsewhere (Welch~e~ al. 1970). Each night was hand-scored on a minute-by-minute basis by three trained experts working independently and using modified Dement--Kleitman critera (Welch et al. 1970). Inter-rater agreement was always better than 90% over the 18 nights of sleep.

Results

For all subjects and for both nights the 1 min epochs of R-to-R intervals were grouped separately according to the hand-scored sleep stages: Stages 0, 1, 2, 3, 4, REM. Stages 0 indicated the awake state. These grouped epochs were converted to the Beatquency Domain to produce spectra representative of each stage of sleep. Fig. 2 illustrates the spectra obtained from one subject (subject SCH). The solid curves represent Night 1 data while the dotted curves represent Night 2 data. As can be seen from Fig. 2 the spectra of the individual stages fell into two morphologically different groups. The first group consisted of spectra of Stages 0, 1 and REM and was characterized by a maximum amplitude peak in the "low b e a t q u e n c y " range and then an exponential-like decay in amplitude with increasing beatquency. The second group was characterized by a high beatquency peak around 0.25--0.35 c/beat and was typical of Stages 2, 3 and 4 spectra. In all subjects the Stage 0 or awake spectrum closely resembled that of Stages 1 and REM. The spectra shown in Fig. 2 were typical of most subjects. It is known that Stages 1 and REM are essentially the same with regard to the EEG patterns (Dement and Kleitman 1957). REM, however, indicates the occurrence of rapid eye movements concurrent with the Stage 1 EEG. Also Stages 2, 3 and 4 are known to contain no rapid eye movement and thus are

171

considered as NREM stages of sleep. Therefore, except for Stage 0, the Beatquency Domain spectra fell into respective REM, 1 (combined REM and Stage 1) and NREM groups. Although Stage 0 (awake) is not typically considered as REM or NREM we chose to include Stage 0 in the REM, 1 group due to its similarity in the Beatquency Domain. To distinguish our definition of REM from the classical definition, we labeled combined Stages 0, 1 and REM as REM+. By pooling the heart rate data of each of the two groups, we produced REM+ and NREM spectra as shown in Fig. 3. These spectra served as templates for our classification procedure. In order to measure intra-subject and intersubject consistency of our analysis we calculated pairwise correlations between the REM+ and NREM template spectrograms. By doing this we were able to determine if the characteristic features of the templates were retained from one night to the next and from one subject to the next. We found the REM+ spectra of Nights 1 and 2 to correlate better than 98% on the average over the 9 subjects. Similarly, the intra-subject correlations of the NREM spectra averaged better than 91%. On an inter-subject basis, pairwise correlations of REM+ spectra from different subjects averaged better than 90%. However, the corresponding correlations of the NREM spectra averaged only 36%. The major discrepancies in the NREM spectra were found to lie in the varying amounts of low beatquency activity the subjects produced, while the high beatquency information tended to be quite consistent. Classification of our sleep data into their respective REM+ and NREM components involved two steps. The first step was to use the Night I data of each subject as the training set. In other words, the REM+ and NREM templates were derived from first night data. Night 2 data served as the test set. The second step was to compare a subject's templates with a beatquency spectrum representative of

172

M.J. LISENBY ET AL

SCH- STAGE 0

S C H - STAGE 2

1.00'

I'00 / bl.I 0.804 r., :3 ~- o.60J

0.80

w Q

0.60I%

..J

~ 0.40"

~E0.40<~

0,20. 0.00 0.00

0.20.

0.I0

0,20

0.'50

0.00 0.00

0.40 0.50

CYCLES / BEAT

SCH - STAGE

3

0"801

I.u r-,

0.60-~ %,

~

%x

/

.J o,. 0.40.

