Unveiling Distinctive Eye-Tracking Markers to Distinguish Toddlers with High-Risk Autism Spectrum Disorder within an Elevated Risk Toddler Population

doi:10.21203/rs.3.rs-3450935/v1

Download PDF

Research Article

Unveiling Distinctive Eye-Tracking Markers to Distinguish Toddlers with High-Risk Autism Spectrum Disorder within an Elevated Risk Toddler Population

https://doi.org/10.21203/rs.3.rs-3450935/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Eye-tracking technology can be utilized to identify early markers of autism. Several atypical features of social attention has been revealed by eye-tracking studies in ASD toddlers. Our objective was to investigate novel eye-tracking variables aimed at distinguishing the ASD group from a population at elevated risk.

Methods

We recorded eye gaze data during the presentation of dynamic social attention stimuli with a self-developed eye-tracking device. We used two types of social attention stimuli: joint attention (ostensive) and preferential looking paradigm. In our research, we analyzed the data of 74 toddlers between 12–30 months (42 nonASD, mean age:22.5 ± 5,0 month; 32 with high-risk ASD (hrASD), mean age 23.3 ± 4,8 month). To examine group differences, we utilized both traditional and innovative eye movement variables, including those based on areas of interest (AOI), Markov variables and gaze retention interval variables (GRI).

Results

As anticipated, we observed significant differences between the two groups across most variables. Typically, within the preferential paradigm, the distinct indicators of the social domain were higher on average in the non-ASD group compared to the hrASD group, while the non-social domain exhibited the opposite trend. In the joint attention paradigm, group differences appeared to be associated with the prompting and ostensive aspects of interpersonal engagement. The hrASD group consistently displayed a lower level of social attention towards individuals, albeit with a smaller effect size compared to the preferential paradigm. The results of correlations with ADOS scores indicated that the social ratio in the preferential paradigm exhibited the strongest negative correlation. Notably, there were higher effect sizes within the hrASD group in comparison to the non-ASD group when correlation with ADOS scores was tested separately for each group.

Conclusions

We have developed variables (Markov and GRI) that are robust enough to use even in the noisy eye movement recordings typical of early childhood. In the preferential paradigm, besides the AOI ratio measures, the Markov and GRI variables show promise in distinguishing ASD toddlers. Furthermore, they exhibit potential for predicting the severity level and the likelihood of autism by correlating with the ADOS score. It is conceivable that, when utilized in conjunction, these variables capture characteristics indicative of early signs of autism.

autism spectrum disorders

toddler

eye tracking

social attention

Eye-tracking methods in research on autism spectrum disorder (ASD) have been popular in recent decades (1–3). Previous research has shown eye-tracking technology to be able to identify many of the ASD characteristics from behavior to neurocognitive functions (1). It is also suitable for infants and toddlers; however, employing eye-tracking-based research in early childhood has a number of challenges (e.g., attention capacity, need for movement and limited understanding of instructions) (4). Nevertheless, several repeated differences were found between the eye-tracking data of ASD and non-ASD toddlers (5–17), and as a result, the possibility of early detection was raised.

One of the most frequently researched areas in ASD is social attention (5). It has been established that people with autism have decreased attention to social stimuli. Furthermore, they have difficulties processing eye gaze information, for example, in a joint attention situation. Although these two aspects are not mutually exclusive, these paradigms are classically used together in research (6).

Decreased attention to social stimuli was demonstrated by applying preferential looking paradigm, in which two videos (or images) were played simultaneously side-by-side, one with socially relevant content and the other with non-social stimulus (e.g., object or geometric images) (7, 8). One of the most studied preferential looking eye-tracking paradigms is the GeoPref test. Pierce and her colleagues (9–11) proved that higher visual attention to geometric images compared to social images can be used as a biomarker for a subgroup of toddlers with autism. In their last research (11), they found highly replicable and reliable results. Their candidate biomarker is useful to recognize those toddlers with autism whose symptom severity is higher and language skill and cognitive ability are low.

The joint attention paradigm is a less uniformly applied procedure. In a typical example of the video stimuli, an adult would attempt to draw the attention of the toddler and then direct their gaze to an object. Various research focus on different aspects: responding (e.g., 12–14), initiation (e.g., 13,15) or the effect of labeling (16, 17). There are studies that concentrate on eye/mouth region preference. There is no clear difference in the amount of focusing on the eyes between ASD and typically developing (TD) toddlers (e.g., 16,17), and there are also ambiguous results about the mouth region (e.g., 17,18). Regarding responding behaviors, Chawarska, Macari and Shic (12) showed that ASD toddlers' attention decreases when the adult initiates joint attention in the video scene, but it is more significant if the adult shows stronger ostensive signs (child-directed speech, eye contact). According to their results, less attention to faces is context dependent. Others (e.g., 13,14) found no clear group differences in responding tasks (e.g., gaze following). In initiating measures, there are different patterns in ASD toddlers' eye-tracking data, mostly in the transition patterns (from object to face) (13, 15).

Our aim was to explore two types of social attention paradigms for some specific eye-tracking variables in toddlers with and without a high risk of ASD. We conducted our research using a custom eye-tracking device designed specifically for studying young children. A further objective was to develop variables that could mitigate the common issue of data loss during testing children under the age of 3; thus, we created more robust indicators. Our primary research question was whether these specific variables are suitable for distinguishing the ASD group from such an elevated risk population in the two prevalent social attention paradigms commonly used in ASD toddler research.

The objective of our study is to test innovative methodological approaches that could effectively distinguish toddlers with ASD from an elevated-risk population. Previous studies explored the potential of early screening for ASD with eye-tracking techniques (21, 10, 22). In various studies based on eye tracking of children with typical and atypical development, a high exclusion rate can be observed (23). A significant challenge lies in the difficulty young children under the age of 3 face in adhering to instructions during examinations, and they also tend to frequently move their head, which often results is data loss and inaccuracy. In the present research, we used a self-developed eye-tracking device that tolerates head movements and does not require calibration.

