Using Baby Facereader to Investigate Facial Expressiveness in at-Risk Populations

Background: Autism spectrum disorders (ASDs) are characterized in varying degrees by difficulties in social interaction, verbal & nonverbal communication (NVC) and repetitive behaviors. Looking at the face as a medium of NVC, researchers have often limited their research scope in studying smiles or laughter at high-risk (HR) populations. While smiling is a facial action with strong relevance to ASD, a detailed investigation into intensity, variability and lability of facial expressions is highly relevant but often avoided due to the cumbersome nature of detailed behavior coding such as the Baby Facial Action Coding System. Automatic facial expression coding solutions in the field of computer vision are a promising approach to overcoming this difficulty. To date, however, such solutions measure facial expression in adults and do not generalize well to infant populations. This demonstration presents Baby FaceReader, a novel solution that automatically measures infant facial expressions. Furthermore, the demonstration shows how Baby FaceReader can be used to investigate facial expressiveness in HR populations.

Objectives: To demonstrate that Baby FaceReader is a valid tool to measure facial expression behavior in terms of action units (AUs) in infants. To use Baby FaceReader to evaluate intensity, variability and lability of facial expression behavior in HR populations.

Methods: To evaluate Baby FaceReader’s validity we collected all images from the Baby FACS manual that have been comprehensively coded for Action Units. After removing duplicates, the dataset consisted of 74 images. We ran Baby FaceReader on the dataset to extract AU classifications and computed F1 scores for each AU. F1 scores are a standard evaluation metric of predictions in pattern recognition and are computed using the formula in figure 1.

To evaluate intensity, variability and lability of facial expression behavior in HR we ran Baby FaceReader on a dataset of videos of two-year infants (20 HR, 5 control) during the social stimuli presentation of an eye-tracking paradigm. To compute (a) intensity, we took the average intensity of all expressed AUs during the social stimuli presentation; (b) variability, we counted the number of different facial configurations produced during the social stimuli; (c) lability, the number of times the infant’s facial configuration changed during the social stimuli.

Results: Based on previous results, we expect to reach an average F1 score of at least 0.7 across all AUs in our evaluation of Baby FaceReader. We have not yet evaluated intensity, variability and lability in our video dataset.

Conclusions: Automatic coding is a promising advance in studying infant facial expression as it allows more time and depth for analysis. Baby FaceReader is one such complete solution that could be used in the study of ASD during development.