Skip to content

5.7. Age-prediction from the NREM EEG

In this section, we apply a previously developed model to predict biological/brain age based on the sleep EEG, using Luna's PREDICT command to apply a model developed by Sun et al (2019) which is based on 13 NREM sleep EEG. Technically, the model was trained on contra-lateral mastoid referenced data, but for the purposes of this walkthrough, using the linked mastoid data will be sufficient.

Applying the basic model

The model files are bundled with the walkthrough data, in the work/data/aux/models/ folder:

  • a Luna script that derives all 13 features and applies the model

  • a file that specifies the features weights for the model and some other parameters

  • a file that contains a normalized training feature matrix, used in kNN missing data imputation

You can review the steps in m1-adult-age-luna.txt:

  • it uses Luna's cache mechanism to store and pass outputs from Luna commands to the PREDICT command at the end of the file

  • it also uses Luna's freeze/thaw mechanism to derive statistics from N1, N2 and N3 stages

  • most features are derived from applications of commands such as MTM, STATS and SPINDLES that match the settings used to build the model

The script expects the following variables to be defined:

  • ${age} which gives the observed age (in years) of each individual; these are defined in work/data/aux/master.txt and so we pass that file to Luna via the vars special command option, so Luna is able to define individual-specific variables (i.e. that are substituted in the script each time it is run for a given person)

  • ${cen} that lists one or more central electrodes

  • ${th} that specifies the SD threshold for missing data

  • ${mpath} that specifies the model folder path (so Luna knows where to look for the model specification/data files)

We can apply this model to all 20 individuals by executing the following:

luna c.lst vars=work/data/aux/master.txt cen=C3,C4 th=3 \
 mpath=work/data/aux/models/ \
 -o out/pad.db < work/data/aux/models/m1-adult-age-luna.txt

We'll save the key predictions into res/pad.base:

destrat out/pad.db +PREDICT > res/pad.base

and the feature-level outputs in res/pad.ftr: 

destrat out/pad.db +PREDICT -r FTR > res/pad.ftr

Looking at the key predictions in R: 

d <- read.table("res/pad.base",header=T,stringsAsFactors=F)
head(d)

   ID        DIFF NF NF_OBS OKAY        Y       Y1 YOBS
1 F01  -9.3203385 13     13    1 35.93861 31.67966   41
2 F02 -13.2381840 13     13    1 27.83963 18.76182   32
3 F03   0.5032096 13     13    1 43.90388 37.50321   37
4 F04 -10.8985415 13     13    1 30.17927 21.10146   32
5 F05 -14.3534051 13     13    1 25.79527 15.64659   30
6 F06  -1.6613576 13     13    1 39.41646 30.33864   32
...

The primary metric is DIFF, which is the predicted age (Y1) minus the observed age (YOBS).

In the context of brain age prediction, this is often called PAD (predicted age difference). As the PREDICT command is generic, and not solely developed for the purpose of age prediction, it is labeled DIFF here instead, but we'll use these two terms interchangeably below.

The average PAD is not significantly different from zero in this sample, suggesting that this is a relatively unbiased estimate when applied to the general population.

 t.test( d$DIFF )

t = -1.1406, df = 19, p-value = 0.2682

In contrast, predicted age (Y1) is significantly correlated with observed age: 

cor.test( d$Y1 , d$YOBS ) 

t = 2.844, df = 18, p-value = 0.01077
      cor 
0.5568107

Sex differences in PAD

We can also look at possible sex differences in the PAD. Rather than merge the phenotype data, we can create a quick indicator of sex, here based on the ID labels (i.e. if starting with M):

d$MALE <- as.integer( grepl( "^M" , d$ID ) )

On average, males and females are well-matched for chronological age, i.e. there is no significant difference in YOBS: 

t.test( d$YOBS ~ d$MALE ) 

