Predicting Performance Under Stressful Conditions Using Galvanic Skin Response

1 Pred icting Per for mance Und er Str essfu l Cond itions Using Galv anic Skin R espons e Carter M undel l, J uan Pab lo V ielma, and Tauh id Zaman * Sl oan School of Man agem ent , Massac hus etts I nstit ute of Technol ogy, C ambri dge, MA 0 2139 USA. *To whom co rresponde nce s hould be address ed. E - mail : zlisto @mit.ed u The rapid grow th of the avail abilit y of w earable biosensors has created the oppo rtunity for using biologica l signals to measu re worker p erforma nce. An important qu estio n is how to use such signal s to not just measu re, but ac tually pr ed ict w orker p erf orm ance on a task under stressful and potentiall y high risk c onditio ns. Here w e sho w that the biological sig nal known as gal vanic skin respo nse (GS R) allo ws such a predic tion. We co nduct an experiment w here subj ects answer arithmet ic questio ns under low and hi gh stress co nditions w hile having their GSR mo nitored using a w earable bio sensor . Using onl y the GSR measured unde r low str ess co ndition s , w e are able to predic t whic h subjects w i ll perform w ell under high stress co ndition s , ac hieving an area under the curve (A UC) of 0.76. If w e try to make si milar predic tions with out using any b iometric sig nals, the AUC barely excee ds 0.50 . Our result sugg ests that p erformance in hig h stress c onditions c an be predicted usi ng signals ob tained from w earable bio sensors in low stress conditio ns . For ma ny j obs , it is difficult to te st a wo rke r’s per for man ce u nd er real - life conditions eith er becau se it is prohibitively e xpens ive or bec ause th e stak es are to o high to ris k an e rror. For instance , it ca n be diff icult to simu late the a ctual pr essur e and risk faced by a financ ial trader on a trading floor. The true ris k of an a ccident is har d to recr eate fo r air line pil ots or truck drivers . A s oldier can only experienc e the real ity o f co mbat o n th e battl efiel d. In each of t hese exam ples, t he w ork er ’s performance is eva luated using some s ort of te sting in a sim ulation or a no ther low stres s e nviro nment. For i nstanc e, a irline pilots a nd soldiers are ofte n tes ted in co mp uter - generate d simulations (Bymer 20 12) (Chang 2013) . Th e perf orm ance o f a work er in a simulator or in an oth er low str es s environment may not be a goo d predictor of hi s performance in a high stre ss situation be cause it is not obvious how he will res pond to the additi onal stres s. So me wor kers m ay thri ve in h igh str ess envi ron ment s, wh ile oth ers m ay suffer a lar ge degr adati on i n their perf orman ce. In th is pap er, show tha t biometric data, in part icular galva nic sk in response (GSR), ca n be used to p red ict p erfo rman ce u nder stres s a nd po te ntially enh ance conventional tes ting methods . Th e rel atio nship bet ween st ress and p erfo rmance w as fir st char acter ized by th e fam ou s Yer kes Do dso n cu rve (Yerk es 1908) whi ch posits th at per for manc e is max imiz ed at a c ertai n level of stres s and d ecr eases as the stres s exc eeds this optima l le ve l. Oth er stu dies h ave fo un d that this relationship holds for a varie ty of task s and typ es of st re ss (Ariely 2009, An ders on 1994, S. L. Bei lock 200 7) . Diff erent th eori es have been pro po sed to try an d expl ain th is rel ationsh ip. On e theor y is that adde d str ess cau ses a shift from automatic t o c ontrolled me ntal proces se s, re sulting in d ecreased perf or m ance (Lange r 1979) , (Ba