1%\

\

,i'\

=E

I

0,20-

0.20"

0,00

0.00 0.10

0.20

0.30 0,40 0.50

0.00

CYCLES/BEAT

0.10

0.20

0.30

0.40

0.50

CYCLES / BEAT

SCH -REM

SCH - S T A G E

1.90'

4

r.o0

tU 0,80

0.80, I.,1.1

I,- 0.60

:~ 0 . 6 0 ' I--J

_.1

Q" 0.40 =E ¢t

Q- 0.40. ~E

0.20•

0.20

0.00 0.00

0.50

1.00,

=E

0.00

0.40

SCH - STAGE

I

0.80. hi 0 D 0.60~ _5

0.20 0.30

CYCLES/BEAT

1.00

(l,, 0.40-

0.10

0.00 0.10

0.20

0.30

CYCLES/BEAT

0.40 0 . 5 0

0.00

0.10

0.20

CYCLES/

0.30 0.40 0.50 BEAT

Fig. 2. E x a m p l e s o f stage s p e c t r a ( s u b j e c t ID = SCH) in t h e B e a t q u e n c y D o m a i n . Solid curve = Night 1; d o t t e d curve = Night 2.

173

CYCLIC S L E E P P H E N O M E N A A N D H E A R T R A T E TABLE I

SCH - R E M +

S u m m a r y o f t h e results o f t h e R E M - - N R E M classification procedure.

I.O0" O80a

D 060-

Subject

Percent accuracy

Correlation coefficient

SAFI* SAF2**

89.8 81.1

0.794 0.616

SCH1 SCH2

84.9 88.7

0.644 0.784

FAR1 FAR2

84.9 84.1

0.669 0.674

CHI1 CHI2

89.4 86.9

0.782 0.720

MOS1 MOS2

78.8 83.8

0.584 0.685

GILl GIL2

83.4 72.2

0.653 0.508

NOR1 NOR2

73.9 62.5

0.435 0.332

PHI1 PHI2

79.8 75.1

0.583 0.564

PAD1 PAD2

80.1 71.5

0.623 0.414

I--

~-J040 ~E <% o 20

0 O0 0 l 00

0 .I0

0.20

0.30

0.40

0.50

CYCLES/BEAT

SCH - N R E M I.O0 I Lu 0-80~ o ~D ~" 0 , 6 0 " ~E 0.40,

0.20

0.00 0.00

O. IO 0.20

0.30 0.40 0.50

CYCLES/BEAT

Fig. 3. E x a m p l e s o f R E M + a n d N R E M B e a t q u e n c y D o m a i n t e m p l a t e s ( S u b j e c t ID = SCH). Solid cure = N i g h t 1 ; d o t t e d c u r v e = Night 2.

* Night 1 ** N i g h t 2

each sequential minute of data throughout both nights of sleep. If this spectrum more closely resembled (resemblance of spectra were measured via calculation of correlation coefficients) the REM+ template then that 1 min epoch was classified as REM+. Similarly, if the test spectrum more closely resembled the NREM template it was classified as NREM, and so on for each 1 min epoch. After each minute was classified as R E M + or NREM a weighted triangular window was used to smooth the results to coincide better with typical sleep patterns. The results of this classification procedure for the subject whose spectra were given earlier is presented graphically in Fig. 4. This figure depicts both the hand-scored progres-

sion of REM+ and NREM and our automated predicted progression. Table I is an overall summary of performing these analyses on each night of our 9 subjects. The " 1 " and " 2 " after each subject label indicated Nights 1 and 2 respectively. The percent accuracy represents the percentage of the total number of 1 min epochs classifted correctly as either REM+ or N~tEM, while the correlation coefficient indicates the similarity in the time progressions of the handand predicted scores. The range of accuracies was 73.9--89.8% with an average of 82.8% for the training nights and a range of 62.5--88.7% with an average of 78.4% for the test nights. The results of the REM+--NREM classifier

174

M.J. LISENBY ET AL. REM-NREM TRAINING

o

eb

~zo

CLASSIFICATION

N I G H T --

leO

SUBJECT SCHI

z4o

3bo

T I M E , (MINS),

HANO

0

6o

i=o

SCORES

,So

z40

300

330

420

4eo

TIME ,(MINS)

