Disentangling vulnerability, state and trait features of neurocognitive impairments in depression

Abstract

Depression is a debilitating disorder that often starts manifesting in early childhood and peaks in onset during adolescence. Neurocognitive impairments have emerged as clinically important characteristics of depression, but it remains controversial which domains specifically index pre-existing vulnerability, state-related or trait-related markers. Here, we disentangled these effects by analysing the Adolescent Brain Cognitive Development dataset (n = 4626). Using information of participants' current and past mental disorders, as well as family mental health history, we identified low-risk healthy (n = 2100), high-risk healthy (n = 2023), remitted depressed (n = 401) and currently depressed children (n = 102). Factor analysis of 11 cognitive variables was performed to elucidate latent structure and canonical correlation analyses conducted to probe regional brain volumes reliably associated with the cognitive factors. Bayesian model comparison of various a priori hypotheses differing in how low-risk healthy, high-risk healthy, remitted depressed and currently depressed children performed in various cognitive domains was performed. Factor analysis revealed three domains: language and reasoning, cognitive flexibility and memory recall. Deficits in language and reasoning ability, as well as in volumes of associated regions such as the middle temporal and superior frontal gyrus, represented state- and trait-related markers of depression but not pre-existing vulnerability. In contrast, there was no compelling evidence of impairments in other domains. These findings-although cross-sectional and specific to 9-10-year-old children-might have important clinical implications, suggesting that cognitive dysfunction may not be useful targets of preventive interventions. Depressed patients, even after remission, might also benefit from less commonly used treatments such as cognitive remediation therapy.

Keywords: affective disorders; child psychiatry; computational psychiatry; depression; imaging.

Figure 1

Group comparison of scores across the three cognitive factors. (A) For the language and reasoning factor, Bayesian model comparison found that Models 7 (LRH=HRH≠RD≠CD) and 3 (LRH=HRH≠RD=CD) performed the best, indicating the absence of any vulnerability impairment. However, more data will be required to determine whether these reflected a non-specific state/trait impairment, or unique state and trait effects. (B) In terms of cognitive flexibility, the null Model 1 (LRH=HRH=RD=CD) was comparable to Model 3 (LRH=HRH≠RD=CD) (BF = 1.01), but outperformed all other alternative models. (C) For memory recall, Models 4 (LRH=HRH=RD≠CD), 6 (LRH≠HRH=RD≠CD) and 3 (LRH=HRH≠RD=CD) were only anecdotally better than the null model (LRH=HRH=RD=CD). The fit of all other models were worse than the null hypothesis. Vertical lines indicate lower and upper bound of the 95% highest posterior density.

Figure 2

Results of canonical correlation analyses. (A) Left: As can be seen by the loadings of the cognitive factors, the first canonical variate along the cognition dimension is dominated by language and reasoning. Right: Correlation of the first brain-cognition canonical pair. (B) Illustration of loadings of regional brain volumes on the first canonical variate. The most strongly loading regions included those implicated in language processing, such as the middle temporal gyrus and inferior frontal cortex, as well as areas thought to be involved in reasoning, such as the superior frontal gyrus and superior parietal cortex. The Desikan-Killiany and automatic subcortical atlases in Freesurfer were used.

Figure 3

Group comparison of scores along the first canonical variate. (A) For structural MRI (sMRI) CV1, Bayesian model comparison found that Models 4 (LRH=HRH=RD≠CD), 7 (LRH=HRH≠RD≠CD) and 3 (LRH=HRH≠RD=CD) provided the most parsimonious account of the data. Thus, a state deficit, but no pre-existing vulnerability, is present. However, more data are required to determine if deficiencies in these areas might be a trait marker. (B) The results for cognition CV1 are similar to those for the language and reasoning factor. Models 7 (LRH=HRH≠RD≠CD) and 3 (LRH=HRH≠RD=CD) performed the best, indicating the absence of any pre-existing vulnerability, but more data will be required to determine whether these reflected a non-specific state/trait impairment, or unique state and trait effects. Models included a random effect of site. Vertical lines indicate lower and upper bound of the 95% highest posterior density.