Moreover, we established variables that are sufficiently robust. In similar research investigating fixations, the subject exclusion rate was 11–26%, which was frequently due to a significant (40–50%) lack of data (11, 18, 24). In a meta-analysis, Von Holzen and Bergmann (25) reported dropout rates ranging from 0%-69% in mispronunciation detection studies with one-to-four-year olds using the preferential looking paradigm. They draw attention to the fact that it would be important to establish the length of the fixations theoretically and to compare only those results where a similar duration was used.

Instead of utilizing fixations, we opted for gaze retention interval (GRI) variables, which cover a broader spatial and temporal range. Additionally, we incorporated variables based on areas of interest (AOI) and variables that analyse the dynamic aspect of eye movements (Markov variables).

Participants

Participants were recruited through social media advertising and various early childhood institutions (e.g., primary care or diagnostic centers). Our recruitment process highlighted whether the suspicion of autism arose in the parents or professionals working with the family. As a result, most of the toddler’s parents had some suspicion or curiosity about autism. Thus, subjects were not screened from an average population but from an elevated risk sample. Our control group consisted of children for whom the suspicion of ASD was not confirmed by our examiners. The data collection took place from September 2022 to March 2023. Finally our research included 74 toddlers between 12–30 months. For the descriptive statistics of the subjects see the chapter “Study population” and Table 1.

Procedure

Parents registering for research filled out a questionnaire about their child’s anamnestic data (e.g., developmental problems) online. The child’s participation began with the eye-tracking procedure (max. 5 minutes), followed by ADOS-2 and ophthalmological screening (the two together max. 45–60 minutes). The parent completed the M-CHAT questionnaire during these examinations. Families received a 5000 HUF (approx. 14.2 USD) voucher upon their participation.

Clinical Measures

ADOS-2 Toddler module

The Autism Diagnostic Observation Schedule 2 (ADOS-2) is a semistructured, standardized assessment for the diagnosis of autism spectrum disorder. In our study, we utilized the Toddler module (26), which is similar to the original tasks of ADOS Module 1 but is intended for children under 30 months of age. The observation was conducted by our specialists who are certified in the method. To assess interrater reliability, 10 toddlers (M_age=23,5, SD = 4; 3 nonASD, 7 hrASD) were rated both by our specialists and a clinical psychologist with several years of experience. The interrater agreement between the observers on sample was 86–100% for algorithm items (M = 93,6%). There was no difference between the two specialists in the classification of the toddlers. Therefore, the following ADOS evaluations were performed by our own specialists.

We used two different algorithms for verbal and nonverbal children to calibrate the severity score. To determine the cut-off points, we used the results of Hong et al. (27). In Algorithm 1. the 0–9 severity score is “nonspectrum” (non-ASD), and 10 or over was classified in the “high-risk ASD” group (hrASD). In Algorithm 2. scores of 0–7 are “nonspectrum”, and scores of 8 or over were classified in the hrASD group.

Ophthalmological screening

Ophthalmological screening procedures are aimed at identifying prevalent ophthalmic issues in cases of atypical development. The screenings were administered by our examiners, who received training from a specialist in pediatric ophthalmology and strabismus. The ophthalmic screening tests applied included guided eye movement, cover and uncover test, and Lang-II test.

M-CHAT

The M-CHAT is a standardized questionnaire to be completed by parents, which helps to screen children aged 18–24 months for autism (28). As part of the research, the questionnaire was used in the entire 12–30-month age group due to its risk-free nature and content of relevant information.

Eye-tracking methods

Stimuli and presentation

Doing eye tracking research on toddlers has a number of challenges: attention capacity, need for movement and limited understanding of instructions. Therefore, two iterative stages preceded the present study (N₁ = 34, N₂ = 35), on the basis of which we selected the most suitable ones from several stimulus materials (types and items), and we also tested our eye tracking protocol on toddlers with typical development.

We decided on two major types of stimuli in the presented material: joint attention (subsequently ostensive) scenes and preferential stimuli (with social and non-social sides) (see Fig. 1 for examples).

In ostensive video, the actress first performed ostensive communication - looked at the camera, smiled and addressed the child (e.g., says ‘hello’) - and then directed the child’s attention to one of two similar objects (toys) by shifting her gaze (and head). The duration of these videos was between 9 and 13 seconds.

In preferential stimuli, two dynamic videos were presented side-by-side. There was always a social video with an adult (who shows ostensive communication) or child (who moves) or children's interactions (two or three children moving, dancing or playing together) and a non-social video with moving objects or nonfigurative scenes (they are all real video recordings). The social scene occurs half-and-half on the right or left side. The duration of these videos is between 8 and 13 seconds.

Through the examination videos followed each other in a quasirandom order: seven joint attention videos and 16 preferential videos.

During the presentation, the videos began with animation (zoom-out) and finished with animation and a sound signal (zoom-in and short beep). They were used to keep the children’s interest and direct the children's attention to the screen.

Eye tracker

For our measures, we used our custom-built eye tracker, which consists of a uEye UI-3860LE-M (SONY IMX-290 sensor) monochrome camera equipped with 16 mm optics and four 850 nm LED emitters spaced in a rectangular fashion approximately 5–5 cm inwards from the corners of the virtual stimuli presentation surface. The lighting and stimuli were combined by a selective wavelength semitransparent mirror. The infant subjects were seated in the lap of their caretaker spaced 50 cm from the front of the housing of the device, spaced 100 cm from the front lens of the optics and 110 cm from the LED emitters. The camera produced 50 frames per second at a resolution of 1936x1096 with a single channel. Note that the LEDs used were measured by an appropriate independent laboratory to emit unharmful levels of infrared light both in terms of corneal and retinal irradiation in the full range of the possible movement of subjects towards the emitters (min. 50 cm).

The eye tracker we built was specifically constructed to tolerate head movements present in infant behavior and to accommodate it using custom built software by our team. The eye tracker does not need any calibration to operate. The head movements are tolerated until the head and at least one eye remains within the field of view of the camera, which varies by the distance of the subject. The optics are calibrated in such a way that the eye tracker tolerates moving 30 cm towards the camera from the original position and 10 cm further from the camera and still maintains gaze capture. Note that backwards movement is more limited due to the subject sitting in the caretaker’s lap.