  Welch Two Sample t-test

  t = -0.67903, df = 16.611, p-value = 0.5065

  mean in group 0 mean in group 1 
             35.5            37.2

However, we do see a difference in PAD, with males having profiles of NREM sleep that are typically more consistent with older individuals (a PAD of +2.1 years on average), whereas females have profiles that are typically more consistent with younger individuals (a PAD of -6.1 years on average):

t.test( d$DIFF ~ d$MALE ) 

  Welch Two Sample t-test
  t = -2.6077, df = 16.895, p-value = 0.01845

   mean in group 0 mean in group 1 
         -6.211031        2.071965

Plotting all the data colored by sex:

lim = range( d$YOBS , d$Y1 ) 
plot( d$YOBS , d$Y1 , pch=20, col = ifelse( d$MALE , "blue" , "red" ) ,
      xlim = lim , ylim = lim , xlab="Observed age" , ylab="Predicted age" ) 
lines( c(0,100),c(0,100))

img

Feature-level analysis

Looking at the feature level outputs:

d <- read.table("res/pad.ftr",header=T,stringsAsFactors=F)
head(d)

We can tabulate the 13 features: 

table( d$FTR ) 

alpha_bandpower_kurtosis_C_N2     alpha_bandpower_mean_C_N1 
                           20                            20 
              COUPL_OVERLAP_C         delta_alpha_mean_C_N3 
                           20                            20 
delta_bandpower_kurtosis_C_N2     delta_bandpower_mean_C_N3 
                           20                            20 
        delta_theta_mean_C_N3                        DENS_C 
                           20                            20 
                kurtosis_N2_C                 kurtosis_N3_C 
                           20                            20 
sigma_bandpower_kurtosis_C_N2 theta_bandpower_kurtosis_C_N2 
                           20                            20 
theta_bandpower_kurtosis_C_N3 
                           20

Per individual, the primary metrics are X (the raw value for that feature) and Z, the value derived after normalization (and potentially missing data imputation). The B metric is the model weight (and so the same for every individual), which indicates whether the feature is expected to increase or decrease with age. Also, we won't dwell on this here, but only two feature values were re-imputed (REIMP) based on the 3 SD unit threshold, so overall these metrics appear consistent with expectations.

d$MALE <- as.integer( grepl( "^M" , d$ID ) )


for (ftr in unique( d$FTR ) )
{
 tt <- t.test( Z ~ MALE , data = d , subset = FTR == ftr ) 
 cat ( ftr , round( tt$statistic,2) , round(tt$p.value,3) , "\n" , sep="\t" ) 
}

COUPL_OVERLAP_C                 -1.37    0.189      
DENS_C                          -0.54    0.596      
alpha_bandpower_kurtosis_C_N2    0.58    0.569      
alpha_bandpower_mean_C_N1        0.14    0.893      
delta_alpha_mean_C_N3             2.1    0.056 .     
delta_bandpower_kurtosis_C_N2    2.76    0.013 *     
delta_bandpower_mean_C_N3        2.17    0.049 *     
delta_theta_mean_C_N3            2.04    0.062 .     
kurtosis_N2_C                    3.98    0.001 *     
kurtosis_N3_C                     2.3    0.036 *     
sigma_bandpower_kurtosis_C_N2   -1.67    0.114      
theta_bandpower_kurtosis_C_N2    1.79    0.095 .    
theta_bandpower_kurtosis_C_N3    2.33    0.033 *

We see the apparent sex difference in PAD was not driven by a single, extreme feature: 5 of the 13 features show nominal (p<0.05) male/female differences, and 3 more show p<0.1 trends in this N=20 sample.

Summary

In this section we've seen how a NREM sleep based prediction of biological or brain age can be computed from EEG data, and we found some suggestions in this small sample that males and females show different rates of brain aging.

We've also had a glimpse of Luna's generic PREDICT command, which is designed to allow users to build and apply their own linear models for prediction, although this advanced usage is beyond the scope of this walkthrough.

In the next section we'll consider linear models and multiple test correction.