Heightened sensitivity to low-level visual information in autism during an emotional 1 attentional blink task.

Impairments in facial emotion recognition have been a hallmark of autism, which may 49 contribute to the difficulty in social engagement and interpersonal interaction. Impaired facial 50 emotion recognition in autism could be partly due to the asymmetrical perceptual bias to High 51 Spatial Frequencies (HSF) information observed during visual perception. While Low Spatial 52 Frequencies (LSF) convey coarse information, which would be critical for a fast analysis and 53 categorization of emotional faces, HSF convey local information, which may serve a critical 54 role in visual consciousness. However, to our knowledge, the effect of HSF on visual 55 consciousness in autism has not been specifically studied so far. compare Our sample was adult, high functioning and mainly late diagnosed. Therefore, our findings may not generalize to the whole autistic population. Results confirm that HSF plays a critical role in visual consciousness in both TD and autistic 75 participants. More importantly, autistic participants demonstrated impaired target detections 76 after filtered distractors, suggesting that they have enhanced sensitivity for low-level 77 characteristics, such as high and low spatial frequencies filtering. These findings are discussed 78 in the context of the Enhanced Perceptual Functioning theory and predictive coding 79 frameworks. The present study shed light on the influence of spatial frequencies on visual consciousness in both autistic and TD persons. The results of the study extend our understanding of the attentional blink phenomenon. Moreover, our findings suggest increased sensitivity to low- level visual characteristics in autism. If HSF sensitivity was expected in autism, LSF sensitivity was more surprising. These results can be explained by the Enhanced Perceptual Functioning framework, which is an influential model of autism, and by superior visual abilities in autism. They can be also explained by predictive coding frameworks, which are emerging models for explaining autism characteristics. However, further studies are required to tease apart low-level processing from predictive coding processing in autism. Results of the current research could have important implications for our understanding of atypical visual processing in autism, which can partly contribute to social deficits.

(non-meaningful scrambled faces) in their experiment enhanced salience of targets among 165 distractors. In other words, they supposed that low similarity between targets and distractors 166 facilitated target detection by intensifying the emotional valence of faces, thereby enhancing 167 performances in autistic participants and reducing differences between the groups on emotional 168 stimuli. On the contrary, high similarity between targets and distractors can account for an 169 increased AB as shown in TD [21,30]. 170

171
The goal of the current research is to examine visual consciousness of happy faces in ASD and 172 TD participants using the AB task. In order to determine which spatial frequency channels 173 influence preferentially visual consciousness in ASD as compared to TD, we used stimuli in 174 different spatial frequency ranges. In our AB task, participants had to detect two unfiltered 175 happy faces (T1 and T2) in a stream of angry faces distractors, which were either low-pass 176 filtered (i.e., LSF), high-pass filtered (i.e., HSF) or unfiltered. We also capitalized on the finding 177 that spatial-frequency filtered distractors would allow us to manipulate similarity between 178 targets and distractors. In this experiment, we tested three major hypotheses. First, we predicted 179 a strong AB effect in both groups, as distractors have a high degree of similarity with targets 180 because they are all faces. Secondly, given a pivotal role that HSF plays in visual consciousness 181 during face identification [9], we expected a stronger AB after HSF and unfiltered distractors 182 (as it also contains HSF) compared to LSF distractors in both groups. Lastly, due to the 183 perceptual bias favoring fine and detailed information observed in ASD, autistic participants 184 would be significantly more disturbed by HSF distractors than TD participants. 