Neural processing of the same, behaviourally relevant face features is delayed by 40 ms in healthy ageing

Fast and accurate face perception is critical for successful human social interactions. Face perception declines with age both in behavioural and neural responses, although we do not yet understand why. Here, we tested the hypothesis that early brain mechanisms involved with face information processing are delayed in older participants. Using face detection - the most basic task for social interaction – we sampled visual information from faces (vs. noise) and reconstructed the features (mainly, the left eye) associated with detection behaviour in young (20-36 years) and older (60-86 years) adults. We then compared behavioural results to neural representations of face features revealed with simultaneously recorded EEG on the N170, an event-related potential associated with visual categorization. Whereas the right hemisphere N170 latency and amplitude represented the left eye in young participants, it was mostly amplitude that represented the eye with a 40 ms delay in older adults. Our results demonstrate that face processing speed declines in ageing with a delay in the early stages that process the visual information important for behaviour.


INTRODUCTION
Ageing impairs social tasks such as face perception, including face detection ( capacities, such as acuity or contrast sensitivity, is also likely involved (Boutet et al., 2015). 12 Indeed, neuroimaging studies provide evidence for many age-related changes in the human 13 visual system, which in turn could lead to slower processing (Rousselet et al., 2009(Rousselet et al., , 2010  In addition to delayed visual processing, evoked responses in older adults might also reflect 28 increased de-differentiated processing-e.g. increased processing of task-irrelevant 29 information (e.g. non-diagnostic face features), or increased false alarms (e.g. noise textures 30 processed as meaningful stimuli). Indeed, some studies have suggested an age-related 31 increase in brain responses to non-preferred stimuli in visual areas that respond 32 11]; group difference = -11 PP [-14, -7]). Specifically, older participants had higher numbers 23 of noise responses on face trials, while both groups were similarly accurate on noise trials 24 (face trials, young = 91% correct vs. older = 75% correct; group difference = 15 PP [8,19]; 25 noise trials, young = 95% vs. older = 92%; group difference = 4 PP [1,13]; group difference 26 of face-noise difference = 7 PP [0, 15]). 27 To understand the face information associated with behaviour, we used MI to compute the 28 relationship between pixel visibility and correct vs. incorrect responses. Eye region pixels 29 were strongly associated with correct responses in only a few young participants and almost 30 all older participants, suggesting that young participants used any feature to do the task, in 1 contrast to older participants who needed to see the eye region to correctly detect a face (N 2 young = 4/17, N older = 16/18, Figure 1A-B, third panel). This reliance on the eyes was 3 confirmed in the average classification image of the difference between the groups (see 4 Figure 1C, third panel). 5 6    were different between the groups, with 12 older participants showing stronger MI than the maximum 2 across young participants for MI(PIX, CORRECT); and 4 older participants for MI(PIX, RT). Images on 3 the right display the differences between young and older average MI maps for each response 4 measure and stimulus category.

6
We confirmed these results with a reverse analysis (see Materials  older participants ( Figure 2). For the right eye, these gains were respectively 17 ms [10,24] 10 and 7 PP [4,11] in young participants; and 16 ms [2,30] and 28 PP [19,36] in older 11 participants (see Table S3 in Supplementary Material for effect sizes of group differences).   To demonstrate a difference in behavioural strategy between the groups, Figure 3 (bottom 1 panel) shows that on trials without any eye visibility (either left or right), young participants 2 could still detect faces accurately, whereas older participants could not (bin 1, young: min = 3 56%, median = 72%, max = 92%; older: min = 9%, median = 31%, max = 94%; see also 4 Supplementary Figures S2 -S4 for reaction times, and for noise trials). 5 In sum, our behavioural RT and accuracy results reveal that all participants used the left eye 6 to detect faces from noise. However, whereas the eyes were all older participants could use 7 to make correct responses, young participants could also use any other face feature; 8 demonstrating a strategy difference in older participants.    We used MI to reveal the features that modulated single-trial EEG responses on the left and 8 right lateral-occipital electrodes (i.e. LE and RE) and midline occipital electrode (Oz) 9 between 0 and 400 ms post-stimulus. As shown in Figure 5, presence of the eye 10 contralateral to the recording electrode modulated EEG responses (see Figures 5A and 5B), 11 with a stronger response to the left eye recorded at RE in both groups. 12 Eye sensitivity exceeded a family-wise error rate permutation threshold in a smaller number 13 of older than young participants and was overall weaker in older participants (see Figures Figure 5A, bottom panels) revealed weaker 18 associations and sensitivity to various face features (eyes, chin, mouth, nose, and forehead) 19 in some participants in both age groups on face trials. 20 To rule out a mere effect of spatial attention (rather than eyes per se), we computed MI on 21 noise trials and found no systematic sensitivity to the eyes in either group ( Figure 5), 22 although there was some sensitivity to the left cheek area in a few older participants. 23 We also ensured not to miss any effects by computing the same classification images across 24 all electrodes; they showed sensitivity to the left eye region in both groups (see Figure S4 in 25

