Computational NeuroscienceCreating multimodal predictors using missing data: Classifying and subtyping autism spectrum disorder

- Novel classifiers for multimodality imaging data that can handle missing data.
- Solves the real life issue of incomplete studies/missing data in clinical studies.
- Shows clear distinction in classifying autism spectrum disorder from normal controls as well as in stratifying Autism based on language impairment based on MEG and DTI features.
- Ranks and specifies the distinctive features involved in the classification for further investigations.

BackgroundAutism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by wide range of symptoms and severity including domains such as language impairment (LI). This study aims to create a quantifiable marker of ASD and a stratification marker for LI using multimodality imaging data that can handle missing data by including subjects that fail to complete all the aspects of a multimodality imaging study, obviating the need to remove subjects with incomplete data, as is done by conventional methods.MethodsAn ensemble of classifiers with several subsets of complete data is employed. The outputs from such subset classifiers are fused using a weighted aggregation giving an aggregate probabilistic score for each subject. Such fusion classifiers are created to obtain a marker for ASD and to stratify LI using three categories of features, two extracted from separate auditory tasks using magnetoencephalography (MEG) and the third extracted from diffusion tensor imaging (DTI).ResultsA clear distinction between ASD and neurotypical controls (5-fold accuracy of 83.3% and testing accuracy of 87%) and between ASD/+LI and ASD/−LI (5-fold accuracy of 70.1% and testing accuracy of 61.1%) was obtained. One of the MEG features, mismatch field (MMF) latency contributed the most to group discrimination, followed by DTI features from superior temporal white matter and superior longitudinal fasciculus as determined by feature ranking.Comparison with existing methodsHigher classification accuracy was achieved in comparison with single modality classifiers.ConclusionThis methodology can be readily applied in large studies where high percentage of missing data is expected.