We measured the accuracy and precision of the eye tracking gaze-points on a sample of 11 adult persons with healthy vision. On a central target, average accuracy in X and Y directions was measured as 0.74° and 1.25°, while average precision was 0.37° and 0.53° respectively. The eye tracker produces raw gaze position data in a manner that is calculated from the two eye positions. In cases when only a single eye/gaze was visible/detectable, an algorithm from a previously learned relationship between the detectable eye and the binocular gaze point was applied to yield the gaze point.

Eye-tracking variables

The eye tracker generated a row-based table in which each row defined a gaze point in time for both eyes and a singular gaze point, also revealing the temporal state of the video stimulus that was presented at the time.

The stimuli were sectioned into temporal parts that represented similar situations across multiple similar stimuli. In the preferential paradigm, we used only the core part of the stimuli without the animations at the beginning and the end. In the ostensive stimuli, we used the postturn-towards-target parts of videos.

The areas of interest (AOI) were developed using the Voronoi method (29). In preferential stimuli, the social and non-social parts were defined in space and time, and in ostensive stimuli, there were a person, a target and a non-target AOI (see examples in Fig. 1.). Our AOI variables relied on the gaze-time ratio. We used five different AOI variables: ratio of social, ratio of non-social AOI; and in ostensive situations: ratio of person, ratio of target and ratio of non-target.

State change dynamics were captured by features that were Markovian Chain like (30), representing the changes of the gazed AOIs. Staying in the previous state was measured by a 5 second window as a unit (allowing fractional counts), while switching to another AOI was measured discretely by counting each represented switch as one. From this point for each person within a domain/item a Markov Chain was calculated and the probabilities of this Markov chain were used as the features. E.g. for a preferential item consisting of social and non-social AOIs, there were 4 Markov features: social to social, social to non-social, non-social to non-social, and non-social to social.

For detecting gaze retention intervals (GRI) a common algorithm was used, the identification by dispersion threshold (IDT) algorithm (31–34). With the conventional parameter settings, the algorithm detects fixations. However, with modified (more tolerant) parameters, it is able to detect the more dispersed, longer duration GRI. The IDT algorithm uses the x- and y-data of the eye positions and two fixed thresholds: the maximum fixation dispersion and the minimum fixation duration. To be a GRI, eye-movement data samples have to be within a spatial area not exceeding the dispersion threshold (300 pix) for at least the duration threshold (300 msec). The samples fulfilling these criteria were marked as GRIs. Based on GRI detections, we calculated density (number of GRI in a minute), duration (average length of GRI in msec), gaze time (duration of all GRI in msec), gaze stuck (duration of GRI in msec after the end of an AOI), GRI in AOI ratio (number of GRI in an AOI/number of all GRI during the AOI time) and latency (time in msec from the start of an AOI to the first GRI in the AOI) parameters.

For each subject, the values of the eye movement and gaze retention interval (GRI) variables were calculated in the preferential and ostensive situations in each AOI for each stimulus. Markov probability variables were also calculated. The values of variables from different stimuli but belonging to the same AOI were averaged. This resulted in robust variables where the effects of occasional missing data due to not watching the stimuli were mitigated.

Variables used in statistical tests

Simple AOI based eye-movement variables used in preferential situations: ratio of social (%) and ratio of non-social (%). AOI based eye-movement variables used in ostensive situations: ratio of person (%), ratio of target (%) and ratio of non-target (%). Markov variables used in preferential situations: staying in a social state, staying in a non-social state, transitioning from a non-social to a social state, and transitioning from a social to a non-social state. Markov variables used in ostensive situations: staying in person state, transitioning from person to non-target state, transitioning from person to target state. GRI variables used for social and non-social areas: density, duration, gaze time, gaze stuck, GRI in AOI ratio and latency.

Statistical method

To determine the optimal sample size, we performed power calculations for independent sample t-tests. A sample size of N ≥ 34 was found to be required for a power of at least 0.8 to detect a 0.7 effect at p < 0.05. Therefore, we decided that our sample size would be near N = 34 in the non-ASD and hrASD groups.

Non-ASD and hrASD groups were compared using independent sample t-tests (α = 0.05), and Hedge’s g was used to calculate the effect size of the differences between the groups. Pearson's point-moment correlation coefficient was computed between the ADOS score and the eye-movement parameters. Correlations were computed for all subjects and separately for hrASD and non-ASD subjects. The difference (expressed in Fisher's z score) between the correlations in the non-ASD and hrASD groups and the significance of the difference were also calculated. The Holm‒Bonferroni method of false detection rate correction was performed to correct for multiple comparisons. Statistical tests were carried out in G*Power 3.1.9.7 software (Franz Faul, Univedrsitat Kiel, Germany, 1992–2020) for power analysis. TIBCO STATISTICA 14.0.1.25 software (TIBCO Software Inc., 1984–2020) was used for independent sample t-tests and correlation calculations. Microsoft Excel Professional 2021 was used for Hedge’s g effect size and Holm‒Bonferroni calculations. An online tool (VassarStats: Website for Statistical Computation, Richard Lowry 2001–2023; http://vassarstats.net/rdiff.html) was used to compute the significance of the difference between two correlation coefficients using the Fisher r-to-z transformation.

Study population

Eighty parents of toddlers applied for our research call. From the final sample, we excluded 6 toddlers (7.5%). Subjects who had a visual acuity problem of 4 + dioptres (N = 1), who failed the eye movement tracking task (N = 2), or who could not be tested with ADOS (N = 3) were excluded from the experiment. It is important to emphasize that we excluded only those who could not be seated in front of the screen at all (not even on their mother's lap). The final sample of 74 toddlers (aged from 12 to 30 months; M = 22.8, SD = 4.9) met our inclusion criteria for the study.

We classified toddlers into two subgroups based on their ADOS scores. In the nonautism spectrum subgroup (non-ASD), we included those with ADOS scores below the cut-off points (N = 42). For a detailed description of the study population, see Table 1. Different algorithms (Algorithm 1 or Algorithm 2) were used to determine the ADOS score depending on the age and speaking ability of the subjects. The mean age was 20.8 months for Algorithm 1 (N = 42, hrASD ratio = 57.1%) and 25.5 months for Algorithm 2 (hrASD ratio = 25.0%).

