Employing K-Means Clustering to Identify Key Performance Indicators for Autism: Data from Eye-Tracking in Joint Attention Tasks

Background: Joint attention might be shown in the developmental deficits of children with autism spectrum disorder (ASD). To quantify performance of joint attention, eye-tracking data analysis is often considered, where there are several performance indicators, namely the proportion of fixation time in the area of interests (referred to as fixation time), the latency of the first fixation (referred to as the first fixation) and the longest common subsequence (LCS) similarity (to quantify how the participant's gaze dynamically follows the given logic of the stimulus). However, it is still a research question which performance indicators can properly reflect the difference of joint attention skills between children with autism and typical developing (TD) children. K-means clustering is a popular method for cluster analysis in data mining and could be potentially applied to identify the most relevant performance indicators for eye-tracking data analysis in joint attention tasks.

Objectives: This study aimed at implementing K-means clustering to explore which performance indicators are the most relevant ones to reflect the capacity of joint attention of autism. We did not only focus on the individual performance indicator, but also investigating their correlations.

Methods: 14 ASD (mean age: 4.96±1.10) and 16 age-matched TD children (4.53±0.90) participated in joint attention tasks. Video stimuli were used to induce joint attention behaviors of the participants. The video consisted of a man sitting behind the table together with three objects (left, middle and right side of the table). The man would turn his head to one object, and the eye movement was recorded by the Tobii X3-120 eye-tracker. The flowchart of the proposed K-means clustering algorithm is shown in Fig. 1.

Results: Figure 2 (a) and (d) show the distribution of measured eye-tracking data for ASD and TD children based on one performance indicator (LCS similarity) and a combination of two performance indicators (fixation time and the first fixation), respectively. The corresponding clustering results by applying the proposed K-means algorithm can be seen in Fig. 2(b) and (e), respectively. Figure 2(c) and (f) show the exact number of ASD and TD in the group clustered by the proposed K-means algorithm. Chi-squared test reveals that as an individual performance indicator, the LCS similarity shows the most significant difference (p=0.002), the fixation time (p=0.024) still reach the significance, but the first fixation (p=0.157) is not significant. However, the difference becomes significant when combining the first fixation to the fixation time(p=0.024) as well as the LCS similarity (p=0.037). When all three performance indicators are considered, there is no significant difference (p=0.063).

Conclusions: We applied K-means clustering to identify the key performance indicator for eye-tracking data analysis in joint attention tasks. We found that the LCS similarity was most likely (99.8%) to reflect the difference between ASD and TD children in joint attention tasks. Although the first fixation does not show a significant difference as an individual performance indicator, its correlation with the fixation time and LCS similarity greatly improves the significance.