Poster #44 - Identifying Pivotal Transitions In Early Development For Children At Risk of Autism Using Sparse Coding

• In Event: 1-005 - Poster Session 01
  In Poster Session: PS 01 Section - Developmental Disabilities

Integrative Statement

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder characterized by social communication deficits and restricted interests and repetitive behaviors. Key challenges for early diagnosis of ASD lie in detecting delays and deviations in pivotal transitions that characterize social development over the first years of life, and in designing efficient screening protocols that are sensitive and specific in identifying both the severity and timing of departures from typical development as each individual child's development unfolds.
Densely sampled longitudinal designs have already been used to collect large amounts of behavioral data from infants at risk of ASD and typically developing controls over the first three years of life. Remarkably, research studies using eye-tracking, vocal recording and neuroimaging have now identified biomarkers for autism within the first 12 months. Critically, all of these biomarkers take the form of trajectories that measure change over time. Current methods for longitudinal data analysis used to calculate trajectory biomarkers traditionally rely on least-squares (L2) fitting of growth curve models to data, but these perform poorly when data are sparse or irregularly sampled, and corrupted by natural variability or measurement noise.
The goal of this study is to test the hypothesis that subtle differences in social interaction characteristic of ASD can be detected within the first two years of life. By applying recent advances in compressed sensing, we show that pivotal transitions signaling the onset of autism can be accurately detected from small numbers of carefully randomized sample points, based on a densely sampled reference database.
As part of an NIH Autism Center of Excellence, we followed 34 low-risk and 44 high-risk infants from 0-36 months. Using a digital audio recording device (LENA) worn by each child, we made day-long audio recordings of each child's language environment every month from 2-24 months. Using specialized vocal analysis software, we counted the number of vocalizations per hour for infant and caregiver, as well as the rate of vocal interactions in each recording.
In previous analyses, we used Functional Data Analysis to identify pivotal transitions in vocal contingency, beginning around 12 months. In the current study, we designed tailored dictionaries of biorthogonal basis functions capable of capturing these transitions, and applied sparse coding techniques to reconstruct individual trajectories from a minimal number of components using L1 optimization. By repeating this analysis on permuted subsets of data, we were able to determine the number and distribution of sample points needed to detect significant differences between risk groups. Using only 4 sample points randomly located around 12 months, we were able to achieve the same performance we only obtained previously using the full data set. Furthermore, we found evidence of additional, earlier transitions that were obscured in our initial analysis.
Our results provide further evidence of early differences identifying infants at risk of ASD, and also demonstrate that sparse coding is effective in reducing the number of sample points needed to detect such differences, leading to new directions for designing cost-effective, community-viable screening instruments.

Author

• Gordon Ramsay, Marcus Autism Center, Children's Healthcare of Atlanta & Department of Pediatrics, Emory University School of Medicine
  Presenting Author