Table 1

Descriptive statistics of the study population
		All					Males					Females
		N	Mean	SD	Min	Max	N	Mean	SD	Min	Max	N	Mean	SD	Min	Max
Age (months)	ALL	74	22,8	4,9	12	30	49	23,1	5,1	12	30	25	22,2	4,8	15	30
	Non-ASD	42	22,4	5,0	12	29	25	23,2	4,9	12	29	17	21,3	5,2	15	29
	hrASD	32	23,3	4,8	12	30	24	23,0	5,3	12	30	8	24,3	3,2	19	30
ADOS score	ALL	74	9,5	7,6	0	25	49	10,8	7,8	0	25	25	6,8	6,6	0	24
	Non-ASD	42	3,6	2,4	0	8	25	4,1	2,2	0	8	17	3,0	2,5	0	7
	hrASD	32	17,1	4,9	10	25	24	17,9	4,7	10	25	8	14,8	5,3	10	24

Statistical analysis

For each subject, the values of eye movement and gaze retention interval (GRI) variables were calculated for the total recordings, and then differences between the hrASD and non-ASD groups and Pearson’s correlations with the ADOS score, considered the gold standard, were examined. Correlations were also examined separately within the hrASD and non-ASD groups. No differences between groups were found for total recordings. No significant correlations were found, except for average length, where a weak correlation with ADOS score was observed for all subjects and for non-ASD subjects (see OSF Data Repository, OSFDR).

Differences between non-ASD and hrASD groups

Differences in eye-movement variables between the hrASD and non-ASD groups were tested using independent samples t-tests, and Hedge’s g effect sizes were also calculated (see Table 3). For the preferential situations, the mean of the hrASD subjects is significantly higher than the mean of the non-ASD subjects in the variable of non-social ratio and the Markov variables of staying in non-social state and transitioning from social to non-social state, while it is significantly lower in the variable of social ratio and Markov variable of staying in social state (Fig. 2.). For the ostensive situations, the mean of the hrASD subjects was lower than the mean of the non-ASD subjects in the variable of ratio of person and the Markov variable of staying in person state. Hedge’s g effect sizes were large for all significant differences in the preferential situations and moderate for all significant differences in the ostensive situations. For the preferential situations, all differences remained significant after correction for multiple comparisons; however, for the ostensive situations, the differences did not survive the stringent correction for multiple comparisons (see OSFDR).

Table 2

Independent samples t-test results and effect sizes of eye-movement variables in preferential and ostensive situations
		Mean hrASD	Mean Non-ASD	p	Sig.	Hedges's g effect size
Preferential	Ratio of non-social (%)	28,8%	17,3%	0,0000	*******	1,11
	Ratio of social (%)	51,9%	67,4%	0,0001	*******	1,00
	Markov staying in non-social	0,68	0,56	0,0010	**	0,81
	Markov non-social to social	0,19	0,19	0,8844		0,03
	Markov social to non-social	0,13	0,07	0,0000	*******	1,06
	Markov staying in social	0,83	0,91	0,0026	**	0,74
Ostensive	Ratio of person (%)	46,9%	58,3%	0,0327	*	0,51
	Ratio of target (%)	13,3%	13,2%	0,9550		0,01
	Markov staying in person	0,76	0,86	0,0110	*	0,62
	Markov person to target	0,07	0,07	0,5269		0,15
	Markov person to nontarget	0,06	0,04	0,2564		0,27

Values in bold represent significant differences, and values in italics indicate large Hedge’s g effect sizes.

Differences in GRI variables in preferential situations between the hrASD and non-ASD groups were tested using independent samples t-tests, and Hedge’s g effect sizes were also calculated (see Table 4). For the variables of density, gaze retention in AOI, total gaze retention time and gaze stuck, the mean of the hrASD subjects was significantly lower in the social ranges and significantly higher in the non-social ranges than the mean of the non-ASD subjects. Hedge’s g effect sizes were large/very large for all significant differences except for gaze stuck in social situations (moderate effect). All differences remained significant after correction for multiple comparisons, except for gaze stuck in social situations (see OSFDR).

Table 3

Independent samples t-test results and effect sizes of GRI variables in preferential situations
		Mean hrASD	Mean Non-ASD	p	Sig.	Hedges's g effect size
Density (1/min)	Non-social	20,9	12,7	0,0000	*******	1,19
Density (1/min)	Social	32,5	38,1	0,0064	**	0,66
Duration (msec)	Non-social	763,4	747,0	0,7295		0,08
Duration (msec)	Social	999,4	1076,2	0,3203		0,23
Gaze retention in AOI (%)	Non-social	34,2%	23,0%	0,0000	*******	1,06
Gaze retention in AOI (%)	Social	56,6%	70,5%	0,0000	*******	1,10
Total gaze retention time (msec)	Non-social	2361,3	1363,5	0,0000	*******	1,14
Total gaze retention time (msec)	Social	4308,6	5534,3	0,0004	*******	0,87
Latency (msec)	Non-social	1644,2	1616,3	0,8596		0,04
Latency (msec)	Social	1151,1	1020,9	0,2000		0,30
Gaze stuck (msec)	Non-social	123,6	62,1	0,0060	**	0,66
Gaze stuck (msec)	Social	232,9	311,8	0,0239	*	0,54

GRI: gaze retention interval. Values in bold represent significant differences, and values in italics indicate large Hedge’s g effect sizes.

Pearson’s correlations between ADOS scores and eye-movement variables

Pearson’s correlations between ADOS scores and eye-movement variables were computed in the group of all subjects. In preferential situations, a positive correlation was found between the ADOS score and the non-social ratio and the Markov variables of staying in a non-social state and transitioning from a social to a non-social state, and a negative correlation was found between the ADOS score and the social ratio and the Markov variable of staying in a social state. In ostensive situations, a positive correlation was found between the ADOS score and the Markov variable of transitioning from person to non-target state, and a negative correlation was found between the ADOS score and the variable of person ratio and the Markov variable of staying in person state (Table 5). All significant correlations remained significant after correction for multiple comparisons, except the Markov variable of transitioning from person to non-target state in ostensive situations (see OSFDR).