umeis te r 1984) , (C ame rer 2005) , (Da ndy 2001) . Ano ther t heory i s that the ad ded stress c onsu mes m ental resou rces 2 and interrup t s pr ocedu ral ized rou tines (S . L. Beilock 2007) . Wh ile t hese th eor ies provide differe nt explana tions for the impact of stres s on perform a nce, one criticism is that thes e theor ies are not p redictive because t he o ptim al stress l evel var ies by task an d per son an d als o it ca n be difficult to q uantifiab ly meas ure s tress ( Ne iss 1988) . On e pro po sed way t o m easure st ress inv ol ve s using what is k nown as galvanic skin resp on se (GS R) , wh ich is the electrical c onducta nce of the s kin and de pends on p re - se cretory activity in the sweat glands ( Lader 1962) . GSR w as fir st shown to be co rrel ated wi th str ess in (Jung 1 908) and ma ny studies since have supported this relationship (B oucse in 2012, Pic ard 19 97) . T he recent ubiquity of w earable bio senso rs such as sma rt wa tches a nd wris tbands ha s ma de GSR da ta much easier to co lle ct and has fac ili tated a new bo dy o f experim ental resear ch w hic h uses GS R fo r str ess det ec tio n . It has been shown to be effective for distinguishing varying deg rees of stress in bo th lab studies (Zha i 2 005) a nd real - life se ttings, s uch as highway a nd city driv ing (Heal ey 20 05) . Ex perimen ts have al so sh ow n that GS R can be us ed to infer the diff iculty or cognitive load a ss ociated with a task ( Setz 2010) , (Nou rb a khs h 2012) . Wh ile t he larg e b ody of w ork on GS R h as sho wn that it is c orr elat ed with stress an d task dif ficulty , w h a t has not be en es tab lished is the a bility of GSR to predict the per for manc e o f a w or ker un der stress fu l c onditions . Bein g able to p re di ct performance under s tress using GSR would make it an incredib ly useful signal for a host of a pplications , such as t esting and eval uat ing wo rker s . Here w e sh ow th at GSR measu red while perf ormi ng a ta sk under l ow stre ss is actu all y pre dic tive of per for manc e o f th e same t ask u nd er high str ess . That is, by te sting a pe rson unde r a low stress condition, we can predict how we ll th e y will perfor m when there is in creased str ess . Expe rimenta l setup W e co n duct ed a mu lt i - st a ge exp er ime nt w i t h var yin g d egrees o f str ess , but unif or m task d iff icu lt y t o st udy the re l at i onship b et ween GS R an d per f o rmance . We recru i ted 30 unde rgraduat e studen t s as subj ect s f or t he e xp er ime nt . T h e subjects were t ol d t hey wou l d be part i cipat in g in a stud y of m at h emat ical abi l i t y a nd st ress. Each subject was fi t ted w it h a NEULOG GSR finger se nso r o n t he ind ex a nd mid d le fin ge r s o n t he i r no n - do mina nt ha nd . The G SR si gnal was capt ured through out the wh o l e ex per i men t at a 5 Hz sa mp lin g r at e . T he s u bj ect was fir st asked t o r e lax f o r t h ree min ut e s . A f te r th is rest p e r io d t he y beg a n t hr e e d ist inct ro unds of a c o mputer - bas ed ar i t hmet i c g a me co n s ist ing o f a ser i es o f t hr ee d ig it by two di g i t mu lt i p l i cat ion quest i o n s . Bet ween each ro un d t h e subj ect was gi v en a t hree minu t e r est peri o d. Fig u r e 1 show s a screensho t o f t he game, t h e fi nger senso r used f o r t h e experiment an d a d i agra m depict in g the seque nce o f stages in t he e xpe r ime nt . T he fir st ro und o f t he ga me wa s fo r c a libr at io n a nd designed t o be min ima l ly s t re ss fu l i n o rder to ev aluat e t he su bject s’ ba se l ine per f o r m ance and GSR re spo n se t o t he a r it h me t ic questi ons. Subj ect s answered mu lt i p l ic a t io n questi o ns fo r t en min ut e s . T he y d id no t 3 rece iv e a n y feed back o n w h et h er o r no t t h eir answer s were co rrec t a nd t here was no t ime limi t fo r an swer ing ind ivid u a l quest i ons . Du r ing thi s round, subj ect s were inc e nt iviz e d w it h a s ma l l fin a ncia l r e war d o f $0. 