A High-Level Overview of fMRI-Based Machine Learning Classification across Brain Disorders

Background/Objectives: Functional magnetic resonance imaging (fMRI) is a neuroimaging technique that characterizes dynamic neural activity in real time by measuring blood oxygen level–dependent (BOLD) signals associated with brain function. Recently, machine learning (ML) methods have been increasingly applied to fMRI data for classification tasks of brain disorders. This review summarizes recent ML-based fMRI classification studies, highlights reported performance, and identifies common limitations and future research needs. Methods: A literature search was conducted in PubMed and Web of Science for articles published from November 1, 2022, to November 1, 2025. Search terms included “machine learning”, “fMRI”, “classification”, and “brain disorders”. Studies that applied ML models to fMRI data to perform classification tasks for Alzheimer’s disease (AD), autism spectrum disorder (ASD), schizophrenia (SZ), major depressive disorder (MDD), or brain tumors were targeted. A total of 25 articles is included in this review. Results: We discovered that many studies elected to employ deep learning models, including Convolutional Neural Networks (CNNs), Graph Neural Networks (GNNs), and Graph Convolutional Networks (GCNs), capable of handling high-dimensional fMRI features to perform the classification tasks. Reported accuracy typically ranged from 70% to 95%, depending on disorder type and dataset. However, many of the models faced challenges, including limited data availability, model generalizability, and class imbalance bias. We propose several recommendations to address these challenges and to support researchers in improving the robustness of future studies. Conclusion: ML-based fMRI classification research continues to advance rapidly, showing strong potential for clinical transition. However, certain steps must be taken to address the current limitations first. Future work should focus on standardizing preprocessing, cross-site generalization, and transparent reporting. Addressing these challenges will be essential for developing clinically reliable fMRI-based ML diagnostic tools for brain disorders.