Supplementary Material). 26
In sum, these results extend former results of contralateral eye coding in face detection in 27 young participants (Rousselet et al., 2014;Ince et al., 2016), and add the weaker association 28 for older adults, which contrasts with their stronger reliance on the eyes for behavioural RT 29 and accuracy.       Knowing what face information was associated with ERP responses, we then investigated 12 how this relationship unfolded over time. To this aim, we plotted the maximum MI across 13 pixels in each classification image between 0 and 400 ms post stimulus (see Figure 6B), and 14 computed the MI peak latencies for young and older participants.  We also confirmed weaker eye sensitivity in older adults across time: the peak MI amplitude 20 in older participants was about 58% of that of young participants at LE and 57% at RE (95%    Having quantified an age-related difference in the timing of maximum eye-sensitivity, we 2 wanted to determine if this timing difference was also present in the earliest measurable 3 EEG eye sensitivity, and if this difference could be explained by a non-specific delay in 4 visual activity. To answer these questions, we used causal-filtered data, to more precisely 5 identify the timing of early effects (see Figure 6A; see also Materials and Methods). 6 First, we measured the onsets of MI to the eye features. The results suggest that the earliest 7 eye sensitivity is already delayed in older participants.  Importantly, we found very weak differences in the onsets of ERP STD across the two groups 20 (difference = -0.5 ms [-7, 5]), suggesting no general delay in the onset of visual cortical 21 activity in older participants. 22 As such, our results suggest that the observed delay in the processing of the eye region is 23 not due to a general age-related delay in the initial activation of the occipital-temporal cortex. 24 Instead, the eye processing delay seems to occur later in the visual cortical processing 25 pathway, thus ruling out low-level optical factors as the main contributor to the delay.   To uncover the functional role of the N170 in coding task-relevant information in older adults, 11 using a reverse analysis we sought to directly investigate how eye processing related to the 12