25 f o r each co r rect re spo nse . D ur ing t he rest period f o ll o wing the calibrat i o n ro und, the subjects ’ ave rage re spo nse t ime was calculat ed an d used in the foll ow i ng two rounds f o r the i n ce n tive scheme described bel ow. The s econd an d third rounds of the g a me we re t he lo w an d hi g h stress r o un ds. Both of t h ese ro unds consisted o f 2 0 mu lt ip l ic a t io n quest i on s. Subj ects were giv en 130% o f t hei r average an swer t im e f r o m the cal i brat i o n ro und to answer each quest i o n i n t h ese ro un ds . They were inst ruct ed that a p o r t i o n of the t im e allo wed was “ bonus” tim e, durin g which co r rect an swers were wort h $1 in stead of $0. 25. In the l o w s tr es s r o un d, 85% o f the questi on t ime a l lo w e d was bonus t i me , whereas d urin g t h e hi gh stress ro un d only 50% of t h e t i me was bonus t im e. T h e se percent ages were ch ose n s uc h t h at achieving t h e bo n us pa yo u t in t he lo w st r ess ro und wa s o nly s lig ht ly cha l le ng ing , whereas in t h e hi g h str ess ro und i t was very difficu l t . T he t ime re ma in in g f o r each quest ion was v is ib le t o t he subject s on t h e ga m e scree n a s a ba r t ha t reduced in size e v er y seco n d , w it h bo nu s a nd nor m al t i me represe n t ed by gree n and red shad in g respect i vely , a s s ho wn in Fi gure 1. S ub jec t s w er e a ler t ed whe n t he y ra n o ut o f bo nu s t ime by a n u np lea sa nt buz ze r s o und . Du r ing t hese t wo str ess ro un ds, subject s were al so g ive n feed ba c k ind ic at ing w het he r t he answer t hey s ub m i tte d for each quest ion was corr e ct . Ex p lo ra t o ry Data Analysis and F eature Engine er ing Two different pe rf o rm ance m e t r ic s c a n be de ve lo p ed fr o m ou r exp er ime nt : accurac y on t h e 20 ques t i o ns aske d in ea c h r o und an d m o net ary earning s in e ac h r o und . In o u r ana l ys is , we f o cus on the earnings metr i c as the depende nt vari a bl e o f in t erest because earnings is likely to b e more sens i t ive than a ccur acy t o t h e impact o f st ress on t h e subj ect ’s per f o rm ance . I n ou r Su pplem entar y Mat erial s, w e incl ud e anal ysi s s howing th at th e accur acy is n ot w ell pred icted by e ith er per form ance featu res o r GS R featu res. The ea r ni ngs depend upon how f as t the s ubjec t an swers the in div i dua l questi ons a n d h ow m any ques t i ons are an swere d correct l y, whereas t he accuracy j ust depen ds upo n the nu mb er o f q uesti o ns answered c orrect ly . If a sub j ect i s impact ed by t h e st ress n egat iv e ly , t his wi ll h ave a larg er i m pact on t h e i r ear nings (w hich is se n sit i ve t o speed) t han accu racy (whic h can re ma in h ig h ev en if t he su b jec t slo ws do w n). I f we believe t h at GSR is re lated so mehow to the st ress of the subject , and th at t he str ess i s impact in g perfor m a n ce, t h en we would ex pect the good and b ad perfor mi ng subject s to be m or e easil y d is t inguished in ear nin gs t han in ac curac y . T he s imp le