185 Similar to other AB studies, we controlled for T1 accuracy. Although previous research did not 186 find a difference between individuals with ASD and TD in T1 accuracy when stimuli were 187 worlds or animals [28,29] we expected reduced T1 accuracy in the ASD group because of 188 impaired emotional face recognition usually found in individuals with ASD. Thus, we predicted 189 reduced T1 accuracy in ASD as compared to TD. We also assessed the Full Scale Intelligence 190 Quotient (FSIQ), verbal Intelligence Quotient (VIQ), Performance Intelligence Quotient (PIQ), 191 Autism spectrum Quotient (AQ) and age, to control for these variables [24,28]. control adults (19 females, 15 males, and 1 transgender female-to-male) were recruited for this 203 study. One additional autistic man was enrolled but excluded as the task was too difficult for 204 him and skipped all answers. All participants were aged from 19 to 47 years and had a FSIQ > 205 70, as estimated using a Wechsler Intelligence Scale ( [39]. Their age at diagnosis was situated between 10 and 45 years old (mean = 220 28.8, SD = 10.3). Three of them also received a diagnosis of attention deficit with or without 221 hyperactivity disorder (one of them was treated with methylphenidate); two received a 222 diagnosis of anxiety disorder and two had a history of traumatic brain injury. Two were 223 receiving neuroleptic medication and six were used to take antidepressant when they were 224 enrolled in the study but were stabilized with the treatment. Twenty-one autistic participants 225 had neither other diagnosis nor treatment. 226 TD adults were recruited via advertisements and mailing list. They didn't have any 227 neurological, neurodevelopmental or psychiatric diagnosis. As far as possible, groups were 228 paired on sex, age and education. An estimation of their FSIQ, VIQ, and PIQ was performed. 229 After having checked assumptions (variance homogeneity with Levene's test and normality 230 with histogram, density curves and qq-plot visual inspection, as well as the Shapiro-Wilk test), 231 we performed two-sample t-test, as t-test is robust to small violation of normality. Groups did 232 not differ on age, education, VIQ, and FSIQ. Nevertheless, it is worth noting that differences 233 between groups on FSIQ is approaching the significance threshold and may be related to the 234 significant difference between groups on PIQ. As predicted, groups differed on the AQ. All 235 relevant statistics regarding groups description are set out in Table 1.   participant in order to ensure that instructions had been correctly understood and to reassure the 284 participant if the task's difficulty generated anxiety. After this training session, the experimenter 285 left the room. Each trial began with a fixation cross for approximately 1,000 milliseconds (ms), 286 followed by a sequence of four, five, six or seven unfiltered angry faces followed by T1 (an 287 unfiltered happy face). The second unfiltered happy face T2 appeared after a second serial of 288 two (Lag 3), four (Lag 5) or six (Lag 7) distractors, which were either in BSF, LSF, or HSF. 289 Lag 0 corresponds to the onset of T1. Each stimulus had a duration of 133 ms. Accordingly, 290 stimulus onset asynchrony was either 399 ms (Lag 3), 665 ms (Lag 5), or 931 ms (Lag 7). As 291 emotional faces are complex stimuli, longer presentation times have been required, compared 292 to experiments using letters, in order to reduce task difficulty. For each trial, distractors were 293 similarly filtered and five unfiltered distractors remained after T2. After this rapid serial visual 294 presentation, a screen appeared for approximately 4,000 ms with three happy faces on the upper 295 half. Right-handed participants had to choose which of these three faces corresponded to T1 by 296 pressing one of the three corresponding buttons on the right side of the Chronos® device 297 (Psychology Software Tools). Importantly, participants were instructed not to answer randomly 298 and had to press the far-left button if they had no idea about the answer. For left-handed 299 participants, the buttons were situated on the left side to facilitate the motor response. Then, a 300 second screen appeared for the same duration with three unfiltered happy faces displayed on 301 the lower half; this was made to ensure participants would notice the change in case they were 302 distracted during this latter. Participants were instructed to find T2 among the three faces. The 303 schematic of one trial is shown in Figure 2. 304 [Insert Figure 2] 305 306 307 308 Figure 2 Example of a trial with HSF as distractors. Target 1 has to be selected among other faces on the Answer 1 slide and Target 2 has to be selected among other faces on the Answer 2 slide.

