Stress is an everyday experience and maladaptive responses play a crucial role in the etiology of affective disorders. Despite its ubiquity, the neural underpinnings of subjective stress experiences have not yet been elucidated, particularly at an individual level. In an important advance, Goldfarb et al. showed recently that subjective stress and arousal levels in response to threatening stimuli were successfully predicted based on changes in hippocampal connectivity during the task using a machine learning approach. Crucially, stress responses were predicted by interpretable hippocampal connectivity networks, shedding new light on the role of the hippocampus in regulating stress reactivity. However, the authors induced stress by displaying aversive pictures, while stress research often relies on the extensively validated Trier social stress task (TSST). The TSST incorporates crucial factors such as unpredictability of success and the social-evaluative threat of the stressor thereby eliciting cortisol responses more robustly compared to threatening images. Towards generalization, cross validation within a sample as conducted by Goldfarb et al. or independent replications are important steps, but the generalizability to different stressors allows to draw broader conclusions about the potential use of hippocampal connectivity to predict subjective stress. Arguably, translating these findings to clinical applications would require a broad generalization of the results or the prediction algorithm to psychosocial stress. Here, we assessed the predictive performance of Goldfarb et al.’s algorithm for subjective stress in an independent sample using an MR adaption of the TSST. In line with Goldfarb et al., we observed robust stress-induced changes in hippocampal connectivity. However, the spatial correlation of the changes in connectivity was low indicating little convergence across alleged stress paradigms. Critically, stress-induced changes of hippocampal connectivity were not robustly predictive of subjective stress across a multiverse of analyses based on connectivity changes. Collectively, this indicates that the generalizability of the reported stress connectivity fingerprint to other stressors is limited at best, suggesting that specific tasks might require tailored algorithms to robustly predict stress above chance levels.
bioRxiv Subject Collection: Neuroscience