Generalized Prediction Model for Detection of Psychiatric Disorders

Computer aided diagnosis of mental disorders like Attention Deficit Hyperactivity Disorder (ADHD) and Autism is a primary step towards automated detection and prognosis of these psychiatric diseases. This dissertation applies analyses based on learning models that use structural texture and functional connectivity to diagnose ADHD and also Autism, from (structural) 3-dimensional magnetic resonance imaging (MRI) and 4-dimensional restingstate functional magnetic resonance imaging (fMRI) scans of subjects. One model learns texture-based filters that are used to extract features from MRI scans. Using these learned features, the model achieves 0.6257 (baseline 0.5497) accuracy on the ADHD-200 hold-out dataset for differentiating between healthy control vs ADHD patients and also achieves 0.6173 (baseline 0.5157) accuracy on the ABIDE (Autism) hold-out test for differentiating between healthy control vs Autism patients. Our next model examines temporal sequence of fMRI activation levels at various brain locations in order to make a diagnosis from fMRI scans. This incorporates spatial nonstationary independent component analysis of the fMRI scans in order to extract the uncorrelated components and decomposes fMRI scans into common spatial components and corresponding time courses. Using individual time courses of 45 independent components as features, our algorithm learns a classifier that yields an accuracy of 0.6491 on the ADHD-200 hold-out dataset, and 0.6233 accuracy on the ABIDE hold-out test. This result is higher (0.0231 for ADHD and 0.0233 for Autism) than previously published accuracies on these datasets using fMRI scans. Finally a combination of multimodal features yields 0.6725 diagnosis accuracy on ADHD-200 and 0.6431 accuracy on ABIDE. This result is significantly higher (0.0465 for ADHD with one sided p = 0.01 and 0.0431 for Autism with one sided p = 1.6172e-06) than previously published hold-out accuracies on these datasets using only imaging data. Our results indicate that combining multimodal features yields good classification accuracy for diagnosis of ADHD and Autism, which is an important step towards computer aided diagnosis of these psychiatric diseases.