PREDICTED SCORES RiEM+ - STAGES 0 ,I , 5 NREM- STAGES :) ,3,4

PREDICT'ION ACCURACY, 84.9 % PREDICTION CORRELATION = 0.644

o

~o

TEST

NIGHT

I;~0

180

-

240 300 TIME,(MINS)

HAND REM+ 1 NREM i O

6o

n

,io

leo

SUBJECT

i

SCH 2

360

420

480

4"zo

~o

SCORES

z;4o

TIME,

3o0

rl

3so

(MINS)

PREDICTED

SCORES

PREDiCTiON ACCURACY. 88.7 Ore PREDICTION CORRELATION = 0.784

REM+--SlrAGES NREM--STAGES

0,1 ,S 2,3,4

Fig. 4. Graphical results o f R E M - - N R E M classification o n a m i n u t e - b y - m i n u t e basis. H a n d scores were derived f r o m t h e EEG data.

were then submitted to a cycle detector algorithm. Although this algorithm has not yet been optimized, preliminary results on all 9 subjects indicate significant success. Table II represents the results of applying the cycle detector to the data of subject SCH.

Discussion Heart rate patterns during sleep differ from sleep cycle to sleep cycle. Therefore, based on

past experience in using heart rate to classify sleep patterns, we found it necessary first to separate cycles and then model the cycles separately. The computer-based algorithm outlined in this report was designed to separate sleep cylces by detecting the occurrence of REM periods which denote the beginnings of cycles. However, the results of the Beatquency Domain analyses indicated that the spectral content of Heart Beat Domain signals of REM and Stages 1 and 0 (awake) were indistinguishable. Stage 1 is often combined

CYCLIC S L E E P P H E N O M E N A A N D H E A R T R A T E

175

T A B L E II C o m p a r i s o n o f h a n d - s c o r e d a n d p r e d i c t e d cycle d e t e c t i o n s for s u b j e c t SCH. S t a r t a n d e n d i n d i c a t e m i n u t e n u m bers, while No. o f m i n i n d i c a t e s cycle lengths. Cycle No.

Hand-scored

Predicted

Start

End

No. of rain

Start

End

No. o f m i n

N~h t 1 0 1 2 3 4

1 95 206 284 383

94 205 283 382 450*

94 111 78 99 68

12 94 206 278 439

93 205 277 438 450*

82 112 72 161 12

N~h t 2 0 1 2 3 4

1 94 179 263 367

93 178 262 366 450*

93 85 84 104 84

1 94 176 266 367

93 175 265 366 450*

93 82 90 101 84

* E n d o f night.

with REM; however, the awake state is usually defined separately. Nevertheless, other researchers agree that the REM and awake states do exhibit some similar characteristics (Aserinsky 1967; Spreng et al. 1968). Also, since our purpose was to distinguish sleep cycles, we felt that long periods of arousal should also be used to denote beginnings (or ends) of a cycle. Although considerable thought has been devoted to determining the physiological meaning of the peaks in the Beatquency Domain spectra, we feel that any attempts of explanation at this time would only be speculative. Considering that beat-by-beat heart rate was our only criteria we feel that our REM+-NREM classification model performed with considerable success. In general the algorithm classified the major periods of REM+ correctly and at the correct time as evidenced by the predictor correlation coefficients. Even though in most subjects accuracies better than 60% could be obtained by simply classifying the entire night as NREM, this procedure would result in very low predictor correlations and would produce errors all of one

type. On the contrary, the errors tended to be evenly distributed between misclassifications of REM+ and NREM in most subjects. Another measure of our success was the high degree of consistency found among the Beatquency Domain spectra. The characteristic features of the REM+ and NREM groups were maintained both in intra-subject and inter-subject comparisons as evidenced by correlation analysis. Although comparison of the NREM spectra among the subjects produced low correlations, the characteristic high beatquency peak was distinguishable in all subjects. With regard to overall efficiency of the algorithm, it must be remembered that the model is designed only to aid in separating sleep cycles, thereby having only to detect correctly the major periods of REM+. The minute-by-minute accuracy is of more consequence in algorithms designed to distinguish sleep stages. Automated detection of the beginnings of sleep cycles were performed by detecting the well defined transitions from the NREM to the REM+ states. Although visual examination of the EEG patterns for cycling produced

