A two-step temporal data augmentation and supervised learning framework for predicting autism diagnosis at 36 months

Cancan Zhang1, Runqiu Wang2, Jamie K Capal3

  • 1Division of General Medicine, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA.

Insights

Early autism spectrum disorder (ASD) prediction in children with tuberous sclerosis complex (TSC) is possible by combining diffusion tensor imaging (DTI) and behavioral data. Machine learning models effectively identified key early biomarkers for ASD outcomes.

Area of Science:

  • Neuroscience
  • Developmental Pediatrics
  • Machine Learning

Background:

  • Autism spectrum disorder (ASD) affects a significant portion of children with tuberous sclerosis complex (TSC).
  • Early ASD identification in this high-risk group is challenging due to complex factors.
  • Timely intervention is crucial for improving outcomes in children with TSC and ASD.

Purpose of the Study:

  • To integrate longitudinal diffusion tensor imaging (DTI) metrics with early behavioral features.
  • To predict autism spectrum disorder (ASD) outcomes at 36 months using supervised learning.
  • To identify key early biomarkers for ASD in children with TSC.

Main Methods:

  • Utilized DTI metrics (axial diffusivity, fractional anisotropy, mean diffusivity, radial diffusivity) from 27 white matter tracts.
  • Developed a data augmentation algorithm to standardize DTI data to ages 12, 24, and 36 months.
  • Included 9 behavioral features from ADOS-2 and ADI-R assessments at 24 months; trained supervised learning models.

Main Results:

  • Regularized logistic regression models (LASSO, Elastic Net) showed balanced performance.
  • Compared two input settings: DTI at 24 months + behavior vs. DTI at 12 & 24 months + behavior.
  • Setting 1 (24-month DTI + behavior) performed comparably or better in predicting ASD diagnosis.

Conclusions:

  • Integrating early neuroimaging (DTI) and behavioral data shows promise for predicting ASD outcomes in children with TSC.
  • A multimodal machine learning approach highlights 24-month DTI and behavioral measures as key early biomarkers.
  • Regularized regression techniques are effective for analyzing small, heterogeneous clinical datasets.
Abstract

Related Concept Videos