st h y po t hes is is t ha t s ub jec t s w ho pe rfo r m we ll in t he ca l ibra t i o n a nd lo w st ress ro unds w ill also per f o r m w el l in t he hig h st r ess ro und. I f t hi s is t he cas e, t hen pe rf orm an ce f eat ures f ro m t he calibrat i o n an d l ow s t r ess ro u nd s sh oul d be highl y co r re l ated w it h e ar n ing s in t he hig h st r e ss r o u nd . However, we fi nd t hat th es e f eat ure s are no t h ig h l y c o r r e la t e d w it h p erfor m ance i n the hig h er st ress rounds. We plot in Fi gure 2 each subject ’s perfor m ance metr i c s , bot h accur acy and earn i ng s, in calibrat i o n an d lo w st ress ro unds ver sus ea rn ing s in t he hig h str ess ro und. There is n o c l ear ly vis ib le r e la t io n s h i p bet w ee n t he pe r f o r m a nc e me t r ic s across ro un ds . This i s f urt h er suppo r t ed by t h e corr el at i o n coe ff icient of these met r ics, w hic h are n ot st a ti st i cally s i gnifi ca n t at the 4 five percent l e vel fo r e ach per f o rmance m et ric ( see Su pplem entar y Mater ials fo r de t ails ) . T his su gg est s t h at to pred i ct per f o rma n ce under high str ess condit i o ns, we w il l lik ely need m o re inf o r m at ion t han s im ply per f or m ance i n l o wer st ress co n diti o ns . T h is a lso supports o ur specu l at i o n that s imu lat i o n a nd t est ing methods designed to me a su r e a candidat e’s abili t y i n l o wer st ress condi t i ons ar e not necessar ily go o d in dicat ors of rea l - wor l d per f orm ance by themselves . We ne xt lo o k at t he p r ed i ct ive po w er of t he G S R s ig na l w e measure d f o r each sub j ect . Figure 3 s h o ws the tr aj ect o ry of t he raw GSR s ig n a l o ve r the co u r se o f t he e xp er i me nt fo r t w o subj ects, who in t hi s cas e are t he hi g hest and l o we st ea rne rs in t he h ig h str ess ro und . In t h e fi gure, e ac h r o u nd of the trial is s ep arat ed by a s h o rt rest p e r io d , dem ar cated wi t h red lines, and eac h ro un d an d rest peri od o f t he e xp er ime nt is la bele d . T h e G S R s ig na l i s mea su r ed in micr o s ie me ns ( µ S ) whi c h is a uni t of conduct ance. The figure s ho w s t hat th e GS R s i gnal c an ex h i b i t a variet y o f be h av i o rs, such as upwar d an d do wnward trends and rap i d o scil lat i ons. There are also s i g nifica nt diff ere nces in the GSR sig n a l pat h f o r t h ese two subj ect s. We wa n t to de ter mi ne what featur es of t h e GSR si g nal w ill pred i ct perf o r m a n ce i n t he hi g h st r es s ro u nd. T he G SR s i g na l is ver y h ig h d ime ns io na l, w it h o ver 1 0 ,000 sam p l ed val ues per subjec t . To o b tain a useful char act erizat i o n the GSR si gnal f o r each user, we reduce d t he f u l l h ig h d im e n s io na l s i g n a l t o a s e t o f lo w d im e n s io n a l fe a t u r es. O ne o f t he featu re is d rif t whi c h i s defi ned as th e abso l ut e change in t he GSR s ig na l f r o m t h e be g i nning to the en d o f a s in g le r o u nd in t he e xpe r i me nt . D rif t measures any sl o w variat i o n s in t h e GSR s i g n al o v er t he cour se of a round. Ano t her featur e i s the max imum i ncre ase w h ic h is t he a bso l ut e d iffer en ce in t he GS R s ig n a l betwee n i t s min i mu m po int in t he ro u nd t o t he e nd o f t h e r o und . Maximu m i n crease measur es any tr end in the GSR s i g n a l t hat i s p re se nt at t he end o f t he ro und . We co n s ide r ed ot her f eat ures as well , but no ne o f t he m pr o ved s ig n ific a nt . D et ails