Classification of neural disorders from fMRI data using machine learning

Abstract

Attention Deficit/Hyperactivity Disorder (ADHD) is a neural disorder prevalent in 5-8% of children worldwide. The traditional procedure followed to diagnose children with this disorder is tedious and lacks homogeneity and objectivity. Based on the hypothesis that children with ADHD may have neuroanatomical and cognitive differences from neurotypical children, this study aimed at taking advantage of these differences to identify functional connectivity patterns unique to children with ADHD. Functional Magnetic Resonance Imaging (fMRI) is a neuroimaging technique used to analyse functional connectivity by obtaining brain scans over a period of time during which a subject must complete a simple task. An fMRI dataset was sourced, visualized and pre-processed. Three Machine Learning (ML) models were developed and trained on a subset of the pre-processed fMRI dataset. Their performance was then evaluated on an unseen testing set, to evaluate whether the models succeeded in accurately classifying children with ADHD and controls. The techniques used during the development included Principal-Component Analysis (PCA), grid searches and cross-validation. Three ML models were developed: Support Vector Machines (SVM), a Multi-Layer Perceptron (MLP) and a Long Short-Term Memory (LSTM) network. The results indicated that the models learned well and were able to generalize. When evaluated on unseen data, the model accuracies range between 60-70%. These results are in line with other studies reviewed in literature, and indicate acceptable performance from ML models trained on small datasets. Further developments can be made on this study to improve the results. These include testing on a larger dataset, looking into different parcellation atlases, and attempting alternative ML techniques such as transfer learning.