c STAGE, Automated Sleep Scoring: Development and

Transcription

1 Sleep, 17(8): American Sleep Disorders Association and Sleep Research Society J c STAGE, Automated Sleep Scoring: Development and Comparison With Human Sleep Scoring for Healthy Older Men and Women *tpatricia N. Prinz, *tlawrence H. Larsen, tkaren E. Moe, teric M: Dulberg and tmichael V. Vitiello *Department of Veterans Affairs Medical Research Service, American Lake VAMC, Tacoma, Washington, U.S.A.; and tdepartment of Psychiatry and Behavioral Sciences, University of Washington, Seattle, Washington, U.S.A, Summary: Using the sleep records of200 men and women (age years), we have developed a human-assisted computer scoring system, C STAGE. The system can have many applications, including quantitative electroencephalographic (EEG) analysis during specific stages of sleep. C STAGE classifies sleep/wake stages using power spectral analysis and other techniques applied to one channel of EEG data, Here we report comparability data between C ST AGE- and human-rated sleep-stage scoring using Rechtschaffen and Kales criteria for 70 normal subjects (a subset of the 200), Because the method was developed using these subjects, we also report comparability data for an independent validation sample of 45 normal older men and women. For waking measures, sleep stages 3 and 4, and total sleep time, C STAGE yielded ratings comparable with the human rater (r = ; P < 0.001). For sleep stages 1 and 2 and REM sleep, C STAGE correlated less well with human ratings (r = ; p < 0.00 I). Overall, these correlations compare well with other currently available computer stage-scoring methods. Epoch-by-epoch comparisons in the validation sample revealed a mean proportion of agreement of 0.74 and a mean Kappa coefficient of 0.57, indicating the two methods provide reasonable agreement on an epoch-by-epoch basis. We conclude that C STAGE is a valid sleep/waking scoring system for healthy older adults. Key Words: Automated sleep scoring-validation studies-?? The typical method of studying sleep involves recording and reduction of massive amounts of electrophysiological information, including the electroencephalogram (EEG), electrooculogram (EOG) and electromyogram (EMG), into stages of sleep by human visual scoring. This method, although still considered the gold standard, is costly, time consuming and emphasizes sleep as a static series of stages rather than a continuous variation (1). In the last 10 years a number of automated sleep EEG analysis packages have been introduced (e.g. 2-6). C STAGE is an automated, human-assisted, sleepscoring system developed in our laboratory based on over 200 records that had previously been rated by humans (7,8). This sample included subjects with Alzheimer's disease, major depressive disorders and multi-infarct dementia who were otherwise healthy, as well as healthy normal controls. Their ages ranged from 50 Accepted for publication July Address correspondence and reprint requests to Patricia N. Prinz, Ph.D., Department of Psychiatry and Behavioral Sciences, RP-IO, University of Washington, Seattle, WA 98195, U.S.A. 711 to 90 years. Our aim in developing C STAGE was to optimize quantification of sleep EEG delta activity and EEG spectral features during tonic rapid eye movement (REM) sleep. Here we report on C STAGE's validity against human sleep rating for a subgroup of 70 control subjects. We also report on validation of the C STAGE methodology using an independent sample of 45 healthy elderly men and women, including an epoch-by-epoch comparison of CST AGE with human sleep scoring. Subjects METHODS All subjects were community-recruited volunteers and were extensively screened for physical and psychological health by clinical interview, physical examination and laboratory screening tests. All subjects reported normal sleep patterns, were nonsmokers and were not taking any central nervous system-active medications for at least 2 weeks preceding their participation in the study (7-9).