In addition, Pearson’s correlations between ADOS scores and eye-movement variables were examined separately in the group of hrASD and non-ASD subjects. The difference (expressed in Fisher's z score) between the correlations in the non-ASD and hrASD groups and the significance of the difference were also analyzed (Table 5). In the non-ASD group in preferential situations, a weak positive correlation was found between the ADOS score and the Markov variable of transitioning from social to non-social state, and a weak negative correlation was found between the ADOS score and the variable of social ratio. No significant correlation was found in the ostensive situations. In contrast, in the hrASD group in preferential situations, a strong positive correlation was found between the ADOS score and the non-social ratio and the Markov variable of staying in a non-social state and transitioning from a social to a non-social state. A strong negative correlation was found between the ADOS score and the variable of social ratio and the Markov variable of staying in social state. In ostensive situations, a negative correlation was found between the ADOS score and the variable of person ratio and the Markov variable of staying in person state (Fig. 5). The differences between the correlations of hrASD and non-ASD subjects were not significant, except for the Markov variable of transitioning from non-social to social state. All significant correlations in the group of hrASD subjects remained significant after correction for multiple comparisons; however, the two significant correlations in the group of non-ASD subjects did not survive the stringent correction for multiple comparisons (see OSFDR).

Table 4

Correlations between ADOS scores and eye-movement variables in preferential and ostensive situations
		ALL		Non-ASD			hrASD		Comparison of correlations
		r(X,Y)	p	r(X,Y)		p	r(X,Y)	p	Fisher'z	p
Preferential	Non-social ratio (%)	0,59	0,0000	0,09	0,5677		0,51	0,0030	-1,90	0,0574
	Social ratio (%)	-0,63	0,0000	-0,35	0,0250		-0,66	0,0000	1,74	0,0818
	Staying in non-social (Markov)	0,44	0,0001	-0,01	0,9592		0,41	0,0187	-1,82	0,0688
	Staying in social (Markov)	-0,49	0,0000	-0,28	0,0763		-0,51	0,0031	1,10	0,2714
	Transitioning from non-social to social (Markov)	0,00	0,9919	0,36	0,0201		-0,26	0,1566	2,60	0,0094
	Transitioning from social to non-social (Markov)	0,59	0,0000	0,31	0,0522		0,45	0,0098	-0,69	0,4902
Ostensive	Ratio of person (%)	-0,41	0,0004	-0,30	0,0569		-0,45	0,0092	0,73	0,4654
	Ratio of target (%)	0,05	0,7009	0,11	0,4908		0,07	0,6913	0,15	0,8808
	Staying in person (Markov)	-0,44	0,0001	-0,22		0,1672	-0,45	0,0107	1,04	0,2984
	Transitioning from person to target (Markov)	0,02	0,8700	0,01		0,9411	-0,14	0,4448	0,62	0,5352
	Transitioning from person to non-target (Markov)	0,25	0,0309	0,15		0,3501	0,34	0,0542	-0,84	0,4010

In each column, the r and p values of the correlations between ADOS score and eye movement variables are presented for all subjects (hrASD and non-ASD together), for non-ASD subjects and for hrASD subjects. The last two columns show the difference (expressed in Fisher's z score) between the correlations of non-ASD and hrASD subjects and the significance of the difference. Values in bold represent significant correlations/differences.

Pearson’s correlations between ADOS scores and GRI variables were examined in the group of all subjects in preferential situations. A significant correlation was found between the ADOS score and the variables of density, gaze retention in AOI, total gaze retention time and gaze stuck. The correlations were positive in non-social situations and negative in social situations. In contrast, latency correlated positively in social situations (Table 6).

In addition, Pearson’s correlations between ADOS scores and GRI variables were examined separately in the group of hrASD and non-ASD subjects. The difference (expressed in Fisher's z score) between the correlations in the non-ASD and hrASD groups and the significance of the difference were also analyzed (Table 6). In the non-ASD group in social situations, a weak negative correlation was found between the ADOS score and the variables of duration, gaze retention in AOI, total gaze retention time and gaze stuck. In the hrASD group, a strong significant correlation was found between the ADOS score and the variables of density, gaze retention in AOI, total gaze retention time and gaze stuck. The correlations were positive in non-social situations and negative in social situations. In contrast, latency correlated positively in social situations and negatively in non-social situations. Duration correlated only in non-social situations, and the correlation was positive (Figs. 6 and 7). The difference between the correlations of hrASD and non-ASD subjects was significant in non-social situations for the variables of duration, gaze retention in AOI and total gaze retention time and in social situation for the variables of density and gaze retention in AOI. All significant correlations in the group of hrASD subjects remained significant after correction for multiple comparisons; however, significant correlations in the group of non-ASD subjects did not survive the stringent correction for multiple comparisons (see OSFDR).

Table 6

Correlations between ADOS scores and gaze retention interval (GRI) variables in preferential situations
		ALL		Non-ASD		hrASD		Comparison of correlations
		r(X,Y)	p	r(X,Y)	p	r(X,Y)	p	Fisher'z	p
Density (1/min)	Non-social	0,58	0,0000	0,11	0,4830	0,41	0,0185	-1,34	0,1800
Density (1/min)	Social	-0,52	0,0000	0,00	0,9935	-0,77	0,0000	4,12	0,0000
Duration (msec)	Non-social	0,17	0,1378	-0,10	0,5250	0,57	0,0006	-3,08	0,0020
Duration (msec)	Social	-0,21	0,0733	-0,35	0,0244	-0,17	0,3509	-0,77	0,4380
Gaze retention in AOI (%)	Non-social	0,60	0,0000	0,14	0,3839	0,57	0,0007	-2,07	0,0380
Gaze retention in AOI (%)	Social	-0,68	0,0000	-0,36	0,0181	-0,71	0,0000	2,03	0,0420
Total gaze retention time (msec)	Non-social	0,60	0,0000	0,05	0,7433	0,54	0,0014	-2,26	0,0240
Total gaze retention time (msec)	Social	-0,58	0,0000	-0,38	0,0125	-0,63	0,0001	1,37	0,1720
Latency (msec)	Non-social	-0,13	0,2801	-0,09	0,5831	-0,44	0,0122	1,56	0,1200
Latency (msec)	Social	0,30	0,0102	0,07	0,6467	0,49	0,0041	-1,91	0,0560
Gaze stuck (msec)	Non-social	0,44	0,0001	0,06	0,6913	0,45	0,0099	-1,71	0,0860
Gaze stuck (msec)	Social	-0,41	0,0003	-0,38	0,0126	-0,42	0,0166	0,19	0,8520