ca n be fou nd in S up p le me nt a r y Mat e r i a ls. Classif ying p e rfo rman ce Our predi ct i o n task i s to di st in guish between subject s w ho perf o rm well in t h e hi gh stress r o und fro m t ho se w ho do no t usi n g data fro m t h e cal ibrat i o n a n d l ow st ress ro un ds. T h is requ i res us to assi gn a binar y label t o each subject in dicat ing if he is a goo d or b ad p e r fo r me r i n t he h i g h st r e ss r o u nd . As a s imp le la bel in g sc he me, we sa y a subject is a good perf orm e r if his per f o rma n ce metr i c is st rict ly a b ove t h e m ed ian v alue f or al l subject s, otherw i se the subject is a bad p erf orm e r. W e use the m ed ian per f o rma n ce metr i c as a t h resho l d b e cause i t i s ro b ust to o u tliers who have extr em e values o f the metr i c. We c h ecked t h e ro bus t nes s o f o ur fi nd in gs w i t h re s pect to the cl a ssifi cat i o n t hr es ho ld a nd w e a ls o f o un d s i milar s i g nifica nce resu l t s f or reg ressi o n mode l s f or co nt inuo u s ear nings (see Su pp le me nt a ry M at er ia l fo r d et ails) . To g ai n a b ett er underst an ding of our data, we show a scatt e r pl ot o f t h e goo d and b ad perfor m ers i n a two dimensi o nal feat ure space in Fi gure 5 . W e use t h e f eatu res drif t an d maximu m i ncrease b ecau se these set of f e atur es l ead to a goo d separati o n o f t he go od and bad per f o r m ers. I n t h e scatt er pl o t the goo d an d bad perfor m ers are in dicat ed wi t h d i fferent sh aped markers. We see t hat each ty pe of perfor m er o ccup ie s a separate regi o n 5 o f t hi s featur e space. The bad perf o rm er s have a sm a ller m a ximum increase a nd a l ar ger d r i ft . It is not clear wh y t he f eat ures we hav e used in Figure 5 are ef fect iv e at separat in g t h e good an d ba d perf or m ers. From t h e fi gure, i t can be see n t h at the b ad per f o rmers t yp i c a l l y ha ve a l ow maximu m increase a n d a drif t that i s n ear zer o o r ne ga t ive in t he ca lib r at io n ro und . T his m ea ns t ha t du r ing t he ca libr a t io n ro u nd t he GS R s i gn al doe s n ot increa se very m uc h . I f we f o ll o w t heories which suggest that increases in GSR in dicat e an increa se in str ess o r co gni tive l o ad , t he n t he b ad per f o r m er s are not suffering ver y much str es s or c ognit i v e load i n the c alibrat i o n round. In c ontrast, t h e g ood perform ers have a l arge maximu m increase r el at iv e to t h e ir dr i f t . Thes e su bj ect s are experiencing a higher l eve l o f st ress o r cognit i ve l o ad . O ne po s s i ble e xp la n at i o n is t ha t t he ba d perf or m ers do not get them selves f o cus ed in t he a bs e nce o f a ny st re ss, a nd t he n w he n t he st ress i s added, t h ey are n ot abl e to perfor m we ll . T h e goo d perfor m er s, on the ot h er ha nd, a re get t ing t he m s e lv es m e nt a lly focu sed nat urally w i t hout an y add i t i ona l str ess. Then, whe n the exter n al stress i s added, the y are not affect ed by i t bec ause they h av e alread y pr epared thems e l ve s m ent ally. This i s on l y o ne po ssible ex p l an ati o n f o r what we o bse r ve a nd t her e ma y be o t he r e xp la nat io ns. Ho we ver , we do not fo cu s o n t he m because o ur goal is not to exp l ain the re l at ionship bet ween GSR a n d per f o rmance, b ut rat her to pr edi ct who will per f o rm well u n der st ress. There f o re, we will not di scuss t hes e exp l a nat i o n s fu rt her , but i nst ead buil d a pr edict i ve m ode l for perfor