participants. 17
Specifically, the N170 reconstructed from trials with high contralateral eye visibility on RE 18 preceded and was larger than the N170 reconstructed from trials with low contralateral eye 19 visibility (latency effect, young = 24 ms [17,31] vs. older = 5 ms [-2, 11]; for amplitude 20 effects, see Table 5). This latency effect was weaker at LE (young = 12 ms [8,17]; older = 1 21 ms [- 6,9]), stronger in young than in older participants (for effect size estimates, see Table  22 4), and stronger at RE than LE in young compared to older participants (young = -7 ms [-12, 23    To understand visual information processing in ageing, we must start by asking what 2 information the aged brain processes and when. Here, for the first time in a sample of older 3 participants, we address these two questions by using reverse correlation to link facial 4 stimulus space to behavioural and brain responses. 5 In terms of behaviour, older adults used pixels in the eye region to detect faces, similarly to 6 young adults. In particular, pixels in the left eye region were associated with faster reaction 7 times in both young and older participants, although the association was stronger in older 8 than in young participants. Both groups were also more accurate when the left eye was 9 visible. However, whereas young participants performed well above chance even when there 10 was no eye visibility on a given trial, older adults struggled to respond correctly on those 11 trials and performed below chance, with a bias towards reporting face absence. As such, 12 young adults were able to do the task based on any feature revealed through Bubble masks, 13 whereas older adults were heavily dependent on the presence of the eyes to detect faces. Goeckner, 1989). Although we tested each participant's contrast sensitivity and visual acuity, 32 unfortunately at present we cannot rule out a level of blur for the older group, given that 1 participants wore their habitual visual correction -which might have been insufficient to 2 allow clear vision at the viewing distance in our study. However, the fact that we observed 3 larger N170 to noise textures in older than in young participants suggests that blur is an 4 unlikely factor, as it should affect all responses irrespectively of their category. In addition, 5 blur should affect all response latencies. However, cortical onsets were very similar between 6 groups, providing another argument against the effect of blurred vision on observed 7 behavioural or neural responses. Having established what information participants use to perform a face detection task, we 9 quantified when and where that information modulated brain activity. In young and older 10 participants alike, we found that single-trial ERPs are mostly associated with the presence of 11 eye pixels contralateral to the lateral-occipital recording electrodes. This association 12 (measured with MI) was also stronger at right hemisphere electrodes in both groups, in line 13 with the right hemisphere dominance for face processing (Sergent, Ohta, & MacDonald, 14 1992). However, MI was, on average, weaker in older participants. MI time courses also 15 peaked about 40 ms later in older than in young participants suggesting that sensitivity to the 16 same face feature is weaker and delayed in ageing. Importantly, there was no general delay 17 in the onset of visual cortical activity in older participants, suggesting that the delay observed 18 at lateral-occipital electrodes occurred at cortical information-processing stages and was 19 unlikely to be due to retinal factors (Bieniek et al., 2013(Bieniek et al., , 2015. 20 The eye sensitivity peaked about 10 ms before the peak of the N170 in young participants, implying that the functional role of the N170 remains the same across the two groups. 33 However, this processing is associated with a different temporal pattern, where information 1 peaks at the same time as the N170 in older participants and is only associated with a very 2 weak change in latency. 3 Our behavioural and EEG results suggest a double-dissociation in age-related differences in 4 face processing: a stronger reliance on the eyes in making behavioural judgments is coupled 5 with weaker and delayed brain sensitivity to these features in older adults, relative to young 6 adults. A similar dissociation was reported in an fMRI study investigating face perception 7 from images degraded with noise in young and older adults (Grady, Randy McIntosh, 8 Horwitz, & Rapoport, 2000). The highest correlation between brain activity and behavioural 9 performance in a face matching task was found in the fusiform gyrus in young participants, 10 but in posterior occipital regions in the older adults (Grady et al., 2000). In addition, two other 11 areas -thalamus and hippocampus -showed positive associations with behaviour in a 12 sample of older participants only, suggesting functional plasticity in the recruitment of brain 13 areas responsible for face processing in old age. In line with these findings, a more recent Specifically, if other brain areas contributed to processing of faces in older adults in the 24 current study, then restricting the analyses to occipital-temporal sensors in the left and the 25 right hemisphere could lead to missing or poorly quantifying effects. We ensured we did not 26 miss any local effects by running the MI analysis on all electrodes and visualising maximum 27 MI across electrodes. Whole-scalp results, however, were very similar to those obtained at 28 the electrodes of interest analysis, suggesting that occipital-lateral electrodes showed 29 maximum sensitivity to the eye region in both young and older participants. Furthermore, 30 results from a multivariate MI analysis between eye sampling and top PCA components 31 showed that considering whole-scalp distribution of EEG activity in the two groups did not 32 alter the difference in maximum MI or age-related delay. Altogether, analyses restricted to a 33 single lateral-occipital electrode in each hemisphere were sufficient to describe age-related 1 differences in processing of facial information in our study. 2 The age-related delay in processing of the eye could not be attributed to the presence of 3 Bubble masks either. Bubbles can be thought of as a form of masking procedure that 4 degrades the visual input and has been suggested to entail object completion (Tang et al., 5 2014). Processing occluded stimuli by the visual system may require additional resources to 6 perform the task, leading to longer processing times (Sekuler, Gold, Murray, & Bennett, 7 2000). As such, any delay observed in a sample of older adults could be due to a 8 combination of factors: a genuine slowing down of processing speed, as well as an increase 9 in the time needed to process the occluded stimulus with respect to young adults. However, 10 our ERP results show that the processing time of Bubbled images compared with full images 11 was not different in young and in older participants. Specifically, even though processing of 12 the Bubbled stimuli was delayed with respect to full images by about 20 ms in both young 13 and older participants, there was no interaction between age and masking condition. In both 14 practice (unmasked) and Bubble (masked) trials, the N170 latency to face images in older older adults were less accurate on Bubble trials compared with practice trials, but the drop in 28 performance was much more pronounced across older participants. 29 The reason for such age-related deterioration in performance on tasks involving perception 30 of fragmented pictures or perceptual closure remains elusive. It has been suggested that 31 perceptual difficulties arise as a consequence of heightened noise or variability associated 32 with internal stimulus representation in the neural system (Salthouse & Lichty, 1985), or as a 1 result of a deficit in inhibitory control of interfering/irrelevant information (Lindfield et al., 2 1994). In our study, ERP variance across Bubble trials, measured at the time point of max 3 MI was slightly lower in older than in young participants (see Supplementary Figure S6 To summarize, our results provide the first functional account that advancing age involves 26 differences in the earlier stages of processing visual information important for behaviour. 27 Specifically, we show for the first time that the information content of early visual ERPs in 28 older adults does not differ from that of young adults. While the contralateral eye region 29 modulates ERPs in young and older adults alike, information processing is weaker and 30 delayed in ageing. Furthermore, ageing affects coding of the eye by the N170 differentially: 31 whereas eye visibility is associated with an amplitude change in older adults, it is associated 32 with both a latency and amplitude change in young adults. These ERP findings are coupled 33 with an increased reliance on the presence of the eyes to produce behavioural responses in 1 older adults, suggesting a change in strategy with age. 2

