5.1. Macro-architecture

We've previously reviewed the HYPNO command in the context of QC. Here we'll output a range of macro-architecture metrics, testing for sex differences.

Unlike the previous steps where we typically used a single temporary output database (out.db) here we'll archive outputs more systematically, as we'll compile some of these outputs for association analyses in subsequent steps. We'll make two folders: out/ for all Luna .db files generated, and res/ for the text files extracted from those output databases:

mkdir out res

Running HYPNO

All metrics generated in this section are from the HYPNO command. We'll run with the epoch option to request epoch-level outputs:

luna c.lst -o out/hypno.db -s HYPNO epoch

For the subsequent association analyses we'll extract three result files from out/hypno.db: first, the base-line per-individual metrics:

destrat  out/hypno.db +HYPNO > res/hypno.base

Second, metrics stratified by sleep stage:

destrat  out/hypno.db +HYPNO -r SS/N1,N2,N3,R,W > res/hypno.stage

Third, metrics stratified by NREM cycle, considering only cycles 1 through 4 (as we'll see below, not many individuals in this sample had more than 4 cycles):

destrat  out/hypno.db +HYPNO -r C/1,2,3,4  > res/hypno.cycle

Baseline metrics

In R, we'll review the outputs, loading the full database (although we could equivalently have loaded the text file res/hypno.base here instead):

library(luna)
k <- ldb("out/hypno.db")

We extract the baseline (BL) strata of results into a data-frame d:

d <- k$HYPNO$BL

and then merge this table with the demographic data:

p <- read.table("work/data/aux/master.txt", header=T, stringsAsFactors=F)
d <- merge( d , p , by="ID" ) 

The base-line variables are represented in the columns of d:

names(d)

[1] "ID"                  "CONF"                "E0_START"           
 [4] "E1_LIGHTS_OFF"       "E2_SLEEP_ONSET"      "E3_SLEEP_MIDPOINT"  
 [7] "E4_FINAL_WAKE"       "E5_LIGHTS_ON"        "E6_STOP"            
[10] "EINS"                "FIXED_LIGHTS"        "FIXED_SLEEP"        
[13] "FIXED_WAKE"          "FWT"                 "HMS0_START"         
[16] "HMS1_LIGHTS_OFF"     "HMS2_SLEEP_ONSET"    "HMS3_SLEEP_MIDPOINT"
[19] "HMS4_FINAL_WAKE"     "HMS5_LIGHTS_ON"      "HMS6_STOP"          
[22] "LOST"                "LOT"                 "LZW"                
[25] "LZW3"                "NREMC"               "NREMC_MINS"         
[28] "OTHR"                "POST"                "PRE"                
[31] "REM_LAT"             "REM_LAT2"            "SE"                 
[34] "SFI"                 "SINS"                "SME"                
[37] "SOL"                 "SOL_PER"             "SPT"                
[40] "SPT_PER"             "T0_START"            "T1_LIGHTS_OFF"      
[43] "T2_SLEEP_ONSET"      "T3_SLEEP_MIDPOINT"   "T4_FINAL_WAKE"      
[46] "T5_LIGHTS_ON"        "T6_STOP"             "TGT"                
[49] "TIB"                 "TI_RNR"              "TI_S"               
[52] "TI_S3"               "TRT"                 "TST"                
[55] "TST_PER"             "TWT"                 "WASO"               
[58] "male"                "age"                

Plotting an initial representative metric, the sleep period time (SPT), the time between sleep onset and final wake), by males and females:

boxplot( d$SPT ~ rep(c("Female","Male"),each=10) , col = c("red","blue" ), 
         xlab = "Sex" , ylab = "Sleep Period Time (SPT)" ) 

There appears to be evidence for a sex difference, which we can formally test, for example via a simple linear regression also controlling for age (and removing any potential outliers (default +/- 3 SD units) from the dependent variable):

summary(lm( outliers( SPT )  ~ age + male , data = d ) )

Coefficients:
            Estimate Std. Error t value Pr(>|t|)    
(Intercept) 479.9224    32.0737  14.963 3.22e-11 ***
age          -1.9894     0.8839  -2.251 0.037936 *  
male         38.1819     9.5082   4.016 0.000896 ***
---
Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 20.99 on 17 degrees of freedom
Multiple R-squared:  0.5252,    Adjusted R-squared:  0.4693 
F-statistic: 9.402 on 2 and 17 DF,  p-value: 0.00178

