A two-step temporal data augmentation and supervised learning framework for predicting autism diagnosis at 36 months in patients with tuberous sclerosis complex.
Authors
Affiliations (10)
Affiliations (10)
- Division of General Medicine, Beth Israel Deaconess Medical Center, Harvard Medical School, Boston, MA, USA.
- Department of Neurology, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA.
- Division of Neurology, Cincinnati Children's Hospital Medical Center, Department of Pediatrics, University of Cincinnati College of Medicine, Cincinnati, OH, USA.
- Department of Neurology, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA; Rosamund Stone Zander Translational Neuroscience Center, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA.
- Rosamund Stone Zander Translational Neuroscience Center, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA.
- Department of Neurology, University of Alabama School of Medicine, Birmingham, AL, USA.
- Department of Pediatrics, McGovern Medical School at the University of Texas Health Science Center at Houston and Children's Memorial Hermann Hospital, Houston, TX, USA.
- Northwestern University Feinberg School of Medicine and Ann & Robert H. Lurie Children's Hospital of Chicago, Chicago, IL, USA.
- Computational Radiology Laboratory, Department of Radiology, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA.
- Department of Neurology, Boston Children's Hospital, Harvard Medical School, Boston, MA, USA. Electronic address: [email protected].
Abstract
Autism spectrum disorder (ASD) affects approximately 25-50% of children with tuberous sclerosis complex (TSC). Early identification of ASD in this high-risk population is crucial for timely intervention but remains challenging due to the heterogeneous clinical presentation and complex interplay of genetic, neurological, and environmental factors. This study aimed to integrate longitudinal diffusion tensor imaging (DTI) metrics with early behavioral features using supervised learning algorithms to predict ASD outcomes at 36 months. Data were obtained from the children enrolled in the TSC Autism Center of Excellence Research Network study. Four DTI metrics: axial diffusivity, fractional anisotropy, mean diffusivity, and radial diffusivity, were measured across 27 major white matter tracts at up to four irregular time points. To account for variability in acquisition timing, we developed a two-step data augmentation algorithm to interpolate each subject's data to standardized ages of 12, 24, and 36 months. In addition, 9 behavioral features from the ADOS-2 and ADI-R assessments at 24 months were included. Supervised learning algorithms were trained to predict ASD diagnosis at 36 months under two input settings. Performance of the supervised learning algorithms was evaluated with accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve as performance metrics. Regularized logistic regression models, least absolute shrinkage and selection operator and elastic net, demonstrated the most balanced overall performance across most evaluation metrics. Comparing input settings, Setting 1 (DTI at 24 months + behavioral features) achieved comparable or slightly improved performance relative to Setting 2 (DTI at 12 and 24 months + behavioral features) in predicting ASD diagnosis. Integrating early neuroimaging and behavioral data suggests potential for prediction of ASD outcomes at 36 months in children with TSC. This multimodal machine learning framework highlights 24-month DTI and behavioral measures as key early biomarkers and demonstrates the effectiveness of regularized regression techniques for small-sample, heterogeneous clinical datasets.