m a nce under st ress. To eva l uat e the st a tis t ical sig nificance o f these appar en t relat i o nsh i ps, we de v el o ped lo g i s t ic regr essi o n m o dels f o r t he above me d i a n h ig h s t re ss e ar n ing s la be l s ho wn in Figure 3. For t h e per for m a nce f eat ures, we considered accuracy, earning s a nd aver age answer t im e a nd fo r t he GSR f eatu res, we c onsidered d r i ft and ma x imu m incr e as e, a ll fro m bo t h ca lib r at i o n a nd l o w s t r ess ro u nds. T a b le 1 sho ws est im at ion r esu l t s for m od els w it h t hr ee d i fferent f eat ure set s: per f o rm a nce f eat ures, GSR f eatur es, and pe rf orm an ce and GSR feat ures co m bined. We co m par ed mode ls accor din g t o featur e si gnific ance and Akaike Infor m at i on Cr i t er i on (AIC) and repo r t in the tab l e t he two s t r o n gest m o dels f ro m each f eat ure set. We see t h at the m ode ls b ui l t w i t h performa n ce feat ures al o ne cont ain featu res from the ca lib rat ion ro un d o nly — no perf or m a nce features fro m the l ow st ress ro un d were f o und to b e signifi cant pr ed i ct ors o f high st ress earnings . The best p e r fo r ma n c e f eatur e mode l s co nta in onl y o n e weak ly sig nifi c ant pr edi cto r each. T h e t w o best GSR f eatu re m ode ls pai r dr if t a n d maximu m increa se f eat ures fro m t he same period. The calibrat ion per i o d m od el exhibits b ot h greater st a t i st ical s ignificance f o r both pred i ct o rs an d a l o wer AIC ( w h ic h in d ic a t e s b e t t e r mo d e l f it ) . F ina l ly, t he be st co mb ine d featu re m o de ls are t he same as t he b es t GS R f eatu re mo d e ls w it h t h e a d d it io n o f t h e calibrat ion accurac y feat ure, whi ch is no t st at ist ic a lly s ig ni fic a nt in e i t he r mo de l . Ho w ev er, t h e A IC f o r t h e combi ned m ode ls doe s imp r o ve s lig ht ly o ver t he GS R fea t u re mo d e l s. Predicti ve Power To estimate ho w m uc h o f an impro v e m ent GSR featu res o f f er o ver perf o r m a n ce f eat ures f or c l a ssifying good a nd ba d per f o r m er s, we eval uate t he p red i ct ion accurac y f o r t he six mo de ls s ho w n in T ab le 1. The predictiv e po we r of o ur m o del s is evalua t ed usi ng 6 st r at ifie d, fo u r - fo ld c r o s s - va lid at io n. T hi s t ec hnique inv o l ves s eparat in g the data i n t o f o ur separat e sets o r f o lds wi t h b ala nced numbers o f eac h out co m e c la s s i n a ll fo ld s. T he m ode l being t est ed i s fi t o n t hree of the f our f olds and predictions are g en erat ed o n t h e fo ur t h he l d- o ut fo ld . Th is fi t and pred i ct pro cess is repeat ed four t i mes, with a di f fere n t f o l d being h eld out i n o rder to gene rat e a p redicti o n for each dat a po int . T he n, w e use t he se ho l do ut p r e d ic t io n s t o d e v e lo p r eceiver o perating charact er i st ic (ROC ) cu rv e , a v isu a liz at i o n o f the t r ade - off betwee n t he t ru e pos it ive and false pos i t iv e rat es of y ou r pr e di c ti ons gi v en dif f e ren t va lu e s fo r t he d e c is io n r u le . To assess the pred i ct iv e accurac y, we calcu l at e t h e area under t he curve (AUC) . For context , an AUC o f 0. 