Pre-Infection Mental Health, but Not Brain Volumetry, Predicts Risk of Post-COVID Condition: A Population-Based Cohort Study in the German National Cohort (NAKO)

Abstract

Background: Post-COVID Condition (PCC) affected 10-30% of individuals after SARS-CoV-2 infection, and pre-infection predictors of who goes on to develop it are not well established. We compared pre-infection structural brain variation and baseline psychiatric phenotype as candidate predictors of PCC within a common multi-modal framework in the German National Cohort (NAKO). Both are candidate "first hits" under the second-hit hypothesis, which motivates the comparison; the vulnerability-by-infection interaction that hypothesis turns on is not identifiable among infected participants, so this is a prediction study rather than a test of the hypothesis. Methods: In 8,464 SARS-CoV-2-infected NAKO neuroimaging participants, of whom 2,304 (27.2%) met PCC criteria (weighted post-COVID syndrome (PCS) score > 10.75) and the rest were symptom-free controls, ten pre-infection modalities (T1-weighted volumetric brain MRI across six parcellations, baseline mental health (PHQ-9, GAD-7), demographics, and seven further biomedical and socioeconomic domains) entered a stacked ensemble with 10x5 nested cross-validation (primary metric: the area under the precision-recall curve, PR-AUC). Pre-specified analyses assessed transportability to a non-imaging cohort and robustness to symptom trajectory and unmeasured confounding. SARS-CoV-2 infection status and the PCC outcome were ascertained by self-report; no clinically confirmed diagnoses were available. Results: Baseline mental health, assessed 3-8 years before infection, was the strongest predictor (38.2% of feature importance; standalone ROC-AUC = 0.640), whereas all six brain-MRI parcellations performed at or near chance (ROC-AUC 0.496-0.516). The multi-modal model achieved moderate, well-calibrated discrimination (ROC-AUC = 0.664, 95% CI 0.652-0.677; PR-AUC = 0.413, 0.392-0.434; ECE = 0.015). Applied without retraining to the non-imaging cohort it retained discrimination (ROC-AUC = 0.660, 0.654-0.665), meeting two of three pre-specified equivalence criteria. Adjusting for the full symptom trajectory shrank the baseline mental-health odds ratio from 2.06 (1.83-2.32) to a conservative lower bound of 1.38 (1.20-1.58; E-value 2.66) while leaving it independently significant, and was essentially unchanged under an alternative control definition (2.07). Baseline mental health did not predict objectively measured hyposmia in participants screened before their infection (0.89) while predicting self-reported smell loss in the same participants (1.92). Conclusions: Baseline mental health years before infection is the strongest pre-infection predictor of PCC among the infected, and the association held across the pre-specified sensitivity analyses. Whether it acts specifically on COVID-19 sequelae is a separate question this design cannot answer, and the indirect evidence points away from specificity: the association is undiminished after mild infection but absent among the hospitalised, and baseline mental health predicts current symptom load no more strongly in infected than in non-infected participants. The volumetric structural candidate is not supported: pre-pandemic T1-weighted volumetry carried no predictive signal in this single neuroimaging cohort, consistent with COVID-19-associated brain changes being acute-onset rather than pre-existing. Risk stratification may benefit from incorporating baseline psychiatric phenotype; whether treating it reduces PCC incidence requires interventional study.