2 712 P. N. PRINZ ET AL. Development subsample. For the development stage we used 70 healthy older (55-82 years, x = 67.7 ± 6.7 years; 51 % females) control subjects who participated in a larger study (7-9) ofthe sleep EEG in normal aging, Alzheimer's disease and major depression. Validation sample. For validation studies we used 28 healthy older men (65.9 ± 4.6 years, mean ± SD) and 17 healthy older women (67.1 ± 4.3 years) participating in a study of the effects of fitness training on the subjective and objective sleep quality of older adults (10). Procedures Sleep recording. Data reported here are from the 2nd night of a 3-night baseline study at the Clinical Research Center, University of Washington Medical Center. Night 1 was an adaptation and apnea/myoclonus screening night. None of the subjects reported here manifested evidence of significant levels of sleep apnea or nocturnal myoclonus. All subjects went to bed at their customary bedtimes, determined by I-week sleep log data, and slept until their customary risetime or until spontaneous awakening. Sleep recordings, including EEG, EOG and EMG, were performed following standard procedures (11). EEG electrodes were positioned for conventional sleep recordings at C3, C4, 01 and 02 (international system of measurement) and were referenced to the contralateral mastoids. Data were recorded using a Grass 8-24 polygraph with filter settings of 0.1 and 35 Hz (development sample) or 0.1 and 70 Hz (validation sample). Two or more channels (C3, C4, 01 and 02) were simultaneously digitized. The sampling rates were 128 Hz (development study) and 256 Hz with data averaging and decimation (validation sample), using a 12-bit digitizer installed in an microcomputer and a voltage range of ±2.5 V. All recordings were calibration corrected. Notch filters were not used or needed due to the well-shielded and -grounded recording environment. Human scoring. All paper records were scored by a single, reliable (r = 0.86, p < 0.01), highly trained human rater using standard techniques (11). All scoring criteria were strictly applied, including the 75-/-tV amplitude criterion for scoring stages 3 and 4 sleep. e STAGE development. C STAGE uses the data from a single EEG channel (C4) to score sleep in 16-second epochs. We used C4 rather than C3 because of our other interests in EEG diagnostics based on the nondominant hemisphere. Before any analyses were conducted, robust filter-smoothers were employed to minimize statistical outliers due to artifacts, including spikes, electrocardiogram (EKG) and muscle (12,13). Individual autoregressive (A-R) filters used to mini- mize these and other sources of spectral contamination included 1) a second-order A-R filter with three iterations to the full 16-second epoch to remove spikes, 2) a first-order A-R filter with three iterations applied to each 2-second segment of each epoch to remove EKG artifacts, and 3) a seventh-order A-R filter with up to three iterations applied to each 2-second segment of each epoch to attenuate sharp high frequency transients. Before computing the fast Fourier transform (FFT), the data were multiplied by a 10% cosine bell taper using the Tukey-Hanning window to minimize end effects on the spectral estimates (14). An FFT was then performed and spectral energies (!LV2/Hz) were saved for each 16-second epoch of the record for the frequency bins (0.5-2, 2-4, 4-6, 6-8, 8-9, 9-12, 12-16, 16-18, 18-20, , , 25-30, 30-35, 35-40, 40-45, 45-50, and Hz). Mean bin values for each nine-epoch sliding window average are calculated, taking care to exclude artifactual epochs. These smoothed epochs are then parsed based on these binned energies, using a series of five decision steps that assign each epoch to wake, movement or one ofthe five conventional sleep stages. Figure 1 shows an all-night plot of human-rated sleep stages as well as each of the five steps in the C STAGE process. The criteria applied in these steps were empirically derived to yield stage scoring similar to that obtained by traditional human scoring, with special emphasis on optimizing correspondence to human-rated tonic (but not phasic) REM sleep. Step 1 (Fig. I)-Obvious nonsleep (wake) epochs are identified. Wake is identified