Novel pattern detection in children with Autism Spectrum Disorder using Iterative Subspace Identification

Recent increase in the number of Autism cases has triggered an alarm in our society. Lack of effective diagnostics, interventions and associated cost makes early intervention and long term treatment difficult. In this paper, we describe novel methods to assist management by automatically detecting stereotypical behavioral patterns using accelerometer data. We use the Iterative Subspace Identification (ISI) algorithm to learn subspaces in which the sensor data lives. It extracts orthogonal subspaces which are used to generate dictionaries for clustering and for signal representation. It is also applied to detecting segments from acoustics data. We further improve the algorithm by detecting novel events which were not known to the system during the training. Using these methods, we achieved an average of 83% and 90% of classification rates for flapping and rocking behaviors and 93% for novel behavioral patterns studied in this paper.