310
All figures and analyses were done using R version 3.6.1 and R Studio version 1.2.5019 [42].  We calculated the correct response rate (accuracy) for T2 given that T1 was correctly reported 322 [19,28,29] for each of the experimental conditions. This latter is abbreviated "T2 accuracy" in 323 the following text. Figure 3 represents the T2 accuracy for each group (ASD, TD) as a function 324 of Lag (3, 5, and 7) and Distractor (BSF, HSF, and LSF). All means and standard errors for T2 325 accuracy are reported in Table 2. We conducted a mixed-design ANOVA with Group (ASD vs 326 TD participants) as a between subject factor, and Lag (3, 5 or 7) and Distractor (BSF, LSF or 327 HSF) as within-subject factors. Assumptions of sphericity were tested with Mauchly Test. As 328 there was no violation of sphericity, no correction was required. Normality assumption on 329 residuals were visually inspected with histogram and qq-plot, and tested with a Shapiro-Wilk 330 test. 331 The analysis revealed three significant main effects. The main effect of Group (F (1, 65) = 4.14, 332 p < .05, η 2 = .04), indicated that overall T2 accuracy of ASD participants was significantly 333 lower than overall T2 accuracy of TD participants. A main effect of Lag (F (1, 85) = 104.8, 334 p < .001, η 2 = .17) was also found. Post-hoc paired t-test with Tukey adjustment showed that 335 T2 accuracy at Lag 3 was worse than T2 accuracy at Lag 5 (t (130) = − 8.27, p <.001, 336 d = − 0.65), which was also worse than T2 accuracy at Lag 7 (t (130) = − 6.15, p <.001, 337 d = − 0.43). These findings demonstrate the expected AB phenomenon in our paradigm (both 338 group taken together). Finally, we observed a main effect of Distractor (F (1, 88) = 17.68, 339 p < .001, η 2 = .02). Post-hoc paired t-test with Tukey adjustment revealed that T2 accuracy after 340 HSF distractors was lower than after BSF distractors (t (130) = 3.61, p <.005, d = -0.21), and 341 lower than after LSF distractors (t (130) = -5.90, p <.001, d = -0.32). These findings revealed 342 that HSF produced a greater AB than other distractors when groups were taken together. 343 However, the difference between T2 accuracy after BSF distractors and after LSF distractors 344 was not significant. 345 There was also a significant Lag × Distractor interaction (F (3, 85) = 2.96, p < .05, η 2 = .007), 346 attesting that the effects of distractors were different depending on the Lag. Paired t-test with 347 Tukey adjustment revealed that T2 accuracy was lower for HSF than LSF at Lag 5 348 was not observed at Lag 3, suggesting that the greater AB followed by HSF compared to LSF 350 appeared only at longer lags (5 and 7). Moreover, T2 accuracy was significantly lower for BSF 351 than LSF at Lag 5 (t (390) = − 2.79, p < .02, d = − 0.29), but not at Lag 3, and only marginally 352 significant at Lag 7 (t (390) = − 2.29, p = .058, d = − 0.24). Hence, BSF distractors also produced 353 a greater AB than LSF distractors at longer lags. 354 The data set out in Figure 3 suggests that the lack of significance difference between BSF and 355 LSF at Lag 7, which was not expected, could be explained by a different pattern in each group. 356 This led us to perform within-group paired t-test with Tukey adjustment although the foregoing 357 analysis provided no interaction involving the group factor. In the TD group, a significant 358 difference in T2 accuracy exists between HSF and LSF at between BSF and LSF, neither at Lag 5, nor at Lag 7. Hence, in the ASD group, only HSF 366 distractors produced a greater AB than LSF distractors at longer lags (5 and 7). These 367 differences between groups could therefore explain the lack of significance previously found at 368 Lag 7 between BSF and LSF in the main effect and in the interaction. Moreover, the analysis 369 showed a difference in T2 accuracy between BSF and HSF at Lag 7 (t (390) = 2.39, p < .05, 370 d = 0.30) for the ASD group, suggesting greater AB after HSF distractors compared to BSF 371 distractors at the longest lag. 372 Finally, we tested our main hypothesis: HSF distractors would produce a greater AB than LSF 373 distractors for the ASD group as compared to the control group. We performed between-group 374 paired t-tests with Tukey adjustment for each distractor type and within each Lag. The expected 375 between-group difference on T2 accuracy after HSF distractors was only found at 376 Lag 7 (t (161) = − 2.21, p < .03, d = − 0.46). Differences between groups on T2 accuracy were 377 also found after LSF distractors at Lag 5 (t (161) = − 2.31, p < .03, d = − 0.55) and at 378 Lag 7 (t (161) = − 3.00, p < .005, d = − 0.73). These unexpected findings indicate that LSF 379 distractors also produce a greater AB for the ASD group as compared to the TD group at longer 380 lags (5 and 7). No between-group differences in T2 accuracy were found after BSF distractors. 