Feature selection and classification of imbalanced datasets: application to PET images of children with autistic spectrum disorders.

摘要

Learning with discriminative methods is generally based on minimizing the misclassification of training samples, which may be unsuitable for imbalanced datasets where the recognition might be biased in favor of the most numerous class. This problem can be addressed with a generative approach, which typically requires more parameters to be determined leading to reduced performances in high dimension. In such situations, dimension reduction becomes a crucial issue. We propose a feature selection/classification algorithm based on generative methods in order to predict the clinical status of a highly imbalanced dataset made of PET scans of forty-five low-functioning children with autism spectrum disorders (ASD) and thirteen non-ASD low functioning children. ASDs are typically characterized by impaired social interaction, narrow interests, and repetitive behaviors, with a high variability in expression and severity. The numerous findings revealed by brain imaging studies suggest that ASD is associated with a complex and distributed pattern of abnormalities that makes the identification of a shared and common neuroimaging profile a difficult task. In this context, our goal is to identify the rest functional brain imaging abnormalities pattern associated with ASD and to validate its efficiency in individual classification. The proposed feature selection algorithm detected a characteristic pattern in the ASD group that included a hypoperfusion in the right Superior Temporal Sulcus (STS) and a hyperperfusion in the contralateral postcentral area. Our algorithm allowed for a significantly accurate (88%), sensitive (91%) and specific (77%) prediction of clinical category. For this imbalanced dataset, with only 13 control scans, the proposed generative algorithm outperformed other state-of-the-art discriminant methods. The high predictive power of the characteristic pattern, which has been automatically identified on whole brains without any priors, confirms previous findings concerning the role of STS in ASD. This work offers exciting possibilities for early autism detection and/or the evaluation of treatment response in individual patients.