by applying a series of criteria empirically found to characterize wake. These criteria include lower delta or theta energies, higher theta to delta ratios (0.5-2/4-6 Hz or 4-6/6-8 Hz), excessive Hz energies and high alpha (9-12 Hz) energies. Step 2-0bvious nonrapid eye movement (NREM) sleep (stages 1-4) epochs are identified. For all epochs not previously defined as wake, average energies and their standard deviations are calculated for the , 2-4- and 8-12-Hz frequency bands. Epochs within ranges of values corresponding to sleep stages 1-4 are classified into the appropriate NREM sleep stage. If classification criteria are not met, epochs remain unclassified. This step yields definitive classification of stages 3 and 4 sleep and provisional classification of stages 1 and 2 sleep. Stage 3 is scored when the delta band (0.5-4 Hz) has root mean square (rms) amplitude in excess of 36.4!LV and stage 4 is scored when the signal has an rms amplitude in excess of49.5!lv. These correspond to values of 650 and 1,200 power units in the binned EEG spectra. Step 3 - Likely tonic REM sleep epochs are identi- Sleep. Vol. 17, No.8, 1994

3 C STAGE: DEVELOPMENT AND VALIDATION 713 J z..: ::;; I 7..: I..: 8..: I- OJ I I- 9..: I- OJ OJ 10 d!/l ::;; 11..: OJ CI I'.',, '\.Y J'ti'f'v."\.!./"v..v;A ".A 1llIltllIllIl.1,II', _ '< HOURS FIG. 1. Steps in C STAGE rating of sleep/wakefulness. 1. From all epochs across the night's record, identify nonsleep. 2. Identify obvious NREM sleep. 3. Identify potential tonic REM. 4. Check assignments, reassign and polish. 5. Human editing and adjustments. These can be compared with: 6. human-scored sleep/wake stages, 7. alphaenergy(8-12 Hz), 8. theta ratio (6-8 Hz/4-6 Hz), 9. beta energy ( Hz), 10. muscle band energy (40-45 Hz) and II. delta energy (0.5-4 Hz). tied. The average energies and standard deviations for the , 6-9- and Hz frequency bins are calculated for all remaining unclassified epochs. Applying a set of criteria based on these quantities identifies REM epochs and additional stage 2 epochs. Following this initial classification, the remaining unclassified epochs are again examined and classified as either stage 2 or REM, using criteria based on the average energies and standard deviations for , 2-4- and Hz frequency bins. Step 4-Waking, NREM and tonic REM epoch assignments are checked and adjusted. Numerous specific criteria are employed to fine-tune epoch classification. 1. Check tonic REM epochs for possible reassignment to stage 1 or 2. " 2. Check stage 2 epochs for possible reassignment to stage 1 or wake. 3. Check stage 1 for possible reassignment to wake. Step 5-Various components are adjusted and edited by a human operator. This final step consists of pot ential assistance by a human operator who can adjust the time base (start and stop times), align the record with other parameters, edit delta for movement artifact and manually adjust criterion cutoff levels for delta, beta, alpha and muscle (40-45 Hz) energy used by C STAGE in steps 2-4 above. For example, a subject with high all-night beta and/or muscle energy during sleep may require a higher Hz cutoff than an average subject. C STAGE rating. All subjects in this study were C STAGE rated by experienced raters (development sample, L.L.; validation sample, M.V.). Because step 5 involves human interactions that can affect C STAGE ratings, we examined the intra- and interrater reliabilities of the C STAGE, per se, by repeated ratings across all epochs for 10 sleep recordings (five men, five women). Each of the 10 records was scored twice by two raters (L.L. and M.V.), employing all five steps of CST AGE. Repeated ratings by a given rater (intrarater reliability) were made a minimum of 3 days apart. C STAGE comparisons with human ratings. The comparison of C STAGE with human rating was examined using group bivariate mean comparisons with Bonferroni corrections and Pearson regression correlations of human versus C STAGE ratings for the variables listed below. These were measured on study night 2 for each subject. The sleep/wake variables rated by both C STAGE and human raters include (but are not limited to): TWT WASO TST STG1+2 STG3+4 REM SLAT REML Total time spent awake during time in bed. Total