381 These results are represented on Figure 3. We calculated a correct response rate for T1 for each participant. The latter is abbreviated by 390 "T1 accuracy" in the text. We did not find any significant correlation between T2 accuracy and AQ, FSIQ, nor VIQ. A 399 significant negative correlation between age and T2 accuracy was found for the TD group only 400 (r = − .43, p < .01), indicating that the older the TD participants are, the stronger is the AB. We 401 also found a significant negative correlation between age and T1 accuracy in the TD group 402 (r = − .35, p < .05), revealing that the older TD participants are, the lower is their T1 accuracy. 403 Finally, we found a significant positive correlation between T1 accuracy and FSIQ in the TD 404 group (r = .41, p < .05) and in the ASD group (r = .39, p < .05), indicating that the higher FSIQ 405 is, the better is the T1 accuracy. We also found significant positive correlations between T1 406 accuracy and PIQ in the TD group (r = .46, p < .01) and in the ASD group (r = .40, p < .05), 407 revealing that the higher the PIQ is, the better is the T1 accuracy. 408 [insert Figure 5] 409 The current study used an AB paradigm with spatial-frequency filtered distractors to investigate 418 the influence of spatial frequencies on visual consciousness in autistic adults compared to TD 419 control participants. 420

421
Replicating previous results on AB in TD people [19,20] and in ASD [26][27][28][29], our results 422 elicited a strong AB, as T2 accuracy was reduced for shorter lags and increased with longer 423 lags. This finding is consistent with the first hypothesis. We also found a stronger AB after HSF 424 distractors and after BSF distractors as compared to LSF distractors, at Lag 5 and at Lag 7 in 425 both groups taken together, which is consistent with our second hypothesis. This result further 426 confirms that HSF plays a primary role in visual consciousness. As HSF are mainly conveyed 427 by the ventral stream, our findings supports the idea that visual consciousness emerged from 428 the temporal pathway processing, as suggested by other studies [43][44][45]. For instance, a 429 correlation has been shown between event-related potential in the occipito-temporal cortex 430 (N170) and conscious perception of faces [43]. It should be noted that the difference between 431 BSF and LSF at Lag 7 was only marginally significant. This could be explained by differences 432 between groups on LSF distractors, which will be further discussed. No difference between 433 distractors at Lag 3 can be explained by the strength of the AB at that lag. 434

Increased sensitivity to filtered distractors suggests enhanced low-level visual
435 processing in autism 436 Based on previous findings of perceptual bias favoring HSF information in autism [14][15][16] we 437 predicted that autistic participants would show significantly reduced T2 accuracy in response 438 to the HSF content of inter-target distractors compared to TD control participants. While there 439 was no difference between groups on T2 accuracy according to distractor types at Lag 3, 440 probably due to the strong AB effect at lag 3 for both groups, group differences emerged at 441 Lags 5 and 7. Of particular importance is lower T2 accuracy for autistic participants, compared 442 to TD, after HSF distractors at Lag 7, which partially support our hypothesis. Interestingly, at 443 Lags 5 and 7, our analysis also revealed unexpected lower T2 accuracy for the ASD group after 444 LSF distractors as compared to the TD group. Additionally, we did not observed differences 445 between groups on BSF. The latter finding indicate that the main effect of group observed on 446 T2 accuracy in our experiment appears to be related to the filtered nature of the distractors. 447 Thus, it is not likely to be due to the emotional content, which might have reduced the AB in 448 TD participants compared to ASD as it has been observed sometimes [28]. Taken together, 449 these results suggest an increased sensitivity to low-level visual characteristics (i.e., HSF and 450 LSF information) in ASD participants compared to TD control. No difference between groups 451 after BSF could be questionable as BSF contain LSF and HSF. However, BSF also contain 452 other spatial frequency bands. Hence, processing LSF and HSF in BSF stimuli might differ 453 from processing them independently in low-passed or high-passed stimuli, as suggested by 454 other experiments [4]. The sensitivity to low-level visual characteristics in autism found here 455 supports the Enhance Perceptual Functionning theory of autism [14]. According to this theory, 456 the local bias observed in autism should be attributed to enhanced low-level perception, which 457 could explain, in our experiment, the sensitivity to both HSF and LSF, rather than to deficits in 458 processing global information, as suggested by the Weak Central Coherence theory [46]. 459 Indeed, it has already been showed that autistic children are perfectly able to perform global 460 processing in some conditions, depending on the task demands [47]. Our findings are also 461 congruent with the enhanced visual acuity found in autism [12]. Enhanced low-level 462 characteristics processing in autism could partially explain some inconsistencies found in 463 studies implicating spatial frequency filtering, as some found that autistic participants were 464 more sensitive to HSF [17,48] and other to LSF [49]. At a neural level, enhanced low-level 465 visual perception in autism could be related to superior functioning of the most posterior region 466 of the visual brain [14]. At a behavioral level, it could also be related to atypical attention pattern 467 usually reported in autism for social [50] but also non-social stimuli [51]. As eye-tracking 468 experiments demonstrate atypical pattern of attention orientation in autistic participants when 469 processing visual scenes, we can hypothesize that their gaze might be attracted by low-level 470 characteristics of the scene. 471