176

different results as to the exact time of night a particular cycle begins, cycle detection from our predicted o u t p u t s performed with reasonable confidence. We feel that the development of Beatquency Domain analysis has provided us with a new and useful tool in achieving our goal of sleep evaluation based on heart rate measures. It is also a tool that exhibits remarkable intrasubject and inter-subject consistency, a most desirable characteristic of any detection parameter. Summary The development of the Heart Beat Domain and the Fourier transform of the Heart Beat Domain (which we call the Beatquency Domain) has provided new and useful tools for the quantitative analysis of sleep level patterns. This m e t h o d of analysis has produced remarkable intersubject as well as intra-subject consistency and the only physiologic parameter required in the analysis is beat-by-beat heart rate. This analytical tool was designed to aid in the detection of sleep cycles, or more specifically, the rhythmic transitions from REM+ (awake Stages I and REM combined) to NREM (Stage 2, 3 and 4 combined) over a normal night of sleep. Employing this method on minute-by-minute sleep recordings from 9 normal sleep subjects, 2 complete nights each, we were able to distinguish between the REM+ and NREM stages with an average accuracy of approximately 80%. Considering that beat-by-beat heart rate was our only criteria, we felt that the algorithm performed with significant success. R6sum6

Ddtection des phdnomdnes cycliques du sommeil "a l'aide de la frdquence cardiaque instantande Le d~veloppement du domalne des battements cardiaques et la transformation de

M.J. LISENBY ET AL.

Fourier de ce domaine (que nous appelons le "Beatquency Domain") a fourni des outils nouveaux et utiles pour l'analyse quantitative des patterns de niveau de sommeil. Cette m~thode d'analyse aboutit ~ une constance remarquable inter-sujets aussi bien qu'intrasujets et le seul param~tre physiologique n~cessaire ~ cette analyse est la fr~quence cardiaque entre deux battements. Cet outil d'analyse est destin~ ~ aider ~ la d~tection des cycles de sommeil, ou de faqon plus sp~cifique, des transitions rythmiques du sommeil REM+ (combinant l'~veil, le stade I e t le stade REM) au non-REM (combinant les stades II, III et IV) t o u t au long d'une nuit normale de sommeil. A l'aide de cette m~thode sur les enregistrements de sommeil minute par minute chez 9 sujets normaux, ayant ~t~ enregistr~s chacun deux nuits completes, nous avons ~t~ capables de distinguer entre stades REM+ et non-REM avec une precision m o y e n n e d'environ 80% Consid~rant que la fr~quence cardiaque entre deux battements ~tait notre seul crit~re, nous pensons que cet algorythme a fonctionn~ de mani~re satisfaisante.

References Aldredge, J.L. and Welch, A.J. Variations of heart rate during sleep as a function of the sleep cycle. Electroenceph. clin. Neurophysiol., 1973, 35: 193--198. Aserinsky, E. Physiological activity associated with segments of the rapid eye movement period. In Sleep and altered states of consciousness, Vol. 45. Williams and Wilkins, Baltimore, Md., 1967, Ch. 16. Bergland, G.D. A guided tour of the fast fourier transform. IEEE Spectrum, July, 1969: 41--52. Bond, W.C., Bohs, C., Ebey Jr., J. and Wolf, S. Rhythmic heart rate variability (sinus arrhythmia) related to stages of sleep. Condit. Reflex, 1973, 8: 98--107. Brooks, C.McC., Hoffman, B.F., Suckling, E.E., Kleyntjens, F., Koenig, E.H., Coleman, K.S. and Treumann, H.J. Sleep and variations in certain functional activities accompanying cyclic changes in depth of sleep. J. appl. Physiol., 1956, 9: 97--104. Dement, W.C. and Kleitman, N. Cyclic variations in EEG during sleep and their relation to eye move-