time spent awake after sleep onset, defined as at least one (human) or two (C STAGE) epochs of stages 1 through 5. Total time scored as stages 1 through 5. Total time scored as stages 1 plus 2. Total time scored as stages 3 and 4. Total time scored as REM sleep. Time elapsed between lights out and onset of any sleep stage. Time elapsed between SLAT and the onset of REM sleep. Epoch-by-epoch analysis. Data from the development sample were not digitized on-line: Hence epochby-epoch analysis was not possible. However, the validation sample was digitized on-line, and alignment of the human- and C STAGE-rated records was relatively precise (±2 minutes). This relative precision in alignment permitted an epoch-by-epoch comparison of the two scoring methods for this sample. Sleep. Vol. 17. No

4 714 P. N. PRINZ ET AL. TABLE 1. Means and standard deviations of human- and C STAGE-scored sleep variables, and the correlation and regression equation for human versus computer variables for the group of70 older adults (SI or D.S.) Human C STAGE Regression Human vs. Mean SD Mean SD C STAGE" r p TWT ns WASO p < TST ns STG ns STG ns REM ns SLAT ns REML ns The epoch-by-epoch correspondence between the two methods for each subject was examined by cross-tallying epoch assignments to wake (including movement), stages 1 and 2 sleep, stages 3 and 4 sleep, and REM sleep. Proportions of agreement for each subject were then calculated. Because proportion of agreement does not take into account the fact that significant epoch-by-epoch correspondence can occur by chance, Kappa coefficients (15), which correct for chance agreements, we also calculated for each subject. Proportion of agreement and Kappa coefficients may be underestimates of the true correspondence between the methods because C STAGE scores sleep in 16- second epochs, whereas the human rater employs 30- second epochs. In addition, Rechtshaffen and Kales (11) employs numerous "smoothing rules", which tend to yield human ratings with many potential shifts in and out of sleep/wake stages glossed over. To examine this possibility C STAGE data were smoothed using a three-rule smoothing procedure designed empirically and post hoc to enhance congruence with human ratings. The proportions of agreement and Kappa coefficients were recalculated using these smoothed data. The smoothing procedure used three decision rules. a. Fragmentation of REM sleep was reduced in the following way. For an index epoch of REM, if another epoch of REM occurred within 5 minutes of the index, up to six epochs (1.5 minutes) adjacent to the index were rescored as REM. This would fill all gaps of 3 minutes or less between REM scores and would otherwise reduce the size of gaps between epochs more than 12 epochs (3 minutes) apart but less than 5 minutes apart. b. Stage 4 was not modified. c. Moving averages of epochs scored as stages 0 (wake) through 3 were calculated for the 3 minutes surrounding the index point. Epochs of stage 4 and REM were excluded from these moving averages. The average was rounded to the nearest integer, this becoming the "smoothed" value. Sleep. Vol. 17. No TABLE 2. Intra- and interrater reliabilities for human-assisted C STAGE scoring of wake and sleep stages. Ten records were rated twice by two raters (M. v., L.L.). Pearson correlation coefficients are shown for (A) repeat assessments for each rater andfor (B) correlation between ratersfor total wake time (0), stages 1-4 (1-4) and REM sleep (5) it (A) Intrarater L.L M.V (B) Interrater L.L. VS. M.V Development sample RESULTS Mean group data for human and C STAGE ratings of wake and sleep stages for the development subsampie are reported in Table 1. The two rating methods produced similar mean values for all variables except the wakefulness variable, W ASO, which was larger in C STAGE as compared with human ratings. Using all-night mean data from the total sample of 70 subjects, bivariate correlations between C STAGE and human ratings were performed. The correlation coefficientsexceededo.70(p < 0.001) for TWT, WASO, TST, STG3+4 and REM (Table 1). The impact of human assistance on C STAGE ratings (step 5) was examined in several ways. Mean intrarater reliability (Pearson correlation of all-night mean