PARTICIPANTS
2 Eighteen young (9 females, median age = 23, min 20, max 36) and nineteen older adults (7 3 females, median age = 66, min 60, max 86) participated in the study. Results from fifteen of 4 the young participants have been reported previously (Rousselet et al. 2014). All older adults 5 were local residents, recruited through advertising at the University of Glasgow, active age 6 gym classes, and a newspaper article. Volunteers were excluded from participation if they 7 reported any current eye condition (i.e., lazy eye, glaucoma, macular degeneration, 8 cataract), had a history of mental illness, were currently taking psychotropic medications or 9 used to take them, suffered from any neurological condition, had diabetes, or had suffered a 10 stroke or a serious head injury. Volunteers were also excluded from participation if they had 11 their eyes tested more than a year (for older volunteers) or two years (for young volunteers) 12 prior to the study taking place. Two older participants reported having cataracts removed, 13 and one older participant reported having undergone a laser surgery. These participants 14 were included because their corrected vision was within normal limits. Participants' visual 15 acuity and contrast sensitivity were assessed in the lab during the first session using a 16 The study was approved by the local ethics committee at the College of Science and 8 Engineering, University of Glasgow (approval no. FIMS00740), and conducted in line with 9 the British Psychological Society ethics guidelines. Informed written consent was obtained 10 from each participant before the study. Participants were compensated £6/h. 11

12
We used a set of 10 grey-scaled front view photographs of faces, oval cropped to remove 13 external features, and pasted on a uniform grey background (Gold, Bennett, & Sekuler, 14 1999). The pictures were about 9.3° x 9.3° of visual angle; the face oval was about 4.9° x 15 7.0° of visual angle. A unique image was presented on each trial by introducing phase noise 16 (70% phase coherence) into the face images (Rousselet, Pernet, Bennett, & Sekuler, 2008). 17 Textures were created by randomising the phase of the face images (0% phase coherence). 18 All stimuli had the same amplitude spectrum, set to the mean amplitude of the face images.  we asked participants to minimise movement and blinking, or blink only when hitting a 3 response button. The viewing distance of 80 cm was maintained with a chinrest. 4 In each experimental session, participants completed 12 blocks of 100 trials each while 5 seated in a sound-attenuated booth. The first block was a practice block of images without 6 bubble masks. A set of 10 face identities and 10 unique noise textures, each repeated 5 7 times were randomized within each block. Each session lasted about 60 to 75 minutes, 8 including breaks, but excluding EEG electrode application. 9 Within a block of trials, participants were asked to categorise images of faces and textures 10 as fast and accurately as possible by pressing '1' for face, and '2' for texture on the 11 numerical pad of a keyboard, using the index and middle finger of their dominant hand. After 12 each block, participants could take a break, and they received feedback on their 13 performance in the previous block and on their overall performance in the experiment 14 (median reaction time and percentage of correct responses). The next block started after 15 participants pressed a key indicating they were ready to move on. 16 Each trial began with a small black fixation cross (12 x 12 pixels, 0.4° x 0.4° of visual angle) 17 displayed at the centre of the monitor screen for a random time interval of 500 to 1000 ms, 18 followed by an image of a face or a texture presented for 7 frames (~82 ms). After the 19 stimulus, a blank grey screen was displayed until the participant responded. The fixation 20 cross, the stimulus and the blank response screen were all displayed on a uniform grey 21 background with mean luminance of ~43 cd/m 2 . 22