CYCLIC SLEEP PHENOMENA AND HEART RATE ments, body motility, and dreaming. Electroenceph. clin. Neurophysiol., 1957, 9: 678---690. Gaillard, J.M. and Tissot, R. Principles of automatic analysis of sleep records with a hybrid system. Comput. bio-med. Res., 1973, 6: 1--13. Itil, T.M., Shapiro, D.M., Fing, M. and Kassebaum, D. Digital computer classifications of EEG sleep stages. Electroenceph. clin. Neurophysiol., 1969, 27 : 76--83. Johnson, L., Lubin, A., Naitoh, P., Nute, C. and Austin, M. Spectral analysis of the EEG of dominant and non-dominant alpha subjects during waking and sleeping. Electroenceph. clin. Neurophysiol., 1969, 26: 361. Kleitman, N. Sleep and wakefulness (Rev. ed.). Univ. of Chicago Press, Chicago, 1963. Larsen, L.E. and Walter, D.O. On automatic methods of sleep staging by EEG spectra. Electroenceph. clin. Neurophysiol., 1970, 28: 459--467. Lubin, A., Johnson, L.C. and Austin, M.T. Discrimination among states of consciousness using EEG spectra. Psychophysiology, 1969, 6: 122-131. Martin, W.B., Johnson, L.C., Viglione, S.S., Naitoh, P., Joseph, R.D. and Moses, J.D. Pattern recognition of EEG-EOG as a technique for all-night sleep stage scoring. Electroenceph. clin. Neurophysiol., 1972, 32: 417--427. Naitoh, P., Johnson, L.C., Lubin, A. and Nute, C. Computer extraction of an ultradian cycle in sleep from manually scored sleep stages. Int. J. Chronobiol., 1973, 1: 223--234.

177 Roessler, R., Collins, F. and Ostman, R. A period analysis classification of sleep stages. Electroenceph. clin. Neurophysiol., 1970, 29: 358--362. Smith, J.R. and Karacan, I. EEG sleep stage scoring by an automatic hybrid system. Electroenceph. clin. Neurophysiol., 1972, 31 : 231--237. Snyder, F. Autonomic nervous system manifestations during sleep and dreaming. In Sleep and altered states of consciousness, Vol. 45. Williams and Wilkins, Baltimore, Md., 1967, Ch. 20. Snyder, F., Hobson, J.A., Morrison, D.F. and Goldfrank, F. Changes in heart rate and systolic blood pressure in human sleep. J. appl. Physiol., 1964, 19: 417--422. Spreng, L.F., Johnson, L.C. and Lubin, A. Autonomic correlates of eye movement bursts during stage REM sleep. Psychophysiology, 1968, 4: 311--323. Weber, F.J., Welch, A.J., Vogt, F.B. and Richardson, P.C. Detection of REM, 1 sleep stage and eye movement from beat-to-beat heart rate. Tech. Rep. 107, Electronics Research Center, University of Texas at Austin, Austin, Texas, 1973. Welch, A.J. Period analysis of space flight EEG. Aerospace Med., 1971, 42: 601--606. Welch, A.J. and Richardson, P.C. Computer sleep stage classification using heart rate data. Electroenceph, clin. Neurophysiol., 1973, 34: 145--152. Welch, A.J., Richardson, P.C., Thomas, C.W. and Aldredge, J.L. Bandwidth reduction of sleep information. Tech. Rep. 92, Electronics Research Center, University of Texas at Austin, Austin, Texas. 1970.