ALATT-network: automated LSTM-based framework for classification and monitoring of autism spectrum disorder therapy tasks

Autism Spectrum Disorder affects the overall growth and development of children by limiting their social and cognitive skills, where a child can have low, medium, and high-functioning autism. Depending upon the level of autism, different applied behavioral and cognitive therapies are performed by therapists that may continue for months or years depending upon the severity of autism in a child. Applied Behavior Analysis (ABA) therapy is provided to improve a child’s social, communicational, and behavioral skills. Nevertheless, locating highly skilled therapists for an extended duration proves challenging in the realm of ABA therapy. Moreover, the progress of the child needs to be monitored at every stage of the therapy, which further increases the overall cost and time of the therapy. During the ABA sessions, various tasks e.g., imitation, joint attention, and turn taking are key measurements to analyze a child's progress. It's very challenging to monitor these tasks manually by the therapist and predict the overall child's progress. To overcome this problem, the paper presents a novel deep learning framework based on a Long Short-Term Memory network for the classification of ABA therapy tasks (ALATT-Network) i.e., Imitation, Joint Attention, and Turn-Taking. The proposed framework is trained on a large-scale skeleton dataset consisting of spatial and temporal information on autism therapy sessions. The DREAM dataset is employed for the experiments which is the largest publically available dataset for autism therapy sessions. The framework assists caregivers, doctors, and therapists to monitor the ongoing therapy sessions and predict the child’s progress. The proposed ALATT network employs different five optimizers and analyzes the performance of the network. In the experiment, the results show that the proposed ALATT-Network with Adam optimizer effectively learns the temporal and spatial features of skeleton movements and provides an accuracy of 79.3% for classification tasks.