Next-Generation Emotion Recognition Markers for Autism Spectrum Disorder

Background: The factors driving recognition of facial emotion, and the impairment of this ability in some autistic individuals, are poorly understood. This lack of understanding may partially reflect psychometric challenges associated with static faces. Static faces are heterogeneous, have potentially limited dynamic range, and have an underexplored relationship to real, dynamic facial expressions. Each of these challenges motivates innovations in the development of more sensitive, generalizable, and clinically-useful markers of non-verbal receptive communication for autistic people.

Objectives: This work addressed the above challenges by designing a next-generation emotion recognition task (ERT) for use in clinical trials with autistic participants, leveraging computational and “participatory design” methods alike.

Methods: Starting with systematic review of the ERT literature, a candidate task was designed according to those task features most commonly associated with significant individual differences in real-world functioning (as quantified by adaptive behaviour) or with pharmacological effects (focusing on oxytocin/vasopressin manipulations in particular). Optimal stimuli for the task were selected using a Bayesian item-response theory (IRT) model of “norming” data from typically-developing volunteers. A series of focus groups were then conducted with the task, in two different countries, to directly incorporate feedback from individuals with Autism Spectrum Disorder. Finally, adaptive psychophysical algorithms were constructed to address the focus group feedback, and the algorithm’s performance characteristics were confirmed through Monte Carlo simulation.

Results: Our systematic review revealed that the majority of tasks eliciting either an association between ERT performance and a pharmacological intervention or individual differences in real-world functioning (as quantified by adaptive behaviour skills) occupy a relatively restricted design space. A task was proposed on this basis, with adaptations for frequent, low-burden administration on smartphones in a home-based setting. Specifically, individuals are asked to classify non-masked, static faces displaying 7 emotions (including a “neutral” emotion) of varying intensity in a self-paced manner. A Bayesian IRT was developed which correctly predicts static ERT performance from dynamic ERT training data, and further revealed psychometric advantages to using Emotient FACET coding (as compared to nominal emotional intensity ratings) for calibration of individual stimuli. Based on qualitative feedback collected from autistic individuals the task length was shortened by introducing an adaptive psychophysical staircase algorithm for determining the optimal intensity of the displayed emotion stimuli based on individuals’ previous responses, without losses to estimation accuracy (Monte Carlo simulations). The final assessment, involving all of these features, is now deployed to autistic individuals enrolled in two clinical trials (one interventional, one observational).

Conclusions: A combination of modelling methods and participatory design is needed to create next-generation marker tasks which are both powerful and feasible for high-frequency, at-home testing in clinical trials. Our application of these methods has yielded a promising assessment of non-verbal receptive communication, currently being validated in two deployments. Ultimately, we plan to make this assessment available to the field at large.