The geometric preference subtype in ASD: identifying a consistent, early-emerging phenomenon through eye tracking

Background The wide range of ability and disability in ASD creates a need for tools that parse the phenotypic heterogeneity into meaningful subtypes. Using eye tracking, our past studies revealed that when presented with social and geometric images, a subset of ASD toddlers preferred viewing geometric images, and these toddlers also had greater symptom severity than ASD toddlers with greater social attention. This study tests whether this “GeoPref test” effect would generalize across different social stimuli. Methods Two hundred and twenty-seven toddlers (76 ASD) watched a 90-s video, the Complex Social GeoPref test, of dynamic geometric images paired with social images of children interacting and moving. Proportion of visual fixation time and number of saccades per second to both images were calculated. To allow for cross-paradigm comparisons, a subset of 126 toddlers also participated in the original GeoPref test. Measures of cognitive and social functioning (MSEL, ADOS, VABS) were collected and related to eye tracking data. To examine utility as a diagnostic indicator to detect ASD toddlers, validation statistics (e.g., sensitivity, specificity, ROC, AUC) were calculated for the Complex Social GeoPref test alone and when combined with the original GeoPref test. Results ASD toddlers spent a significantly greater amount of time viewing geometric images than any other diagnostic group. Fixation patterns from ASD toddlers who participated in both tests revealed a significant correlation, supporting the idea that these tests identify a phenotypically meaningful ASD subgroup. Combined use of both original and Complex Social GeoPref tests identified a subgroup of about 1 in 3 ASD toddlers from the “GeoPref” subtype (sensitivity 35%, specificity 94%, AUC 0.75.) Replicating our previous studies, more time looking at geometric images was associated with significantly greater ADOS symptom severity. Conclusions Regardless of the complexity of the social images used (low in the original GeoPref test vs high in the new Complex Social GeoPref test), eye tracking of toddlers can accurately identify a specific ASD “GeoPref” subtype with elevated symptom severity. The GeoPref tests are predictive of ASD at the individual subject level and thus potentially useful for various clinical applications (e.g., early identification, prognosis, or development of subtype-specific treatments). Electronic supplementary material The online version of this article (10.1186/s13229-018-0202-z) contains supplementary material, which is available to authorized users.

three positions (top left, top right, and center) immediately after testing to determine whether excessive loss of spatial accuracy has occurred during the testing process. Though we do not have this information for the current dataset, across 240 toddler testing sessions, we have found that 97% of toddlers test within this margin. For adults we routinely obtain by calibration, and maintain during eye tracking, accuracy with error below 1 degree, and this is verified periodically as part of standard lab practices to assure data quality. Further details of the calibration and eye tracking procedures used were largely the same as those described in our prior paper's supplement [38].

Figure S2
below shows data "heat maps" drawn in Tobii Studio of convenience samples for illustrative purposes. The point of gaze fixation is usually near the AOI's center, so the outer perimeter of the AOI constitutes a large margin of error against spatial accuracy that falls outside the accuracy measurement the initial calibration would predict (e.g. due to drift).
A. depicts data from n=11 typically developing children B. depicts data from n=21 children with ASD The means and standard deviations for total looking time (in seconds, out of a possible total of 90 seconds) to either AOI by diagnostic group were as follows: ASD 65.6 (17.2); ASD Features 79.2 (11.5); DD 76.4 (15.1); TD 78.3 (13.0); Other 77.7 (15.9); TypSib 72.9 (21.6).
There was a significant main effect (F5,221 =5.8, p<.001) of diagnosis, and significant after Bonferroni correction post-hoc pairwise differences between ASD and the groups DD (p<.005), TD (p<.001) and Other (p<.05). There were no significant differences found between groups in total looking time to the original GeoPref test [38]. The difference found here can likely be attributed to the total duration of the Complex Social test video being 90 seconds, while the original GeoPref test's duration is 60 seconds.

Saccade Frequency to Geometric vs Complex Social Stimuli
Differences in saccades between clinical groups were reported for the original GeoPref test in our 2016 paper [38], therefore we report whether or not these effects were replicated with more complex social stimuli in the current study. For both the geometric and social stimuli, number of fixations per AOI was divided by sum of fixation time for that AOI to derive saccade frequency as saccades per second. Homogeneity of variance was confirmed then 1-way ANOVAs were performed (diagnostic group (6 levels) X saccades/sec (1 level)) for each AOI, and significant effects were followed by pairwise comparisons with Bonferroni correction. To confirm that differences in data quality were not impacting the reported results, ANCOVAs were performed as well, with 6 diagnostic groups as a fixed factor, saccade/sec as the dependent variable, and a data quality measure (percent of valid samples obtained) as a covariate.
However, in this case, we found significant effects for the percent samples data quality measure. This implies that variable data quality between groups may be confounding true measurement of saccade rate, so results must be interpreted with caution.