5 indicat es that a pred i ct iv e mode l is no bet ter than rando m guessin g and the closer an AUC is to 1 .0 , t he m or e accurat e i t i s co nsidered t o be. We generat e d 500 di fferen t p e r m u ta ti on s of s tra ti fi ed fo ld s to ev aluate the resu l ts ’ sensi t ivi t y to the da ta spli t s a nd r an t h e above t echniques on each o ne . Fi gur e 5 s ho ws b ox pl ots f or th e di s tri b uti on of A UC s f or e ac h m od el ov er th e 500 p e r mu t a t io n s . N ote fro m t he figure, per f o r man ce f eat ure m od els per f o rm no bett er t h an random guess i ng ( AUC = 0.5 ) whereas the GSR f eat ur e m o de l s achie v e a media n AU C o f 0. 76. It i s als o wor th n ot i n g t h at the com bi ned fea t ure m o dels do not appea r to achieve hig h er accur acy t h an the GSR f eat ure m o de l s, suggest i ng t ha t pe rf orm an ce featu res pro vi de ver y li t t l e if any pred i ct ive infor mation not al read y conta in ed in the GS R featu res f or o ur da ta . Disc ussion Our resul t s dem onst rate t h at G S R s ig n a l c a n b e us ed to predi ct t h e perf or m a nce o n co gni t iv e tasks in hig h str ess c on di t i on s . An impo rt an t aspect o f o ur w o rk i s the f a ct t h at t h e GSR s i g nal used f or predict i o n is obtained under ca li brat i o n an d low st ress c on di ti on s . Therefore it is tr uly pred ict iv e o f per f o r m a nce under hi g h stress co ndi t i on s. We do not y et ha v e a clear ex p l a nat i on as to the mechanism of this p heno me no n , althou gh we suspect i t i s re l at ed to t h e f act that good perfor m er s will get f o cused an d menta lly pr epar ed for a t ask wi t h o u t the need f o r ad di t i o nal st ress o r p re ssure. No net he le ss , e ve n w it ho u t a cle ar exp la na t i o n fo r t he m ec ha n is m, w e are st il l a ble to accurately pred i ct the g ood an d b ad perf or m ers usin g GS R . T hi s sugges t s that by using wearable bi o se n sor s, o ne can evaluat e t he p e r fo r manc e of wo rkers un der dan gero us , hi g h r is k , high st ress con di ti ons using data fro m sa f er , l o w r isk , l o w str ess c o n d it io n s . W e ant i c ipate t hat o ur ex per im ent wi ll lead t o mo re ana lys is on t he pred i ct iv e power of GSR fo r worke r perf o rm ance und er s tress in a v ar iet y o f ot her t asks. Referen ces Anderson, K riste n Joan. 1994. "Impulsitivity, ca ffe ine, a nd task difficulty: A within - subjec ts t est of t he Y erke s - Dods on la w. " Pers onality and Indi vidu al Diff erences 813 -- 829. Ari ely, D an and G neez y, Uri an d Lo ewenst ein , Geo rg e and M azar, N ina. 2 009 . "L arge sta kes and big mis takes." Th e R eview of Econom ic S tudie s 451 -- 469. Baume iste r, R oy F. 1984. "C hoking under pre s sure: s elf - co nsci ou sness an d paradoxica l effec ts of incentiv es on skillful performance ." Journ al of person ality an d s ocial psy ch ology. Bei lo ck, Si an L and DeCaro , Mar ci S . 2007 . "From poor perf orman ce to success u nder str ess: wo rki ng m emor y, strat egy sel ectio n, and m athem atic al pr obl em 7 solvi ng under pres sure ." Jou rnal of E xperi menta l Psy cholog y: Learn in g, Me mory, an d Cogni tion . Beilock, Sian L. 200 8. "Math perf ormance in s tre ssf ul situations." Curre nt Dire cti ons in Psych ologic al Sci enc e 339 -- 3 43. Boucse in, Wolfram. 2 012. Electrode rmal acti vi ty. Springer Scienc e & Busine ss Media. Bymer, Maj . Lore n. 2012. Virtual re ality u sed to train Soldie rs in new trai nin g si mulator. August 1. ht tp:// www.a rmy.mi l /a rti cle/844 53 /. Cam erer, Co li n and Loew enstei n, Geo rg e and Pr elec, Dr azen . 200 5. "Neuroeconomics: How neuros cience can inform economics ." Journ al of ec onomi c Li terature 9 -- 64. Cha ng, H iro. 201 3. Simu lators alway s va luable i n mi litary traini ng. April 13. http:/ /www.a r my.m il/art icle/1959 9/simulator s - al ways - v aluabl e - in - military - training/. Dandy, Justine a nd Brewer, Neil and Tott man, Robin. 