472
The sensitivity of autistic participants to filtered distractors could also provide evidence to 473 support predictive coding theories. Predictive coding theoretical frameworks hypothesize 474 hypoactive top-down modulations or hyperactive low-level bottom-up information processing 475 in autistic individuals would result in increased prediction errors. Higher prediction errors could 476 be used as feed-forward inputs for the next stage of processing, whereas those errors should 477 normally be minimized to allow adaptation to an environment [52-55]. In our experiment, the 478 second target appeared after a stream of distractors containing the same low-level 479 characteristics (i.e., the same type of filtering was used for each distractor). It is reasonable to 480 assume that ASD participants became used to the characteristics of the distractors (at Lag 5 and 481 7) and made strong predictions for the next image, based on the current input. Hence, we assume 482 that when distractors were low-pass or high-pass filtered, the appearance of the unfiltered 483 second target T2 violated the expectation generated by the statistical regularity of the stream, 484 leading to greater prediction errors. This mechanism could impair T2 accuracy in autistic 485 participants more than TD as the latter seem more likely to adapt to this change. In the case of 486 BSF distractors, targets and distractors had the same low-level content. Thus, it is possible that 487 autistic people did not make prediction error related to low-level characteristics of the stimuli 488 (i.e., spatial frequencies filtering), explaining the absence of between-group differences for BSF 489 distractors. The predictive coding hypothesis for our results is supported by 490 electrophysiological studies on Mismatch Negativity in autistic adults. The Mismatch 491 Negativity paradigm is used to study brain reactions to an unexpected event occurring among 492 regular events and can be seen as an electrophysiological signal of perceptual prediction errors 493 to other experiments (where T2 accuracy range from approximately 60 % to 90 %), even if they 500 implied autistic participants [28,29]. It could be surprising as an attenuation of the AB (i.e., 501 better T2 accuracy) is usually observed when targets are emotional faces [22,23]. However, the 502 similarity between targets and distractors, which were all emotional faces in our experiment, 503 can explain this result. It confirms past findings that target-distractor similarity would reduce 504 salience of targets among distractors thus increasing the AB [21,29,59]. In addition, the answer 505 choice in our task was presumably harder as compared to other similar tasks and may affect 506 performances. Indeed, our participants have to identify the T1 face between three happy faces 507 (and the same for T2), with the instruction not to answer by chance, having the possibility to 508 choose "I don't know" (thus, the chance level here is much less than 33%). We assume that 509 finding a target face among other is more difficult than a dichotomous yes/no answer on a 510 question such as "Did you see a face?" (chance level = 50 %). 511 The task difficulty could also explain the longer AB observed in our experiment. Despite AB 512 usually occur between 200 ms and 500 ms after T1 onset, it appears to still occur at Lag 5 513 (665 ms) and maybe even at Lag 7 (931 ms). Nevertheless, longer lags would be required to 514 determine the actual duration of the AB. This can extend previous findings showing that the 515 AB magnitude can be modulated by various factors such as task demands [60,61]. 516

517