Saccade Frequency of the 6 Groups to Geometric vs Complex Social Stimuli
For each toddler, each stimulus type (geometric and complex social) was considered separately to calculate saccades/sec. Saccade frequency data were subjected to additional scrutiny as this data quality can be impacted by data loss and lack of precision, while overall percent total fixation duration calculations are not as sensitive. Therefore, we excluded five subjects with saccades/sec values greater than two interquartile ranges from the upper quartile, indicating poor fixation filter performance [49]. While looking at geometric stimuli, there were no statistically significant differences in saccades/sec among the six diagnostic groups (F5,216=.43, p=.8). However, ASD toddlers had significantly more saccades/sec when viewing social stimuli than did TD or DD toddlers (F5,216=3.4, p=.005, partial eta-squared=.07; ASD vs DD, p<.005, Cohen's d=.63; ASD vs TD, p<.05, Cohen's d=.57). This is consistent with a faster rate of saccades occurring when the child is less interested in or attentive to the stimuli. But this could also reflect less ability to correctly measure saccades from some ASD children (also arguably due to being less interested or attentive).

Saccade Frequency Comparisons within ASD Subgroups
For geometric stimuli, the ASD GeoPref subtype had less frequent saccades, 1.33 saccades/sec, while the ASD SocPref subtype had more frequent saccades, 1.87 saccades/sec, (t25.8=2.14, p<.05). For comparison, saccade rates to the same geometric stimuli reported in Pierce et al (2016) for these ASD subgroups were for ASD GeoPref 1.33 saccades/sec, and for ASD SocPref 1.94 saccades/sec. When viewing the new complex social stimuli, the opposite occurred: the ASD GeoPref group had more frequent saccades, 2.38 saccades/sec, and the ASD SocPref group has less frequent saccades, 1.61 saccades/sec, (t18.4=-3.77, p=.001). See Figure S3. This is again consistent with faster rate of saccades when less attentive to the stimulus, but also potentially impacted by difficultly precisely measuring saccade frequency during less attentive behavior.

Figure S3
Saccade Frequencies of ASD Subgroups to Complex Social and Geometric Images Figure S1. Bar graph illustrating comparison of saccade rate (saccades/second) between two ASD subgroups defined by their % Geo scores in the Complex Social GeoPref test: those who viewed geometric images more than 69% of the time, the Geometric Responder (GeoPref) subtype, and those who viewed social images more than 69% of the time, the Social Responder (SocPref) subtype. Group sizes were n=13 geometric responders and n=16 social responders. Error bars represent 95% confidence intervals. Table S1 shows classification validation statistics based on the cutoff for maximizing specificity with the Complex Social GeoPref test, 75% Geo looking time. At this cutoff, where specificity is 99% (more accurately, it is 99.78%), the PPV of the Complex Social GeoPref test slightly exceeds that of the original GeoPref test, at 92%. That is, on the original GeoPref test, at the 69% cutoff for %Geo the specificity is 99%, however the PPV is 90%. PPV is the likelihood that a given positive test is a true positive, and is therefore of particularly strong interest to clinicians.  Table S2 shows an additional potential usage for the Complex Social GeoPref test, ruling out an ASD diagnosis. That is, children with a %Geo score below 11% (i.e. a %Soc score above 89%) are very unlikely to have an ASD, with test classification performance shown below, where a lower score is more positive. This can be particularly useful as a second tier screen applied to toddlers who have already shown a few potentially concerning behaviors, and who may or may not have an urgent need for an autism focused evaluation. These values were derived by defining positive as falling into any diagnostic group other than ASD, and true positive by also have a %Geo score below 11%; negative refers to having an ASD diagnosis, and true negative to having an ASD and a %Geo score of 11% or greater.   Table S3 shows validation classification statistics for detecting children with ASD without the inclusion of typically developing children (i.e. TD or TypSib groups), as distinguishing between ASD children and those with some sort of delay or other challenge that impacts behavior is typically the task facing clinicians. Also, our overall TD group contains some "control" participants, who would not be present in a sample from a non-research clinical setting, so this may be more reflective of a natural clinical sample. Additional clinical severity in relation to eye tracking data information Figure S3 (below) shows the significant correlation between %Geo scores on the Complex Social GeoPref test and ADOS scores for the entire group of 76 ASD study participants (Pearson's r=.46, p<.001).

Experiment 2: Exploratory data on back to back usage of the original and Complex Social GeoPref tests
An additional 162 subjects participated in both GeoPref tests administered back to back on the same day. Between tests a break of several seconds was provided and a brief presentation of fixation crosses at known X-Y coordinates was used to confirm continued accuracy of gaze measurement.
Sixty-six of these participants (41%) were excluded from analysis due primarily to behavior incompatible with data collection (excessive movement, tantrums, etc) when exposed to the combined 2.5 minutes of GeoPref video presentation across two tests. In comparison, incorrect. This value, combined with the large percentage of subjects whose data cannot be used (41%), limits the clinical utility of administering the two GeoPref tests immediately back to back. In the future, the addition of musical sounds to increase willing attentiveness during back to back presentation of two GeoPref videos might be tested.