One of the aim of our research was to create innovative eye-tracking variables that would be effective in tracking young children, in addition to the widely used AOI-based variables. We introduced two novel types of variables, one based on a Markov chain and the other on gaze retention (GRI). Furthermore, we aimed to understand how these specific eye-tracking variables function within two widely adopted social attention paradigms in ASD toddler research. We were particularly interested in assessing whether these new variables can effectively differentiate high-risk ASD toddlers from a population at elevated risk?

First, we examined the differences between the hrASD and non-ASD groups. As expected, we found significant differences between the two groups for most of the variables, although with different effect sizes. Typically, in the preferential paradigm, the distinct indicators of the social area were higher on average in the non-ASD group than in the hrASD group, while the non-social domain exhibited the opposite trend. These results are in accordance with the direction of previous studies (7–11), as well as with our hypotheses. In the joint attention paradigm, the group differences seemed to manifest in relation to the prompting and ostensive aspects of the person. The hrASD group consistently demonstrated a lower level of social attention towards the person, albeit with a smaller effect size compared to the preferential paradigm.

Given the substantial differences observed between the groups in these paradigms, we were interested in assessing the effect size variations among our different indicators. After this assessment we can argue that based on the frequently used AOI ratio variables the preferential paradigm is more effective in differentiating between the hrASD and non-ASD groups compared to the joint attention paradigm.

Our results regarding the Markov variables – which are specifically intended to capture the shift dynamics between two distinct AOIs – yielded intriguing results in the preferential paradigm. Notably, we observed a higher probability of shifts from social to non-social with a higher probability in the hrASD group, while no significant differences were found between the groups in non-social to social shifts. Hence, it appears that frequent shifts between AOIs are not a distinctive characteristic of the ASD group itself. It is important to highlight that although there were two main AOIs (social and non-social) in the preferential paradigm, they did not encompass the entire stimulus area, although they covered a significant section of it.

We examined the differences in the gaze retention interval (GRI) variables in the preferential paradigm, anticipating that certain variables would yield a larger effect size difference compared to the AOI ratio. According to the group comparisons of the GRI variables, the effect sizes of variables related to non-social area ranged between 0.66–1.19, while the effect size of variables related to social area ranged between 0.5–1.10. Similar to the variables based on AOI, the average for the hrASD group was higher in the non-social variables compared to social variables’.

Our overarching objective to identify eye-tracking variables that potentially indicate the suspicion of ASD; therefore, it was fundamental for us to examine the relationship between our variables and the gold standard method for ASD diagnosis, the ADOS scores. We operate under the assumption that a linear relationship between the measured variables and the ADOS scores would validate that the observed group differences are indeed linked to autism.

According to our results, regarding our AOI-based variables, the social ratio in the preferential paradigm had the strongest negative correlation with the ADOS score: the higher the toddler scored on ADOS, the less they spent looking at the social AOI area. We obtained a similar correlation with direct proportionality for the non-social area, and in both cases, this relationship proved to be significant with a higher effect size within the hrASD group compared to the non-ASD group. Hence, our assumption that ADOS scores show a linear relationship with our variables is mostly valid within the hrASD group. This coincides with the fact that the ADOS was created specifically for diagnostic purposes; thus, it is primarily for categorical separation, and as it was indicated by earlier researches (35, 36) it separates different levels of severity within the group of diagnosed individuals.

While the Markov variables exhibited weaker correlations with the ADOS score (0.41–0.51), they do substantiate that the non-social area garners more frequent and heightened attention within the hrASD group, especially with a higher ADOS score. Regarding ostensive stimuli, the attention-related variables towards person demonstrated a negative correlation with ADOS scores, albeit only of moderate strength. In the case of ostensive stimuli, the variables of attention to the person show a negative correlation with the ADOS scores but only a moderate correlation.

With regard to GRI variables, those associated with the non-social area exhibit a positive correlation, whereas variables related to the social area display a negative correlation with the ADOS score. In addition, the correlation was more significant in the hrASD group, although in several cases, it showed a significant correlation even within the non-ASD group (e.g., GRI in AOI – social, total gaze retention time – social, gaze stuck). In summary, the most pronounced effect sizes were observed with variables related to the social domain. Additionally, it's important to emphasize that several GRI variables displayed a higher correlation coefficient compared to the AOI social ratio indicator.

In summary, we conducted testing on our innovative methodologies aimed at differentiating toddlers with ASD from an elevated-risk population. Our approach involved the use of a custom-built eye-tracking device, specifically designed to accommodate young children by allowing for head movements and eliminating the need for calibration.

In our study, we looked at two types of social attention paradigms, and found that the preferential paradigm is more suitable for distinguishing toddlers with autism.

Moreover, we developed sufficiently robust variables through aggregation and their inherent nature. We employed gaze retention interval (GRI), which encompass a broader spatial and temporal range, than fixation. GRI variables proved to be exceptionally good distinguishable variables.

We concluded that in the preferential paradigm, apart from AOI ratio, the Markov and GRI variables can also prove effective in differentiating ASD toddlers. Additionally, they demonstrate potential for predicting the level of severity and the probability of autism based on correlations with the ADOS. It is plausible that, when employed in combination, these variables capture characteristics indicative of early signs of autism. We anticipate that our more robust indicators and results will contribute to early autism risk screening in the future.

ADOS - Autism Diagnostic Observation Schedule

AOI – area of interest

ASD – autism spectrum disorder

GRI – gaze retention interval

IDT - identification by dispersion threshold

OSFDR – OSF Data Repository

We confirm that the manuscript has been read and approved by all named authors and that there are no other persons who satisfied the criteria for authorship but are not listed. We further confirm that the order of authors listed in the manuscript has been approved by all of us.

