Lèvy Flights Search Patterns In Children with ASDs Exploring Social Stimuli

Background: It is known that, in order to efficiently scan a visual scene, humans employ a 'scale free' jump strategy that is geometrically similar to a class of random walks known as Lévy flights (Brockmann et al, 2000), for which the distribution tails are power-laws. This 'scale free' strategy results also in a particular distribution of the clusters of fixation points, that are typically concentrated in the relevant zones of the salience field. Since individuals with Autism Spectrum Disorders (ASDs) seem to be characterized by atypical visual attention toward complex social scenes (Klin et al., 2002), in which the salience of social cues is reduced for them compared to the controls, it might be of interest to investigate if abnormalities of Lèvy flights patterns of fixation points might underlie their atypical focus of visual attention.

Objectives: Investigate the Lèvy flights patterns and the clustering patterns of visual attention of children with ASDs exploring static images of human faces looking objects.

Methods: We looked at the scanpath patterns of  27 children with ASD (18 M; 9 F), aged between 1-8 yrs (mean= 4; SD=1,9), IQ Leiter-R=65 (sd=24), and 70 TDs controls (34 M; 36 F), aged between 1-8 yrs (mean = 3; SD=1,7). The stimuli were presented with a Tobii T60 Eye Tracker and were those used in Fadda et al, 2010: an adult face gazing at one of two objects located laterally on eye level, next to the head. The scene develops a narrow horizontal salience field that generates an x Vs y asymmetry. We measured the absolute jumps distribution both in the x and the y direction. In order to characterize the clustering patterns we performed a k-means analysis. In particular we studied the distortion parameter, a quality parameter for the convergence of the cluster structure, versus the number k of centroids.

Results: The results showed that children with ASDs had qualitatively the same Lévy flight behaviour as the TDs, but with a different slope in the distribution tails (2.08+-0.03 for the ASD vs 1.76+-0.06 for TD, x axis and 2.24+-0.04 for the ASD vs 1.97+-0.04 for TD, y axis). After a crossover region, we found that the ASDs had always on average an higher probability to perform long jumps. The cluster analysis revealed a decay with the number of centroids k, in which systematically the ASDs had a greater value of k for the same distortion, signaling on average a much higher number of clusters covering the scene for them.

Conclusions: In summary, the fundamental mechanism of visual exploration of a static social scene seems to be a Lèvy flight process both for ASDs and TDs, but with different values in the parameter that determines the probability of jumps. Indeed, the pattern of the clusters distribution for the ASDs is much more spreaded, especially in the zone far from the salience field. This latter characteristic points to an over exploratory behavior, even in empty zones, and to a more spotted strategy in probing the salience space.