Decoding Cognitive States from fMRI Using Classical Machine Learning and Temporal Dynamics Analysis: An Interpretable Approach Using the Human Connectome Project

We propose a rigorous and reproducible methodology for analyzing functional MRI data, aimed at: (1) demonstrate their efficiency in classifying task-induced brain states with a limited amount of data, (2) present a methodology to identify brain regions critical for classification and reveal their uniqueness across different states, and (3) show, using strong mathematical methods, that the discriminative power of these regions depends not only on their spatial localization but also on their coordinated temporal activity. Through correlation and temporal structure analyses, we demonstrated that top-ranked regions exhibit stronger, more structured, and richer dependencies than low-ranked regions, underscoring the critical role of temporal dynamics in shaping distinct cognitive brain states. Our work addresses the need for a transparent, accessible, and interpretable framework for studying cognitive processes through neuroimaging data. We analyzed fMRI data from 587 healthy participants from the Human Connectome Project across seven cognitive tasks. Finally, we perform a detailed analysis of the identified brain regions to support further neuroscientific interpretation and discussion. Key PointsO_LIClassical machine learning methods effectively classify task-induced brain states from fMRI data with high accuracy (up to 99% for some tasks), demonstrating that simple, interpretable algorithms can successfully decode complex neuroimaging data without requiring advanced deep learning approaches. C_LIO_LIHigh-accuracy brain states require relatively few significant regions suggesting focal neural signatures, while lower-accuracy states involve more distributed activations across multiple brain areas, revealing different levels of neural organization complexity underlying various cognitive processes. C_LIO_LIThe identified brain regions align with established neuroscientific knowledge, with motor tasks activating contralateral sensorimotor areas, language processing engaging left-hemisphere networks, and social cognition recruiting visual motion processing regions, validating the neurobiological relevance of our machine learning approach. C_LIO_LIRigorous mathematical analyses of temporal dynamics demonstrated that the discriminative power of significant brain regions depends not only on spatial localization but also on their coordinated temporal activity. Correlation, temporal structure analyses consistently showed that top-ranked regions exhibit stronger, more structured, and richer dependencies than low-ranked regions, underscoring the critical role of temporal dynamics in shaping distinct cognitive brain states. C_LI