values) was high (r = ) for both raters (Table 2), with the stage 1 and stage REM correspondences being weakest (r = 0.70, 0.94 and 0.75,0.99). Interrater reliability was high for stages wake, 2, 3 and 4 (0.93 to 1.0). Interrater reliability for stages REM and 1 were lowest (0.04, 0.50). Validation sample Table 3 shows the validation group means for the sleep/wake variables of interest for both human and C STAGE scoring. Group means did not differ significantly for any ofthe sleep/wake variables. Table 3 also shows the results of correlational analyses examining the relationship between the two systems for each of the variables. Correlations for most sleep/wake variables were quite high (r = ; p < 0.001). Correlations for REM-based measures were lower (r = ). Figure 2 shows scatter plots comparing human and C STAGE ratings forwaso, TST, STG3+4 and REM. Epoch-by-epoch analysis revealed that the mean proportion of agreement between the two methods was 0.69 (see Table 4). Table 4 also reports the expected "

5 C STAGE: DEVELOPMENT AND VALIDATION 715 TABLE 3. Means and standard deviations of human- and CST AGE-scored sleep variables (in minutes), p values of two - tailed t tests for each variable (Bonferroni corrected) and correlation coefficients and significance levels for human- vs. C STAGE-scored variables for 45 healthy older adults Human CSTAGE Correlation t test Mean SD Mean SD (p) r p TWT ns WASO ns TST ns ns ns REM ns SLAT ns REML ns proportion of agreement, that is, the agreement expected based solely on chance. The mean Kappa coefficient between the two scoring methods was 0.49 (see Table 4). Table 4 also shows the mean proportion of agreement (0.74) and the mean Kappa coefficient (0.57) between the two methods using the smoothed C STAGE data, as well as the percent improvement TABLE 4. Means and standard deviations of proportions of agreement (observed and expected) and Kappa coefficient between raw and smoothed sleep stage scoring by human rater and C STAGE for 45 healthy older adults Percent change Raw score Smoothed score from raw Mean SD Mean SD score Proportion (observed) Proportion (expected) Q Kappa coefficient in correspondence over the raw data. Smoothing increased proportion of agreement by 6% and Kappa by 15%. DISCUSSION C STAGE was specifically designed to differ in some ways from human ratings. Our intent was to increase A 160 C!) 140 Z i= 120 LU 100 C!) 80 en u IWASol [ r=o.89 [ 20 [ p<.001 [ h 250 h ITST I 200 [ r=0.91 [ [ P <.001 [ C 120 C!) /' Z 100 i= LU 80 C!) en 60.. u '/ /' 40,...,/ '" - 20./,/T [ r-o.59 I FIG. 2. Scatter plots of human versus C STAGE sleep scoring for 45 healthy older subjects for (A) wake after sleep onset (W ASO) in minutes, (B) total sleep time (TST) in minutes, (C) stages 3 and 4 (STG#3+4) sleep in minutes, and (D) rapid eye movement (REM) sleep in minutes. The equivalency line (45 ) (- - -) and the regression line (-) are included for reference. Sleep, Vol. 17. No

6 716 P. N. PRINZ ET AL. detection of tonic REM delta sleep and brief arousals and stage transitions beyond that possible with human ratings. To achieve this, we elected to use a shorter epoch (16 seconds) and more rigorous distinctions regarding muscle, theta, beta and delta activities than is possible in human ratings. As a result, brief arousals or brief early epochs of sleep are more detectable with C STAGE than with human raters. Because our original aim in developing the C STAGE system was to assess tonic rather than phasic REM epochs for use as a diagnostic biomarker of Alzheimer's disease (16,17), C STAGE scores single epochs of tonic REM more frequently than human raters (Fig. 1). These considerations are reflected in the trends observed in the group means for the two rating methods: C STAGE tends to score somewhat more wake, less REM and shorter sleep and REM latencies compared to human rater. However, of the eight sleep/wake variables examined, only W ASO (and only in the development sample) differed significantly from human ratings, indicating that C STAGE yields group means comparable to conventional human ratings. Correlations between C STAGE and human ratings were significant for all of the variables examined and were particularly high (r > 0.7) for