Cognitive Abnormalities in Long COVID: Discovering Patient Subtypes using Machine Learning

Abstract

Long COVID is a complex and heterogeneous condition affecting millions worldwide, with cognitive dysfunction emerging as one of its most debilitating and poorly understood symptoms. While neuroimaging studies reveal brain alterations, they may not fully capture the underlying patterns that traditional statistical analyses often overlook. This study applies machine learning to identify neurocognitive subtypes in Long COVID using multimodal MRI data. We compared four hierarchical clustering methods, grey matter-based, white matter-based, structural (combined grey matter and white matter), and a novel multilayer network approach; using Ward’s linkage to ensure consistency. Standardized neuroimaging preprocessing and dimensionality reduction were applied, with clustering performance evaluated through internal and external validation metrics. Results identified distinct neurocognitive subtypes with varying brain structure alterations, cognitive performance, and health metrics. The multilayer network approach outperformed traditional methods, effectively capturing complex grey and white matter interactions. Significant cognitive impairments were linked to specific brain degeneration patterns in population with higher comorbidity risks and more severe symptoms during the acute infection. These findings highlight the potential of machine learning in refining the understanding of Long COVID by identifying structural markers to distinguish different patient subtypes.