We see that males have significantly longer SPTs in this small sample: 7.4 hrs versus 6.8 hrs (with 0 and 1 indicating female and male below):

tapply( d$SPT , d$male , mean )  / 60 

0        1 
6.821667 7.401667 

We can continue to test a range of key sleep metrics, applying the same linear model with the results tabulated here:

Variable	Metric	Beta	P-value
TRT	Total recording time (mins)	35.26	0.004
TIB	Time-in-bed, here = TRT (mins)	35.26	0.004
SPT	Sleep period time (mins)	38.18	0.001
TST	Total sleep time (mins)	23.61	0.19
TST_PER	Persistent sleep time (mins)	-4.98	0.84
TWT	Total wake time (mins)	11.66	0.62
WASO	Wake after sleep onset (mins)	14.57	0.48
SE	Sleep efficiency (%)	-1.35	0.78
SME	Sleep maintenance efficiency (%)	-2.48	0.58
SOL	Sleep onset latency (mins)	-1.23	0.88
REM_LAT	REM latency, includes WASO (mins)	-16.32	0.38
REM_LAT2	REM latency, excludes WASO (mins)	-13.18	0.38

That is, although males have significantly longer SPT and also longer time in bed (TIB) and total recording time (TRT), there are actually no (significant) group differences in total sleep time (TST) or wake times (TWT or WASO), nor on sleep efficiency or onset latencies.

As we don't have explicit lights-off and lights-on annotations present, the TRT and TIB metrics are identical here (i.e. lights off is assumed to be the start of the recording, etc).

Stage-specific metrics

Next, we'll look at stage-specific metrics, including stage duration. Above, these metrics were saved in the file res/hypno.stage, but we can also extract them from the attached database (k), and merge with the demographic data as before:

d <- k$HYPNO$SS
d <- merge( d , p , by="ID" )

The absolute stage duration is encoded by the MINS variable, which is defined for each stage (N1, N2, N3, R as well as W and WASO) but also for combined stages: (NR: all NREM, S: all sleep). The left/right columns give the means (in minutes) for females and males respectively:

tapply( d$MINS , list( d$SS, d$male  ) , mean ) 

          0      1
?      0.00   0.00
L      0.00   0.00
N1    30.00  41.15
N2   195.75 233.55
N3    60.10  51.40
N4     0.00   0.00
NR   285.85 326.10
R     68.50  49.20
S    354.35 375.30
W     78.15  91.90
WASO  54.95  68.80

These appear broadly as expected, e.g. the majority of time asleep is spent in N2 (over 3 hours), with a average total sleep time around 6 hours (note, these were recorded in a hospital sleep lab, and so for a variety of reasons, sleep durations will likely be shorter then natural sleep time).

Fitting the same linear models as above, we see there appears to be some suggestion of sex differences in a number of stage-specific metrics:

Variable	Metric	Beta	P-value
MINS N1	N1 duration (mins)	11.377	0.145
MINS N2	N2 duration (mins)	39.591	0.038
MINS N3	N3 duration (mins)	-8.832	0.514
MINS NR	NREM duration (mins)	42.136	0.015
MINS R	REM duration (mins)	-18.529	0.024
PCT N1	Relative N1 duration (%)	2.384	0.254
PCT N2	Relative N2 duration (%)	7.412	0.1
PCT N3	Relative N3 duration (%)	-3.738	0.294
PCT NR	Relative NREM duration (%)	6.059	0.003
PCT R	Relative REM duration (%)	-6.059	0.003

That is, males in this sample appear to have more NREM but less REM sleep, in both absolute and relative senses. The effect is primarily due to more N2 (almost 40 minutes more) and about 20 minutes less REM. Of course, whether or not these differences generalize to the larger GRINS sample, let alone to other samples from other comparable, independent populations, is a separate question.

Bout/transition metrics

As well as stage durations, the stage-stratified (SS) outputs also contain some simple summaries of the number and length of contiguous bouts of each stage, as well as metrics that attempt to quantify the rate and nature of inter-stage transitions.

