Structural Alterations and Cognitive Impairment in Late-Onset Depression: A Reverse Correlation Analysis

Late-onset depression (LOD), particularly when accompanied by cognitive impairment, represents a significant risk factor for dementia. Prevailing perspectives emphasize that cognitive impairment arises from interactions among multiple brain regions. However, current approaches to identifying brain network patterns associated with cognitive impairment largely rely on group-level analyses with multiple-comparison corrections, which may obscure complex and interconnected relationships between brain regions. Our previous research demonstrated that alterations in brain network properties in patients with LOD are closely associated with cognitive function. We therefore hypothesised that aberrant interactions among multiple brain regions in LOD lead to changes in network properties and subsequent cognitive dysfunction.

This study aimed to investigate the interregional brain interactions underlying cognitive impairment in LOD by leveraging the robust interpretability of neural network models. Specifically, we sought to: (1) develop a neural network model of LOD-related cognitive impairment based on brain network properties; and (2) apply a reverse correlation approach to identify connectivity features associated with cognitive impairment in LOD.

No statistically significant differences were observed in tthe structural network properties when comparing the LOD and control participant groups across various thresholds. Using a neural network-based reverse correlation method, the most prominent differences were identified in the inferior, middle, and anterior regions of the left temporal pole when comparing patients with LOD with and without mild cognitive impairment (MCI).

Alterations in the internal structure of the temporal lobe may represent potential anatomical biomarkers for the early prediction of Alzheimer's disease, providing novel insights into the pathophysiological mechanisms underlying LOD-related MCI. The research framework proposed in this study effectively addresses the challenge of detecting subtle intergroup anatomical differences in studies with limited sample sizes. Moreover, the reverse correlation approach is not restricted to multilayer neural networks; as machine learning models become increasingly powerful and accessible, this method offers a practical and interpretable alternative for exploratory neuroimaging research.