Early Identification of as Disorder using Machine Learning Based Classifier System Implementation

In this work, we utilized resting-state functional connectivity (FC) data to construct diagnostic classifiers using machine learning classifiers like random forest algorithms on four different autism spectral disorders (ASD) samples with a total of participants, ranging in age from 6 to 18 years. We purposefully altered the makeup of each sample to vary the degrees of gender diversity and symptom intensity, allowing for some overlap between groups and accounting for the wide range of ASD symptoms. Each sample had specific inclusion requirements, such as: all genders with an open-ended severity range; male participants with an open-ended severity; all genders with an emphasis on higher severity levels; and male participants only with higher severity levels. There were subjects in each subset (ASD, TD), carefully matched for head motion and age, of whom were assigned to training and to validation. Within every sample, random forest classifiers were trained. Sample through sample exhibited classification accuracies of corresponding to after testing in validation samples. Important factors affecting classifier performance were shown to vary among sample sets, including the connections within the cingulo-opercular task control (COTC) network and the interactions between COTC ROIs and the default mode and dorsal attention networks. Our findings demonstrate how difficult it can be to develop diagnostic classifiers based on the characteristics of ASD samples. The efficacy of classifiers is enhanced with greater uniformity about gender and symptom intensity. Nonetheless, it is crucial to recognize the inherent variability of ASDs, as relying exclusively on performance metrics may not be sufficient to accurately assess the usefulness of classifiers in practical applications.