23
EEG data were recorded at 512 Hz using a 128-channel Biosemi Active Two EEG system 24 (Biosemi, Amsterdam, the Netherlands). Four additional UltraFlat Active Biosemi electrodes 25 were placed below and at the outer canthi of both eyes. Electrode offsets were kept between 26 ±20 µV. detrended. Two types of filtering were then performed. First, data were band-pass filtered 30 between 1 Hz and 30 Hz using a non-causal fourth order Butterworth filter. Independently, 1 another dataset was created in which data were pre-processed with fourth order Butterworth 2 filters: high-pass causal filter at 2 Hz and low-pass non-causal filter at 30 Hz, to preserve Data from both datasets were then downsampled to 500 Hz, and epoched between -300 and 6 1000 ms around stimulus onset. Mean baseline was removed from the causal-filtered data, 7 and channel mean was removed from each channel in the non-causal-filtered data in order 8 to increase reliability of Independent Component Analysis (ICA) (Groppe, Makeig, & Kutas, 9 2009). Noisy electrodes and trials were then detected by visual inspection of the non-causal 10 dataset, and rejected on a subject-by-subject basis. Following visual inspection, one young 11 participant and one older participant were excluded from further analyses due to noisy EEG 12 signal. Mutual Information (MI) analysis confirmed the lack of sensitivity to any facial features 13 in these participants. The resulting sample size was 17 young and 18 older participants. In 14 this sample, more noisy channels were on average removed from older than from young 15 participants' datasets (older participants: median = 10, min = 0, max = 24; young 16 participants: median = 5, min = 0, max = 28; median difference = 4 [2,7]). More noisy 17 Bubble trials were also removed from older than from young participants' datasets (trials single-trial spherical spline current source density waveforms using the CSD toolbox (J. 33 Kayser, 2009;Tenke & Kayser, 2012). CSD waveforms were computed using parameters 50 1 iterations, m=4, lambda=10 -5 . The head radius was arbitrarily set to 10 cm, so that the ERP 2 units are µV/cm 2 . The CSD transformation is a spatial high-pass filtering of the data, which 3 sharpens ERP topographies and reduces the influence of volume-conducted activity. CSD 4 waveforms also are reference-free.

19
We compared the amplitude and latency of the N170 between the two age groups. To this 20 end, we computed mean ERPs across trials for each participant, separately for face and 21 noise trials, and for practice (without Bubbles) and regular (with Bubbles) trials. For ERPs 22 recorded at the lateral-occipital electrode in the right hemisphere (RE), we defined the N170 23 peak in individual participants as the minimum mean ERP between 110-230 ms, and 24 considered separately its latency and amplitude. We estimated the size of the between-group differences using two robust techniques: Cliff's 2 delta and the median of all pairwise differences. Cliff's delta (Cliff, 1996;Wilcox, 2006) is 3 related to the Wilcoxon-Mann-Whitney U statistic and estimates the probability that a 4 randomly selected observation from one group is larger than a randomly selected 5 observation from another group, minus the reverse probability. Cliff's delta ranges from 1 6 when all values from one group are higher than the values from the other group, to -1 when 7 the reverse is true. Completely overlapping distributions have a Cliff's delta of 0. In line with 8 Cliff's delta approach, we also calculated all pairwise differences between young and older 9 participants on the measures of interest (reaction times, percent corrects, N170 latencies 10 and amplitudes), and took the median of the distribution of these differences. This way of 11 measuring effect sizes enabled us to provide information about the typical difference 12 between any two observations from two groups (Wilcox, 2012). 13