waking measures (TWT, WASO), TST and STG3+4 in both samples, for REM in the development sample, and for STG and SLAT in the validation sample. Weaker correlations were observed for REML and, in the development sample, SLAT, possibly reflecting C STAGE's scoring of single 16-second epochs. It is of interest to note that correlations between human and C STAGE ratings for the sleep/wake variables differ somewhat between the two samples. In the validation study, correlations for waking (TWT and WASO), TST, STG1+2 and STG3+4 tended to be higher than in the developmental sample (mean correlation of 0.84 vs. 0.75), whereas the correlation for REM was lower (0.59 vs. 0.74). These differences may derive from the fact that investigators performing C STAGE scoring focused primarily on detection of REM (development) or wake (validation study). Because C ST AGE allows for human interaction, issues of interrater reliability remain a concern, although intrarater reliability appears to be excellent (r = ) (Table 2). Because all-night mean data are far from ideal for testing comparability across scoring methods, we performed an epoch-by-epoch comparison on the validation sample. This analysis revealed excellent proportion of agreement between the two scoring methods. The use of the Kappa coefficient, a more conservative estimate of correspondence that corrects for chance agreement, also demonstrated that the two methods agree well on an epoch-by-epoch basis (11). When C Sleep. Vol. 17. No STAGE data were smoothed to correct for methodological differences between the human and C STAGE scoring systems, the correspondence between the two methods improved 15%. These analyses indicate C STAGE scoring is valid when compared on an epochby-epoch basis with human sleep scoring. It is noteworthy that (with minor exceptions) C STAGE yielded results comparable to human ratings both for the sample used to develop the scoring program and for a large independent sample (the validation sample). C STAGE appears to be a valid method of quantifying NREM sleep and wake patterns in healthy older adults, particularly for measures of wakefulness and stages 3 and 4 sleep. For applications requiring total REM detection, C STAGE is not the method of choice. C STAGE can be compared with other computer stage-scoring systems that employ multiple channels; in general, stage 1, REM and waking are most problematic for all computerized systems. Correlations coefficients reported for CNS's SL2000 (2) and SASSSY (3) were 0.60 and 0.59 for stage 1 and 0.51 and 0.68 for REM, respectively (all vs. human scorer). Similar weaknesses in rating stage 1, REM and/or waking have been observed for the Sleep lit (4), the Oxford Medilog 9000 system (5) and for Oxford Medi10g's SAC (D. Eder, personal communication). On the brighter side, most of these systems, like C STAGE, correspond well with human-rated STG3+4 and TST (r = ). C STAGE is particularly useful in large sample research applications. C STAGE has the methodological advantages of extensive data preconditioning [artifact removal (12,13)] and data reanalysis potential. C STAGE was developed as an algorithm to aid in locating raw EEG epochs of interest for further analysis. Once data segments of interest have been defined by C STAGE, they can be recovered precisely from the original data and processed quite differently from the C STAGE analysis. This processing can include special spectral analysis during tonic REM sleep to detect early Alzheimer's disease (16,17) and/or delta-wave quantitation as affected by age (18), hormonal (19) or other factors of interest. In summary, C STAGE compares well with human ratings of sleep and wakefulness in healthy older adults. For this population, C STAGE is comparable to other computer sleep-scoring systems; however, the validity of C STAGE in younger and/or clinical populations remains to be demonstrated. Acknowledgements: This study was supported by Public Health Service grants MH45l86, MH33688, RR37 and the Department of Veterans Affairs Medical Research Service. The authors thank David Akers, Maria Bjelke, Michael Davis, Valerie Larson, Kristen Lawrence, Alan Marks, Sharon Roloff, Tameria Veith and Robert Ward for their expert assistance.