The task difficulty probably impaired T1 recognition as well. Indeed, we observed apparent 518 poorer T1 accuracy in our experiment with greater variability as compared to T1 accuracy 519 usually observed in other experiments. 520 We predicted that T1 accuracy would differ between groups, because the task involved 521 emotional faces which are difficult to process by autistic people [2]. However, T1 accuracy was 522 similar in both groups. A meta-analysis on emotions recognition in autism suggests the intact 523 recognition of happiness in autistic individuals [2]. Since our targets were happy faces, autistic 524 participants might have no difficulty in identifying them, which could explain similar T1 525 accuracy in both groups. Alternatively, this result could be related to IQ performances. A 526 correlation exists between T1 accuracy and FSIQ, as well as between T1 and PIQ for both 527 groups, which is in line with previous studies showing that higher IQ can enhance emotion 528 recognition in both TD and ASD participants [62,63]. Thus it is possible that higher IQ of ASD 529 participants compared to the TD control group (statistically significant difference for PIQ, 530 marginally significant difference for FSIQ), can possibly compensate emotion recognition 531 impairment usually found in ASD and mask between-groups differences. Nevertheless, when 532 we remove the few outliers in term of IQ performances (group means ± 2SD), correlations 533 between T1 accuracy and FSIQ and PIQ were not significant anymore. This is in favor of weak 534 FSIQ and PIQ weights in the task. It should be mentioned that correlation between IQ and T2 535 accuracy is not significant, which is coherent as the calculation of T2 accuracy is based on trials 536 where T1 have been correctly reported. It is also congruent with previous research [25]. 537 Moreover, we didn't find any correlation between VIQ and T2 accuracy in ASD (nor in TD), 538 contrary to another study [28]. It may be explained by differences in our paradigm: their targets 539 and distractors were emotional and non-emotional words, which could more imply VIQ than 540 our task based on emotional faces. Taken together, these findings support the fact that IQ 541 variations among our subjects have not impaired our results. Finally, we found a significant 542 negative correlation between accuracy (for T1 and for T2) and age in TD, performances 543 becoming weaker with age, but this is not found in ASD. AB performances tend to improve 544 between 18 and 39 years old, and then to decline [24]. Hence, the negative correlation could be 545 at odds with the fact that only 9 TD participants are aged 39 and above. However, a fast decrease 546 of performances after 39 might explain this result. Surprisingly, this decline is not observed in 547 ASD participants despite they are a bit older than TD participants in our study ( representative of the ASD population. Hence, the results found here cannot be generalized to 563 the whole spectrum. It would be interesting to conduct similar task with autistic participants 564 with IQ in the normal range, or with low IQ or with participants with typical autism, but the 565 task should be adapted because of its difficulty. Moreover, there is a high variability in 566 performances of autistic participants, probably reflecting the highly heterogeneous autism 567 spectrum. Data driven analysis might be conducted to distinguish autistic subgroups [67], but 568 would probably require larger sample size. Another limitation is that a small portion of our 569 autistic participants has comorbidities and/or treatment, which might have affected their 570 performances. Nevertheless, in this case, we can hypothesize that all performances would have 571 been affected, and not only those after LSF and HSF distractors. To confirm or infirm this 572 hypothesis, similar studies should be conducted with other clinical populations. Finally, the 573 correlation found with age could potentially have a slight impact on the result and should 574 conduct to lower the maximal age in further studies. 575 Other limitations are related to the task. Our task was designed to detect happy faces target 576 among angry faces distractors. Hence, it cannot be generalized to other emotional faces. 577 Moreover, the task was difficult and we are aware that accuracy is quite low; nevertheless, as 578 participants were instructed not to answer by chance and to use the "I don't know" button when 579 they were not sure, we think that our results are reliable. 580

582
The present study shed light on the influence of spatial frequencies on visual consciousness in 583 both autistic and TD persons. The results of the study extend our understanding of the 584 attentional blink phenomenon. Moreover, our findings suggest increased sensitivity to low-585 level visual characteristics in autism. If HSF sensitivity was expected in autism, LSF sensitivity 586 was more surprising. These results can be explained by the Enhanced Perceptual Functioning 587 framework, which is an influential model of autism, and by superior visual abilities in autism. 588 They can be also explained by predictive coding frameworks, which are emerging models for 589 explaining autism characteristics. However, further studies are required to tease apart low-level 590 processing from predictive coding processing in autism. Results of the current research could 591 have important implications for our understanding of atypical visual processing in autism, 592 which can partly contribute to social deficits.

Availability of data and materials 624
The datasets generated and/or analyzed during the current study are available in the Open 625 Science Framework repository, https://osf.io/maqvz/. Any other materials are available from 626 the corresponding author on reasonable request. 627