Cognitive Performance and Brain-Predicted Age Difference in Bipolar Disorder

Abstract

Neuroimaging-derived brain-predicted age difference (brain-PAD) is a promising marker of advanced brain aging, but its link to cognitive function in bipolar disorder (BD) is not well understood, especially when comparing across publicly available algorithms trained on diverse, large sample datasets and to algorithms trained on local cohorts with rich multimodal imaging data. Our study compares algorithms used to estimate brain-PAD in terms of their clinical relevance to cognition in BD. We included 44 euthymic BD I individuals and 73 HCs who completed the Delis-Kaplan Executive Function System, and we selected nine scores from this battery for further analyses. Raw scores were log-transformed, scaled, and subjected to PCA; PC1 indexed overall executive function. Four brain-PAD algorithms (PHOTON, BrainageR, DenseNet, Multimodal) were applied to T1-weighted MRI data; the multimodal algorithm also included Diffussion Tensor Imaging (DTI), Arterial Spin Labeling (ASL), functional Magnetic resonance imaging (fMRI) and resting state Magnetic resonance imaging (rsMRI) data. For each algorithm, we regressed brain-PAD on age, sex, and their interaction to obtain residuals, then used those residualized brain-PADs (which we refer to subsequently as brain-PADs throughout the text) to predict PC1. We then directly assessed if there were group differences in the relationship of brain-PAD to cognitive function by including an interaction term between group x brain-PAD. We found no significant group x brain-PAD interaction across all four algorithms. Given that, we then combined BD and HC and explored whether brain-PAD was a meaningful predictor of cognitive performance. Multimodal brain-PAD emerged as a strong negative predictor of cognitive performance (Beta Estimate = -0.084, SE = 0.024, t = -3.50, p < 0.001), indicating that those with older-appearing brains, as indexed by the brain-PAD, scored lower on PC1. BrainageR brain-PAD also significantly predicted PC1 (Beta Estimate = -0.031, SE = 0.0116, t = -2.71, p < 0.01), and DenseNet brain-PAD showed a modest effect (Beta Estimate = -0.0355, SE = 0.0177, t = -2.00, p < 0.05). PHOTON brain-PAD demonstrated a negative trend with PC1 (Beta Estimate = -0.024, SE = 0.0127, t = -1.92, p = 0.06). Residualized brain-PAD, after accounting for age and sex, was inversely associated with a composite metric of executive functioning, particularly for an algorithm integrating a range of imaging modalities. Our findings demonstrate how brain aging patterns captured by a neuroimaging-based, ML-derived composite metric could be associated with cognitive performance across algorithms trained on a variety of data granularity and sample sizes.