14
We used mutual information (MI) to quantify the dependence between stimulus features and 15 behavioural and brain responses. MI is a non-parametric measure that quantifies (in bits) the Here, we calculated several MI quantities in single participants: MI(PIX, RT) to establish the 2 relationship between image pixels and reaction times; MI(PIX, CORRECT) to establish the 3 relationship between image pixels and correct responses; MI(PIX, RESP) between pixels 4 and response category; and MI(PIX, ERP) to establish the relationship between image pixels 5 and ERPs. These quantities were computed separately for face and noise trials. To control 6 for the variable number of trials in each participant arising as a result of EEG preprocessing, 7 we scaled every MI quantity for every participant by a factor of 2Nln2 (Ince,Mazzoni,8 Bartels, Logothetis, & Panzeri, 2012), using the formula: 9 where MI refers to mutual information values, and Nt is the number of trials. MI scaled , 11 therefore, reflects a measure of MI adjusted for a systematic upward bias in the information 12 estimate that might arise due to limited data sampling, especially if the numbers of trials in 13 the two age groups are systematically different. It also converts MI to be the effect size for a 14 log-likelihood test of independence (Sokal & Rohlf, 2012). All group-difference analyses 15 were performed using the scaled MI values. 16

17
We refer to MI between pixels and behaviour or ERPs as classification images: they reveal maxima. As such, clustered pixels will get higher TFCE scores than individual ones, which 1 combined with standard permutation testing alleviates the problem of multiple comparisons 2 across many pixels ( classification images and saved the maximum delta score. This procedure was performed 18 1000 times in order to obtain a distribution of maximum delta scores under the null 19 hypothesis that there are no differences in the classification images of the two age groups. 20 We then compared the original delta scores against the 95 th percentile of the permutation 21 distribution. 22

23
To quantify how the presence of the eyes modulated behavioural and brain responses, we 24 ran a reverse analysis Rousselet et al., 2014). First, we created the eye 25 mask by centring a circle (radius = 15 pixels) on the pixel that showed the maximum MI 26 value in the group-averaged MI(PIX, ERP) classification image, separately for the left and for 27 the right eye. We then summed pixel values revealed through single-trial Bubble masks that 28 fell within the boundaries of each eye mask independently, and within both eye masks 29 together, to provide an estimate of eye region visibility. We then split these values into ten 30 equally populated bins ranging from the lowest to the highest sum values and compute the 31 median RT and the mean percent correct for each bin. Next, we quantified the effect of eye 32 visibility on behavioural judgments by calculating the RT and percent correct difference 1 between the 10 th and the 1 st bin. We then repeated this analysis with single-trial ERP 2 distributions: we averaged the ERPs corresponding to each bin, separately for the left and 3 right lateral electrodes. We then computed the N170 amplitude and latency in every 4 participant and for each eye mask, for the lowest (1 st bin) and the highest (10 th bin) sum 5 values, separately for the left and right electrodes. Given that the N170 on Bubble trials was 6 delayed with respect to that on practice trials in both groups, we defined the N170 as the 7 minimum in the time window 150 to 250 ms following stimulus onset in ERPs low-pass 8 filtered at 20 Hz using a fourth order Butterworth non-causal filter. We then computed the 9 differences between high and low amplitude and latency values for each group separately. 10

11
We quantified ERP onsets using the causal-filtered datasets. To control for multiple course for each individual participant and mean baseline centred it. Then, we localised the 23 first peak whose minimum height was five times the height of any peak in the baseline. 24 Then, using ARESLab toolbox (Jekabsons, 2015), we built a piecewise-linear regression 25 model with three basis functions using the Multivariate Adaptive Regression Splines (MARS) 26 (Friedman, 1991) method. Onsets were defined as the location in time of the first knot. 27 MI onset. We quantified MI onsets using the same technique as with ERP STD onsets. 28

29
Topographic maps for each participant were computed from the whole-scalp MI(PIX, ERP) 30 results at the individual MI peak latency. Individual topographic maps were normalised 31 between 0 and 1, interpolated and rendered in a 67 x 67 pixel image using the EEGLAB 1 function topoplot, and then averaged across participants in each age group. Using the 2 interpolated head maps, we then computed a hemispheric lateralisation index for each 3 participant. First, we saved the maximum pixel intensity in the left and the right hemisphere 4 (lower left and right quadrants of the interpolated image), excluding the midline. Then, we 5 computed the lateralisation index in each group as the ratio (MI left -MI right ) / (MI left + MI right ). 6

15
A reproducibility package with data and code will be available online as soon as possible. 16