Ethical approval

The research had been performed in accordance with the Declaration of Helsinki. The present study was approved by the Scientific and Research Ethics Committee of the Medical Research Council from Hungary (File number: 3502-5/2022/EÜIG). All participants’ legal guardians signed an informed consent form during registration for the study.

Consent for publication

People who appear in the video stimuli – and in the picture of the devices – (or their legal guardians) also sign an informed consent for publication of their identifying images for the purpose of research.

Availability of data and materials

The datasets generated and/or analyzed during the current study are available in the OSF data repository [https://osf.io/hvxu9/?view_only=30732c28ec094e1eba87a9ede464a002].

Competing interests

The authors of this submission received honoraria from ADDAS Ltd for consultancy services and/or other professional engagements. The compensation received from ADDAS Ltd does not influence the research presented in this article, and the authors declare no other conflicts of interest. The study was conducted independently, and the results and interpretations presented here are based on unbiased analysis and findings.

Funding

The research was founded by GINOP Grant (GINOP-2.1.1-15-2016-00857) received by ARD Medical Ltd. and NANOTRADE Ltd.

We declare that additional expenses in relation to our research have been sponsored by ADDAS Technology Ltd.

Author contributions

Orsolya Pachner contributed to the design and implementation of the research and was a major contributor in writing the manuscript.

Patrícia Gerván were involved in planning and supervised the work, contributed to the design and implementation of the research and worked on the manuscript.

Péter Soltész calculated gaze level and part of item level eye-tracking data and worked on the manuscript.

Ferenc Gombos analyzed and interpreted the eye-tracking data and made group statistics, tables and figures and worked on the manuscript.

Dóra Gyebnár contributed to sample preparation, worked on the manuscript.

All authors discussed the results, commented and approved the final manuscript.

Acknowledgements

The authors thank Iulia Savos and Ágnes Arató for their assistance in data collection. The authors wish to express their appreciation to the families and their children for their participation.

Authors information

We understand that Orsolya Pachner is the sole contact for the editorial process. She is responsible for communicating with the other authors about progress, submissions of revisions and final approval of proofs.