Some of these are listed below, with the same beta/p-value estimates from the linear model testing for sex differences (beta: effect in males versus females):

Variable	Metric	Beta	P-value
SFI	Sleep Fragmentation Index (sleep to W count / TST)	0.022	0.068
TI_S	Transition Index (excludes W) N1-N2-N3-R (count / TST)	-0.0058	0.82
TI_S3	3-class Transition Index (NR,R,W (count / SPT)	0.023	0.13
TI_RNR	REM-NREM transitions (count / TST)	-0.013	0.048
BOUT_N N1	Number of N1 bouts	12.087	0.13
BOUT_N N2	Number of N2 bouts	5.668	0.344
BOUT_N N3	Number of N3 bouts	-3.372	0.323
BOUT_N NR	Number of NR bouts	6.858	0.062
BOUT_N R	Number of REM bouts	-0.592	0.836
BOUT_N S	Number of sleep bouts	10.609	0.006
BOUT_N W	Number of wake bouts	10.405	0.007
BOUT_N WASO	Number of WASO bouts	10.609	0.006

boxplot( d$BOUT_N[ d$SS == "S" ]  ~ rep(c("Female","Male"),each=10) ,
         col = c("red","blue" ),
         xlab = "Sex" , ylab = "# of sleep bouts" )

That is, males have a greater number of sleep bouts (i.e. periods of contiguous sleep), or, correspondingly, more wake/WASO) bouts. Given that males and females have approximately equal TSTs, this implies a more fragmented (more sleep-wake changes) in males. This is reflected in the slightly higher sleep fragmentation index (SFI), although that is not significant in this (small) sample. In general, this speaks to the utility in querying multiple related metrics (when coupled with appropriate correction for multiple testing) in the context of biomarker discovery, when there is no strong a priori and specific reason for any one particular metric.

NREM cycle metrics

Next, we'll pull out the cycle-specific metrics: primarily just (REM/NREM) duration per cycle.

d <- k$HYPNO$C
d <- merge( d , p , by="ID" )

After merging the demographic data, we can compute a table of means (in minutes)

tapply( d$NREMC_MINS , list( d$C, d$male  ) , mean )

0         1
1 125.15000 109.35000
2  93.60000 114.00000
3  87.45000  97.72222
4  67.42857  65.00000
5  44.33333  52.12500
6        NA  56.00000

It is important to remember that not all individuals will have all six cycles:

tapply( d$NREMC_MINS , list( d$C, d$male  ) , length )

   0  1
1 10 10
2 10 10
3 10  9
4  7  9
5  3  4
6 NA  1

i.e. we see that N = 19 people have at least 3 NREM cycles defined, and everybody has at least two. Analyses based on the 5th cycle would not be representative of the whole sample, however.

From the baseline metrics (as extracted above), the variables NREMC and NREMC_MINS give the count and average duration of NREM cycles per individual; applying the same linear model as above, we can confirm there are no significant differences between males and females for either metric:

Variable	Metric	Beta	P-value
NREMC	Number of NREM cycles	0.3	0.50
NREMC_MINS	Average NREM cycle duration (mins)	4.32	0.72

Hypnogram visualization

Finally, as we extracted epoch-level data with the epoch option, we can view epoch-level metrics and plot hypnograms, etc. For example:

d <- k$HYPNO$E
lhypno( d$STAGE[ d$ID == "F01" ] )

As we've previously reviewed visualizing hypnograms earlier in this walkthrough, we'll not repeat those steps here.

Summary

Based on extracting a set of common hypnogram-derived metrics describing sleep macro-architecture in males and females, we can conclude that:

• males spent more time in bed, despite not showing significant differences in TST, sleep onset latency, sleep efficiency or WASO

• males had relatively more NREM sleep, females had relatively more REM

• there were no obvious differences in NREM cycle structure

Even here, given the small sample it would be incautious to make strong conclusions, due to the threat of both false positive (multiple testing) and false negative (low power) errors. We'll address both issues later, by a) using methods to correct for multiple testing, and b) querying the larger remaining GRINS control sample for the same tests.

---

In the next section we'll consider time/frequency (spectral) analysis.