2001. " Self - co nsci ousness a nd performance de crements within a sporting cont ext. " T he Jou rnal of soci al psyc hology 150 -- 152. Healey , Jenn i fer, and Rosalind W. P icard. 2005. "Dete cting stres s during real - world driving task s using phys iological sens ors." Inte lli gent T ransporta tion Syste ms, IEEE T ran sacti ons on 156 - 1 66. Jung, Ca rl G usta v. 1908. " On psychophys ical rel ations of the a ss ociative e xperiment." Th e Jou rnal of Abn ormal Ps ych ology 247 -- 255. Lade r, MH an d Montagu, JD . 1962. " The psy cho - galv anic r eflex: a phar maco lo gic al stu dy o f the p erip her al mechan ism. " Journ al of n eurology, neu rosurge ry, a nd psyc hi atry. Lan ger, E ll en J and Imber , Loi s G. 1 979 . "W hen prac tice m akes im perfec t: debilitating e ffec ts of overlea rning." Jou rnal of p erson ality an d soci al psyc hology. Neis s, Rob. 1988. "Reconceptualizi ng arous al: psy chobiological state s in motor per for manc e." Psy chologi cal Bu lletin 345. No urbak hsh, N., Wang , Y., Ch en, F., & Cal vo, R. A. 201 2. "Usin g galvan ic skin resp onse for cognitive load measurement in a rithmetic a nd reading tas ks ." Proc eedi ngs of th e 24th Austra lian C ompute r - H uman I nteraction Co nference 420 - 423. Pic ard, Roalin d. 199 7. A ffect iv e com putin g. MI T pr ess . Setz , C., Arnri ch, B ., Schu mm, J., La Marca, R. , Tr oster, G ., & Ehl ert, U. 20 10. "Discriminating s tress from cognitive load usin g a we arable EDA device." Inf ormati on Te chn ology i n Biome dicin e, IEEE T ransac tions on 410 -- 417. Yer ke s, R . and Dodson J. 1908. "The relation of stre ngth of stimulus to rapidity of habi t - formation." Journ al of c omparative neu rology and psyc hology 459 -- 482. Zhai , J., Barret o, A. B., Ch in, C., & Li, C. 20 05. " Zhai, Jing, et al. "Re aliz ation o f stress det ecti on u s ing psychophys iological signa ls for impr ovement of human - computer intera ctions." South eastC on, 200 5. Pro ce edi ngs. IEEE 415 -- 420. 8 Fig ur e 1. ( a) A sc r een shot o f the computer in terface used for the e xper im en t. (b) T he N eu l o g GSR s ens or u sed f or th e ex p eri me nt. (c) T h e sequence of stage s in the expe r ime nt, beginni ng on the left wi th a res t stage fo r the subje ct and e nding on the r igh t with a h igh stre ss stag e . 9 Fi gur e 2. Sc atter p lots of the perf orma nce m etric s in d iff e rent rounds : (a) ca lib rati on earni ngs verus high stress e arnings, (b) low stre ss earnings ver sus high str ess earn ings, (c) c alib r ation a ccur acy versu s high stress accuracy, and (d) low stre ss accuracy versus h igh stress accuracy. The c orrela tio n coe ffici ent for ea ch p ai r of m etrics plotted i s not s ig nifi cant a t the f ive perc ent level . 10 F ig ur e 3 . Plo ts of the GSR sign als during the e xperiment f or the highe st and lowe st e a rning s ub ject s . T he di ff er e nt r ou nds a nd r est p er iods a r e dem a r cat ed b y t he ve r tical line s on e ach plo t. Fi g ur e 4 . S ca tt er pl ot s of subj ec ts’ calib ratio n d rift v e rsus ca lib ratio n ro u nd m axi mum i nc rease wi th diffe rent marke rs f or goo d an d b ad pe r f o r mers in t he hig h st ress r o und in t e rms of (a) ear nin gs an d (b) accu r acy. 11 Tab le 1 . Pre dic to r c o eff ic ie nts a nd AIC fo r six b es t lo gis tic re gres sio n mo de ls 12 Fi g ur e 5 . Box plo ts of mode ls AU Cs dev elope d w ith 500 permutatio n s o f stratifie d, 4 - fold cr os s - v alidat io n.