Falck-Ytter T, Bölte S, Gredebäck G. Eye tracking in early autism research. J Neurodev Disord. 2013 Dec;5(1):28.
Papagiannopoulou EA, Chitty KM, Hermens DF, Hickie IB, Lagopoulos J. A systematic review and meta-analysis of eye-tracking studies in children with autism spectrum disorders. Soc Neurosci. 2014 Jul 2;1–23.
Setien-Ramos I, Lugo-Marín J, Gisbert-Gustemps L, Díez-Villoria E, Magán-Maganto M, Canal-Bedia R, et al. Eye-Tracking Studies in Adults with Autism Spectrum Disorder: A Systematic Review and Meta-analysis. J Autism Dev Disord. 2023 Jun;53(6):2430–43.
Dalai G, Komandur S, Volden FS. Applications of Eye Tracking in the Diagnosis of Early Stages of Autism Spectrum Disorders: In: Proceedings of the 2nd International Conference on Computer-Human Interaction Research and Applications [Internet]. Seville, Spain: SCITEPRESS - Science and Technology Publications; 2018 [cited 2023 Oct 2]. p. 156–62. Available from: http://www.scitepress.org/DigitalLibrary/Link.aspx?doi=10.5220/0006951201560162
Guillon Q, Hadjikhani N, Baduel S, Rogé B. Visual social attention in autism spectrum disorder: Insights from eye tracking studies. Neurosci Biobehav Rev. 2014 May;42:279–97.
Laskowitz S, Griffin JW, Geier CF, Scherf KS. Cracking the Code of Live Human Social Interactions in Autism: A Review of the Eye-Tracking Literature. Proc Mach Learn Res. 2021 Oct;173:242–64.
New JJ, Schultz RT, Wolf J, Niehaus JL, Klin A, German TC, et al. The scope of social attention deficits in autism: Prioritized orienting to people and animals in static natural scenes. Neuropsychologia. 2010 Jan;48(1):51–9.
Sasson NJ, Touchstone EW. Visual Attention to Competing Social and Object Images by Preschool Children with Autism Spectrum Disorder. J Autism Dev Disord. 2014 Mar;44(3):584–92.
Pierce K, Conant D, Hazin R, Stoner R, Desmond J. Preference for Geometric Patterns Early in Life as a Risk Factor for Autism. Arch Gen Psychiatry. 2011 Jan 3;68(1):101.
Pierce K, Marinero S, Hazin R, McKenna B, Barnes CC, Malige A. Eye Tracking Reveals Abnormal Visual Preference for Geometric Images as an Early Biomarker of an Autism Spectrum Disorder Subtype Associated With Increased Symptom Severity. Biol Psychiatry. 2016 Apr;79(8):657–66.
Wen TH, Cheng A, Andreason C, Zahiri J, Xiao Y, Xu R, et al. Large scale validation of an early-age eye-tracking biomarker of an autism spectrum disorder subtype. Sci Rep. 2022 Mar 11;12(1):4253.
Chawarska K, Macari S, Shic F. Context modulates attention to social scenes in toddlers with autism: Context modulates social attention in autism. J Child Psychol Psychiatry. 2012 Aug;53(8):903–13.
Billeci L, Narzisi A, Campatelli G, Crifaci G, Calderoni S, Gagliano A, et al. Disentangling the initiation from the response in joint attention: an eye-tracking study in toddlers with autism spectrum disorders. Transl Psychiatry. 2016 May 17;6(5):e808–e808.
Parsons JP, Bedford R, Jones EJH, Charman T, Johnson MH, Gliga T. Gaze Following and Attention to Objects in Infants at Familial Risk for ASD. Front Psychol. 2019 Aug 20;10:1799.
Muratori F, Billeci L, Calderoni S, Boncoddo M, Lattarulo C, Costanzo V, et al. How Attention to Faces and Objects Changes Over Time in Toddlers with Autism Spectrum Disorders: Preliminary Evidence from An Eye Tracking Study. Brain Sci. 2019 Nov 27;9(12):344.
Chita-Tegmark M, Arunachalam S, Nelson CA, Tager-Flusberg H. Eye-Tracking Measurements of Language Processing: Developmental Differences in Children at High Risk for ASD. J Autism Dev Disord. 2015 Oct;45(10):3327–38.
Shic F, Wang Q, Macari SL, Chawarska K. The role of limited salience of speech in selective attention to faces in toddlers with autism spectrum disorders. J Child Psychol Psychiatry. 2020 Apr;61(4):459–69.
Kwon MK, Moore A, Barnes CC, Cha D, Pierce K. Typical Levels of Eye-Region Fixation in Toddlers With Autism Spectrum Disorder Across Multiple Contexts. J Am Acad Child Adolesc Psychiatry. 2019 Oct;58(10):1004–15.
Camero R, Martínez V, Gallego C. Gaze Following and Pupil Dilation as Early Diagnostic Markers of Autism in Toddlers. Children. 2021 Feb 5;8(2):113.
Habayeb S, Tsang T, Saulnier C, Klaiman C, Jones W, Klin A, et al. Visual Traces of Language Acquisition in Toddlers with Autism Spectrum Disorder During the Second Year of Life. J Autism Dev Disord. 2021 Jul;51(7):2519–30.
Jones W, Klin A. Attention to eyes is present but in decline in 2–6-month-old infants later diagnosed with autism. Nature. 2013 Dec;504(7480):427–31.
Bradshaw J, Shic F, Holden AN, Horowitz EJ, Barrett AC, German TC, et al. The Use of Eye Tracking as a Biomarker of Treatment Outcome in a Pilot Randomized Clinical Trial for Young Children with Autism. Autism Res. 2019 May;12(5):779–93.
Dalrymple KA, Manner MD, Harmelink KA, Teska EP, Elison JT. An Examination of Recording Accuracy and Precision From Eye Tracking Data From Toddlerhood to Adulthood. Front Psychol. 2018 May 23;9:803.
Bacon EC, Moore A, Lee Q, Carter Barnes C, Courchesne E, Pierce K. Identifying prognostic markers in autism spectrum disorder using eye tracking. Autism. 2020 Apr;24(3):658–69.
Von Holzen K, Bergmann C. The development of infants’ responses to mispronunciations: A meta-analysis. Dev Psychol. 2021 Jan;57(1):1–18.
Luyster R, Gotham K, Guthrie W, Coffing M, Petrak R, Pierce K, et al. The Autism Diagnostic Observation Schedule—Toddler Module: A New Module of a Standardized Diagnostic Measure for Autism Spectrum Disorders. J Autism Dev Disord. 2009 Sep;39(9):1305–20.
Hong JS, Singh V, Kalb L, Ashkar A, Landa R. Replication study of ADOS ‐2 Toddler Module cut‐off scores for autism spectrum disorder classification. Autism Res. 2021 Jun;14(6):1284–95.
Robins DL, Fein D, Barton ML, Green JA. The Modified Checklist for Autism in Toddlers: An Initial Study Investigating the Early Detection of Autism and Pervasive Developmental Disorders. J Autism Dev Disord. 2001;31(2):131–44.
Hessels RS, Kemner C, Van Den Boomen C, Hooge ITC. The area-of-interest problem in eyetracking research: A noise-robust solution for face and sparse stimuli. Behav Res Methods. 2016 Dec;48(4):1694–712.
Xun J. The Research of Markov Chain Application under Two Common Real World Examples. J Phys Conf Ser. 2021 Jun 1;1936(1):012004.
Andersson R, Larsson L, Holmqvist K, Stridh M, Nyström M. One algorithm to rule them all? An evaluation and discussion of ten eye movement event-detection algorithms. Behav Res Methods. 2017 Apr;49(2):616–37.
Salvucci DD, Goldberg JH. Identifying fixations and saccades in eye-tracking protocols. In: Proceedings of the symposium on Eye tracking research & applications - ETRA ’00 [Internet]. Palm Beach Gardens, Florida, United States: ACM Press; 2000 [cited 2023 Oct 2]. p. 71–8. Available from: http://portal.acm.org/citation.cfm?doid=355017.355028
Stark L, Ellis SR. Scanpath revisited:Cognitive models of direct active looking. In: EyeMovements: Cognition and Visual Perception. Hillsdale: NJ: Erlbaum; 1981. p. 193–226.
Widdel H. Operational problems in analysingeye movements. In: Theoretical and Applied Aspects of Eye MovementResearch. New York: Elsevier; 1984. p. 21–9.
Gotham K, Pickles A, Lord C. Standardizing ADOS Scores for a Measure of Severity in Autism Spectrum Disorders. J Autism Dev Disord. 2009 May;39(5):693–705.
Hus V, Gotham K, Lord C. Standardizing ADOS Domain Scores: Separating Severity of Social Affect and Restricted and Repetitive Behaviors. J Autism Dev Disord. 2014 Oct;44(10):2400–12.

Competing interest reported. The authors of this submission received honoraria from ADDAS Ltd for consultancy services and/or other professional engagements. The compensation received from ADDAS Ltd does not influence the research presented in this article, and the authors declare no other conflicts of interest. The study was conducted independently, and the results and interpretations presented here are based on unbiased analysis and findings.

Download PDF

Version 1

posted

You are reading this latest preprint version

Unveiling Distinctive Eye-Tracking Markers to Distinguish Toddlers with High-Risk Autism Spectrum Disorder within an Elevated Risk Toddler Population

Status:

Version 1

Abstract

Background

Methods

Results

Conclusions

Figures

Background

Methods

Participants

Procedure

Clinical Measures

ADOS-2 Toddler module

Ophthalmological screening

M-CHAT

Eye-tracking methods

Stimuli and presentation

Eye tracker

Eye-tracking variables

Variables used in statistical tests

Statistical method

Results

Study population

Statistical analysis

Differences between non-ASD and hrASD groups

Pearson’s correlations between ADOS scores and eye-movement variables

Discussion

Conclusions

Abbreviations

Declarations

References

Additional Declarations

Status:

Version 1