7 C STAGE: DEVELOPMENT AND VALIDATION 717 REFERENCES 1. RoffWarg HP. Automatic scoring, ASDA position statement. Sleep 1990; 13: Becker PM, Forester M, Jamieson AO, et al. Comparisons of the CNS sleep IIT (edited), CNS CASS and human scoring of sleep. Sleep Res 1993;22: Doghramji K, Breuninger W, Gaddy JR, Beck J, Grant R, Mac Donald J. Comparison of Semi-Automated Sleep Staging System (SASSSy) with manual scoring. Sleep Res 1992;21 : Scharf MB, McDannold M. An evaluation of consistency and accuracy of data obtained from an 8-channel physiologic recorder (sleep IIT) versus conventional polysomnography. Sleep Res 1993;22: Ferri R, Ferri P, Colognola RM, Petrella MA Musumeci SA Bergonzi P. Comparison between the results f an automati and a visual scoring of sleep EEG recordings. Sleep : ' 6. Penzel T, Stephan K, Kubicki S, Herrmann WM. Integrated sleep analysis, with emphasis on automated methods. In: Degen R, Rodm EA, eds. Epilepsy, sleep and sleep deprivation, 2nd ed. New York: Elsevier, 1991: Williams DE, Vitiello MV, Ries RK, Bokan J, Prinz PN. Successful recruitment of elderly, community dwelling subjects for Alzhimer.'s disease research: cognitively impaired, major depressive disorder, and normal control groups. J Gerontol (3):M ' 8. Vitiello MV, Prinz PN, Williams DE, Frommlet MS, Ries RK. Sleep disturbances in patients with mild-stage Alzheimer's disease. J GerontoI1990;45:MI Yitiello MV, Prinz PN, Avery DH, et al. Sleep is undisturbed m elderly, depressed individuals who have not sought health care. Bioi Psychiatry 1990;27: Vitiello MV, Prinz PN, Schwartz RS. Slow wave sleep but not overail sleep quality of healthy older men and women is improved by increased aerobic fitness. Sleep Res, 1994 (in press). II. Rechtschaffen A, Kales A, eds. A manual 0/ standardized terminology, techniques, and scoring system/or sleep stages o/human subjects. USPHS Publication No Washington DC: U.S. General Publishing Office, Larsen LH, Prinz PN. EKG artifacts suppression from the EEG. Electroencephalogr Clin Neurophysiol1991 ;79: Larsen LH, Prinz PN, Moe KE. Quantitative analysis of the EEG during tonic REM sleep-methodology. Electroencephalogr Clin NeurophysiolI992;83: Priestley MB. Spectral analysis and time series, Vol. 1: univariate series. Vol. 2: multivariate series, prediction and control. New York: Academic Press, Fleiss JL. Statistical methods for rates and proportions, 2nd ed. New York: Wiley & Sons, Prinz PN, Larsen LH, Moe KE, Vitiello MV. EEG markers of early Alzheimer's disease in computer selected tonic REM sleep. Electroencephalogr Clin NeurophysioI1992;83: Moe KE, Larsen LH, Prinz PN, Vitiello MV. Major unipolar depression and mild Alzheimer's disease: differentiation by quantitative tonic REM EEG. Electroencephalogr Clin Neurophysiol 1993;86: Larsen LH, Moe KE, Prinz PN, Dulberg EM, Vitiello MV. The use of normalization techniques to reveal age effects in the sleep EEG. J Sleep Res (in press). 19. Prinz PN, Moe KE, Dulberg EM, Larsen LH, Vitiello MV, Toivola B, Merriam GR. Higher plasma IGF-I levels are associated with increased delta sleep in healthy older men. J Gerontol Med Sci, 1994 (in press). Sleep, Vol. 17, No.8, 1994