[PPT] - Sensitivity to Sleep Deprivation. Analysis from Short to Long Time PowerPoint Presentation

SLIDE 1

New PVT Metrics with an Improved Sensitivity to Sleep Deprivation.

Analysis from Short to Long Time Intervals.

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 1

LATOUR Philippe and VAN DROOGENBROECK Marc University of Liège – Belgium

SLIDE 2

Interested in analyzing, studying, designing and/or assessing performances

… of automatic « instantaneous » alertness monitoring and drowsiness detection problems/systems … by using (especially) PVTs (Psychomotor Vigilance Test)

General context

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 2

SLIDE 3

22 Subjects (11 males, 11 females, mean 22.2y., range 19-34 years):

Arrival at the laboratory at 8h30, day 1. Non-SDP PVT 1 : at 9h30, day 1; this is the reference “Non Sleep Deprived” PVT. Go home or at work, with an actigraph and back to the laboratory at 20h30, day1. SDP PVT 2 : at 02h30, during the night; this is the first Sleep Deprived PVT. SDP PVT 3 : at 10h30, day 2; this is the second Sleep Deprived PVT.

Our PVT Protocol (22 Subjects & 3 PVTs)

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 3 PVT Tests 1 2 3 Night 1 Day 1 Night 2 Day 2 At home In lab At home + actigraphy In lab Sleep No sleep No stimulant 23h 10h 7h 4h 2h 23h 20h30 11h 9h 7h 12h

Figure adapted, with permission, from C. François & al., “Tests of a new drowsiness characterization and monitoring system based on ocular parameters”, in Int. J. Environ. Res. Public Health, Vol. 13, n°2, 2016, pp. 174-183

SLIDE 4

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 4

Temporal sub-interval … …

PVT

' k

RT

1 k

RT 

k

RT

1 k n

RT  

Physiological Signals

1 k

RT 

k n

RT 

SLIDE 5

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 5

Temporal sub-interval … …

PVT

' k

RT

1 k

RT 

k

RT

1 k n

RT  

Physiological Signals Alertness Monitoring & Drowsiness Detection Metrics computation

1 k

RT 

k n

RT 

SLIDE 6

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 6

Temporal sub-interval … …

PVT

1 ' '

1 k n

k k k

RT n RT

  





1 ' '

1 1

k n k k k

n R RS T

  





 

'

# ' 500 ' 5 ,

k

k RT ms k k LN k n     

' k

RT

1 k

RT 

k

RT

1 k n

RT  

Physiological Signals Alertness Monitoring & Drowsiness Detection Metrics computation Alertness level

1 k

RT 

k n

RT 

SLIDE 7

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 7

Temporal sub-interval … …

PVT

1 ' '

1 k n

k k k

RT n RT

  





1 ' '

1 1

k n k k k

n R RS T

  





 

'

# ' 500 ' 5 ,

k

k RT ms k k LN k n     

' k

RT

1 k

RT 

k

RT

1 k n

RT  

Physiological Signals Alertness Monitoring & Drowsiness Detection Metrics computation Alertness level Compare Correlate

1 k

RT 

k n

RT 

SLIDE 8

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 8

Difficulties … PVT metrics sensitivity to sleep deprivation has been demonstrated mainly when computed on the full length PVT, not on (much) shorter temporal sub-interval. RT distribution (and then also any metrics distribution) are strongly dependent on the subject. How is the alertness level retated to sleep deprivation?

SLIDE 9

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 9

Difficulties … PVT metrics sensitivity to sleep deprivation has been demonstrated mainly when computed on the full length PVT, not on (much) shorter temporal sub-interval.

Will be discussed later … stay tuned!

RT distribution (and then also any metrics distribution) are strongly dependent on the subject.

It’s Now

How is the alertness level retated to sleep deprivation?

Not Today!

SLIDE 10

Lapse Number

Use a threshold adapted to the subject; instead of 500ms for everyone.

meanRT / meanRS

Not obvious at first sight Could consider to apply a kind of normalizing function to RT/RS before summing or averaging

How to normalize metrics?

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 10

SLIDE 11

Lapse Number

Use a threshold adapted to the subject; instead of 500ms for everyone.

meanRT / meanRS

Not obvious at first sight Could consider to apply a kind of normalizing function to RT/RS before summing or averaging

RT

… …

25( 3) Q Sbj

SLIDE 17

LNQ25 : Normalized Lapse Number

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 17 The entire Non-SDP PVT1 (10min) of Subject 3

… …

k

RT

2

RT

1

RT

N

RT

Metrics computation

… …

Sub-interval in a PVT of Subject 3

… …

' k

RT

k

RT

1 k n

RT  

… …

1 k

RT 

k n

RT 

SLIDE 18

LNQ25 : Normalized Lapse Number

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 18 The entire Non-SDP PVT1 (10min) of Subject 3

… …

'

1

# ' , 25 25 ) ' ( 3

k

RT

LNQ Q Sbj k k k k n

         

    

k

RT

2

RT

1

RT

N

RT

Metrics computation

… …

Sub-interval in a PVT of Subject 3

… …

' k

RT

k

RT

1 k n

RT  

… …

1 k

RT 

k n

RT 

25( 3) Q Sbj

SLIDE 19

Lapse Number

Use the 25% quantile of the RS distribution in the Non-SDP PVT for each subject as a lapse threshold.

meanRT / meanRS

Could consider to apply a kind of normalizing function to RT/RS before summing or averaging

How to normalize metrics?

