Identifying Children Autism Spectrum Disorder via Machine Learning based Behavior Analysis

Autism Spectrum Disorder (ASD) is a group of lifelong neurodevelopmental disorders with complicated causes. A key symptom of ASD patients is their impaired interpersonal communication ability. In this thesis, we consider the problem of screening/diagnosing children autism spectrum disorder by analyzing their behaviors through machine learning. 

Our work is motivated by the broad spectrum of previous research which indicates that children with ASD are often characterized by certain behaviors such as abnormal visual attention, lack of response to names, and impaired interpersonal communication abilities. Such behavior-level signs motivate us to analyze and identify these abnormalities under a variety of different modalities with data-driven approaches. Specifically, we begin by identifying ASD children based on their face scanning patterns. We consider using a bag-of-words (BoW) model to encode face scanning patterns, and further propose a novel dictionary learning method via discriminative mode seeking to improve the BoW representation and the identification accuracy. 

To render more natural and spontaneous children reactions, we further consider an interactive diagnostic procedure under a multi-camera, multimodal system where children activities are recorded with minimum constraints. Three assessment protocols originated from the Autism Diagnostic Observation Schedule-Generic (ADOS-G) are designed: 1. response to name, 2. separation and reunion, 3. response to non-social sound stimulation. We comprehensively analyze the children behaviors through a number of computer vision, speech processing, and general machine learning approaches. Some typical problems we consider include preprocessing steps such as person detection/re-identification, pose estimation, as well as feature extraction/score prediction on top of preprocessing. 

Comprehensive experimental results show that the proposed frameworks not only can effectively identify ASD, but also help human diagnosis by providing an auxiliary view with mid-level machine features/scores. Although the proposed work is yet too preliminary to directly replace existing autism assessment methods in clinical practice, it shed light on future applications of machine learning methods in early screening of the disease.