Creating multimodal predictors using missing data: classifying and subtyping autism spectrum disorder.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Autism spectrum disorder (ASD) is a neurodevelopmental disorder with varied symptoms, including language impairment (LI).
• Conventional methods often exclude subjects with incomplete imaging data, limiting study scope.
• This study addresses missing data challenges in ASD and LI research.

Purpose of the Study:

• To create a quantifiable marker for ASD using multimodal imaging.
• To develop a stratification marker for LI within ASD.
• To accommodate subjects with incomplete data in multimodality imaging studies.

Main Methods:

• Ensemble of classifiers trained on data subsets.
• Weighted aggregation of classifier outputs for probabilistic scoring.
• Feature extraction from magnetoencephalography (MEG) auditory tasks and diffusion tensor imaging (DTI).

Main Results:

• Achieved 87% testing accuracy for ASD vs. neurotypical controls.
• Attained 61.1% testing accuracy for ASD with/without LI.
• Mismatch field (MMF) latency from MEG and DTI features were key discriminators.

Conclusions:

• The proposed methodology effectively handles missing data in large studies.
• Ensemble classification with multimodal data surpasses single-modality performance.
• This approach offers a robust tool for ASD and LI research.