A preliminary evaluation of still face images by deep learning: A potential screening test for childhood developmental disabilities

Abstract

Most developmental disorders are defined by their clinical symptoms and many disorders share common features. The main objective of this research is to evaluate still facial images as a potential screening test for childhood developmental disabilities, which is free of any biases of subjective judgments of human observers. Via supervised machine learning, a classifier of convolution neural network (CNN) was built by using 908 facial images, half of those were photos of children labeled with “autism”, which may include some developmental disorders with autism-like features. Then face images were generated for two categories of photos. Above all, the most important discovery of this research is that face images labeled “autism” and normal controls populate two quite distinctive manifolds. Different pattern was found to be distributed in the eyes and mouth in the generated photos for two categories of faces by deep learning. It is showed that supervised machine learning can obtain facial features, which could possibly be applicable to improve early screening for childhood developmental disabilities by facial expression. A simple computer-based screening test of still face images may prove to be a useful adjunct in many clinical settings.