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 19

SLIDE 20

Lapse Number

Use the 25% quantile of the RS distribution in the Non-SDP PVT for each subject as a lapse threshold.

meanRT / meanRS

Could consider to apply a kind of normalizing function to RT/RS before summing or averaging

RT

… …

 

LpPr 1

RS RS 



      

 

SLIDE 24

ELN: Expected Lapse Number

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 24 The entire Non-SDP PVT1 (10min) of Subject 3

… …

k

RT

2

RT

1

RT

N

RT

Metrics computation

… …

Sub-interval in a PVT of Subject 3

… …

' k

RT

k

RT

1 k n

RT  

… …

1 k

RT 

k n

RT 

SLIDE 25

ELN: Expected Lapse Number

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 25 The entire Non-SDP PVT1 (10min) of Subject 3

… …

1 ' '

L 1 pPr

k n k k k

RT

ELN

  

     

 

k

RT

2

RT

1

RT

N

RT

Metrics computation

… …

Sub-interval in a PVT of Subject 3

… …

' k

RT

k

RT

1 k n

RT  

… …

1 k

RT 

k n

RT   

LpPr 1

RS RS 



      

 

SLIDE 26

Lapse Number

Use the 25% quantile of the RS distribution in the Non-SDP PVT for each subject as a lapse threshold.

meanRT / meanRS

Use an estimation of the Lapse Probability as a normalizing function for RS.

How to normalize metrics?

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 26

SLIDE 27

Alertness Monitoring & Drowsiness Detection

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 27

Temporal sub-interval

PVT

RT RS 500 LN

Metrics computation

Measure the Sensitivity to Sleep Deprivation ?

25 LNQ ELN

… …

' k

RT

1 k

RT 

k

RT

1 k n

RT  

1 k

RT 

k n

RT 

SLIDE 28

A metric mi

p,s,l is computed

n the temporal Interval ‘i’
f Length ‘l’

from the PVT ‘p’

f the Subject ‘s’

How to measure sensitivy to Sleep Deprivation

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 28

Metrics computation

Interval ‘i’

f Length ‘l’

… …

' k

RT

k

RT

1 k n

RT  

… …

1 k

RT 

k n

RT 

PVT ‘p’ of Subject ‘s’

, ,l i p s

m

l i

Itv

SLIDE 29

We would like, for all Subject ‘s’, Length ‘l’ and interval ‘i’ & ‘j’, mi

p=(2,3),s,l for the SDP PVT ‘2’ or ‘3’

should be significantly larger than mj

p=1,s,l for the Non-SDP PVT ‘1’

How to measure sensitivy to Sleep Deprivation

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 29

 

2 , , 1 , ,3 ,

?

l p s p s j l i

m m

 



SLIDE 30

We would like, for all Subject ‘s’, Length ‘l’ and interval ‘i’ & ‘j’, Dmij

p=(2,3),s,l = ( mi p=(2,3),s,l - mj p=1,s,l )

should be significantly positive.

How to measure sensitivy to Sleep Deprivation

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 30

   

 

2,3 2, , , , , 3 , 1 ,

?

ij i j p s p s p l l l s

m m m

  

D   

SLIDE 31

We would like, for a given interval Length ‘l’, The mean value Dm

l of the difference of metric Dmij p,s,l

for all SDP-PVT, all Subjects and all Intervals of the given Length l

Should be significantly positive

How to measure sensitivy to Sleep Deprivation

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 31

 

, , , , , , , , , ,

# # # 1 #

Mean

?

ij ij i j i j i j l l l p s p s p s p s p m

s

m m D  D  D 



SLIDE 32

We would like, for a given interval Length ‘l’, the Effect Size ESDm

l ,

which is the ratio of the mean value Dm

l by the standard

deviation Dm

l of the differences of metric Dmij p,s,l

for all SDP-PVT, all Subjects and all Intervals of the given Length l

Should be as large as possible

How to measure sensitivy to Sleep Deprivation

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 32

   

, , , , , , , , , ,

Mean as large as possible StdDev

i l l l l l j i j ij i p s p s p s p s m j m m

m ES m  

D D D

D   D

SLIDE 33

Sensitivity to Sleep Deprivation of Metrics

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 33

SLIDE 34

We observe absolute ES values that are lower than those

btained by Basner & Dinges. The possible reasons are:

Our protocol slightly differs and might lower the average sleepiness in

ur population.

We use less PVTs; 3 instead of 17 (1 instead of 7 in non-SDP, and 2 instead of 10 in SDP).

A drawback of LNQ25, ELN compared to the meanRS is that a reference distribution of the RS is necessary to compute them.

Discussion

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 34

SLIDE 35

On time interval greater than 3 min, LNQ25, ELN clearly

utperform the now standard meanRS

On time interval greater than 2 min, the ES 95% confidence interval of LNQ25 is greater than 1.0 and LNQ25 is (just) greater than 1.0 after 1 minute. On very short interval duration (1 or 2 minutes), the covering of the confidence intervals asks us to remain cautious before drawing definitive conclusions.

Discussion

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 35

SLIDE 36

LNQ25, ELN and meanRS are the most sensitive metrics to sleep deprivation. PVT metrics should not be computed on time interval smaller than 2 or 3 min, for keeping the Sleep Deprivation sensitivity as large as possible. On our data, LNQ25 (& ELN) outperform meanRS and should certainly be preferred, if we can accept their additional complexity.

Conclusions

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 36

SLIDE 37

Thank you for your attention

20 March 2017 10th Int. Conf. on Managing Fatigue (San Diego, USA) 37