Introduction: Neurodevelopmental disorders comprise a group of neuropsychiatric conditions. Presently, behavior-based diagnostic approaches are utilized in clinical settings, but the overlapping features among these disorders obscure their recognition and management. Attention deficit hyperactivity disorder (ADHD) and autism spectrum disorder (ASD) have common characteristics across various levels, from genes to symptoms. Designing a computational framework based on the neuroimaging findings could provide a discriminative tool for ultimate more efficient treatment. Machine learning approaches, specifically classification methods are among the most applied techniques to reach this goal.
Methods: We applied a novel two-level multi-class data maximum dispersion classifier (DMDC) algorithm to classify the functional neuroimaging data (utilizing datasets: ADHD-200 and autism brain imaging data exchange (ABIDE)) into two categories: Neurodevelopmental disorders (ASD and ADHD) or healthy participants, based on calculated functional connectivity values (statistical temporal correlation).
Results: Our model achieved a total accuracy of 62% for healthy controls. Specifically, it demonstrated an accuracy of 51% for healthy subjects, 61% for autism spectrum disorder, and 84% for ADHD. The support vector machine (SVM) model achieved an accuracy of 46% for both the healthy control and ASD groups, while the ADHD group classification accuracy was estimated to be 84%. These two models showed similar classification indices for the ADHD group. However, the discrimination power was higher in the ASD class. 
Conclusion: The method employed in this study demonstrated acceptable performance in classifying disorders and healthy conditions compared to the more commonly used SVM method. Notably, functional connections associated with the cerebellum showed discriminative power.