Applying Machine Learning to Kinematic and Eye Movement Features of a Movement Imitation Task to Predict Autism Diagnosis

Autism is a developmental condition currently identified by experts using observation, interview, and questionnaire techniques and primarily assessing social and communication deficits. Motor function and movement imitation are also altered in autism and can be measured more objectively. In this study, motion and eye tracking data from a movement imitation task were combined with supervised machine learning methods to classify 22 autistic and 22 non-autistic adults. The focus was on a reliable machine learning application. We have used nested validation to develop models and further tested the models with an independent data sample. Feature selection was aimed at selection stability to assure result interpretability. Our models predicted diagnosis with 73% accuracy from kinematic features, 70% accuracy from eye movement features and 78% accuracy from combined features. We further explored features which were most important for predictions to better understand movement imitation differences in autism. Consistent with the behavioural results, most discriminative features were from the experimental condition in which non-autistic individuals tended to successfully imitate unusual movement kinematics while autistic individuals tended to fail. Machine learning results show promise that future work could aid in the diagnosis process by providing quantitative tests to supplement current qualitative ones.

. Breakdown of eye movement behaviour features. Whole movement -movement between two locations of interest, static finger --part of the movement when the finger was static on the start/end locations, dynamic movement --part of the movement when the finger was moving between the start/end locations, SD --standard deviation, IA --interest area.

1/3
Supplementary Note. Discriminative eye movement behaviour and combined features.
Interpretation of why selected features were salient for classification can shed some light on the problem at hand. In the main article text, we gave our interpretation for the selected ten discriminative features from the kinematic feature set. Here we explore the most discriminative features from the eye movement and combined datasets selected using the Wrapped t-test method.
Eye movement dataset Table S1 shows that all selected eye movement features were based on visual attention to the finger performing movement sequences on the screen for participants to imitate (Fig. 1a in the main text). All ten selected features were similar and between feature inter-correlation of r = 0.70 was even higher than the inter-correlation in a full 48 feature eye movement dataset (r = 0.48). All except one feature were variability measures, all showing greater variability in visual attention in autistic compared to non-autistic individuals. The similarity of features did not allow us a detailed analysis of individual features similarly as for kinematic features (see section Discriminative features in the main manuscript). The only inference we could make was that greater variability in visual attention exhibited by autistic individuals was important for classification. Dynamic finger, horizontal offset from fingertip, SD A>N, p<0.001 Peak acceleration (direct condition), SD A>N, p<0.001 Table S1. Features selected with Wrapped t-test selection method in eye movement and combined datasets. Mean difference column shows whether the mean for a particular feature was greater for autistic (A) or non-autistic (N) class and gives a p-value of two-sample t-test. SD -standard deviation, diff. -difference.

Combined dataset
In a dataset in which both kinematic and eye movement features were combined, among most discriminative features nine out of ten were eye movement features (Table S1). All of the nine eye-movement features which were selected from the combined dataset were also selected in the eye movement dataset alone (nine features in the left and right parts of the table match). The single selected kinematic feature was a feature which was also most discriminative when Wrapped t-test selection was used on kinematic dataset alone (Fig. 7c in the main text). This observed consistency of selected features in different datasets signifies the stability of the Wrapped t-test method.
The single selected kinematic feature measured peak acceleration variability in the direct experimental condition. This is consistent with behavioural results in previous imitation studies which have shown significant between-group movement differences between direct and elevated trials 1-3 and with our finding that the majority of most discriminative features from the kinematic dataset were based on acceleration (see section Discriminative features in the main manuscript). Features selected using Wrapped t-test from a combined dataset gave better classification performance than features selected from either 2/3