Subjects

This study is based on samples of five independent studies that implemented the ScanSTRESS protocol on the same scanner with identical parameters and used the same assay for cortisol determination across all studies (Bärtl et al., Reference Bärtl, Henze, Peter, Giglberger, Bohmann, Speicher and Kudielka2024; Giglberger et al., Reference Giglberger, Peter, Henze, Kraus, Bärtl, Konzok and Wüst2023; Henze et al., Reference Henze, Konzok, Kreuzpointner, Bärtl, Peter, Giglberger and Wüst2020; Konzok et al., Reference Konzok, Henze, Peter, Giglberger, Bärtl, Massau and Kudielka2021; Speicher et al., Reference Speicher, Henze, Bärtl, Böhm, Sommer, Wüst and Kudielka2023). From the Regensburg Burnout Project sample, only healthy subjects were included (Bärtl et al., Reference Bärtl, Henze, Peter, Giglberger, Bohmann, Speicher and Kudielka2024). Additionally, we included data exclusively from subjects who underwent the original version of ScanSTRESS (Speicher et al., Reference Speicher, Henze, Bärtl, Böhm, Sommer, Wüst and Kudielka2023). In total, $ 291 $ subjects $ \left(18-62\hskip0.42em \mathrm{years},\mathrm{mean}\hskip0.52em \mathrm{age}\hskip0.52em 26.16\pm 8.95\right) $ were analyzed, comprising 134 males $ \left(18-60\hskip0.5em \mathrm{years},\mathrm{mean}\hskip0.5em \mathrm{age}\hskip0.5em 26.65\pm 8.23\right) $ and 157 females $ \left(18-62\hskip0.42em \mathrm{years},\mathrm{mean}\hskip0.42em \mathrm{age}\hskip0.42em 25.74\pm 9.53\right) $ . Of those female subjects, $ 81 $ were tested during their luteal phase of the menstrual cycle, $ 64 $ were taking hormonal contraceptives, and $ 12 $ were post-menopausal.

Subjects were recruited via flyers and social media platforms. General exclusion criteria across all studies comprised the following: self-reported history of one or more current psychiatric, neurological, or endocrine disorders; treatment with psychotropic drugs or other medications affecting central nervous system or endocrine function; daily alcohol consumption; incompatibilities with MRI (e.g. metal parts, pregnancy); and regular night shifts. Study-specific exclusion criteria are detailed in the original research reports. All subjects provided written informed consent and received financial compensation for their participation. All studies were approved by the local ethics committee of the University of Regensburg.

General procedure

The test protocol sequence, an overview of ScanSTRESS, and its block design are shown in Figure 1. Test sessions took place between 1 and 6 pm to minimize the influence of the circadian rhythm of cortisol secretion (Kudielka, Schommer, Hellhammer, & Kirschbaum, Reference Kudielka, Schommer, Hellhammer and Kirschbaum2004; Zänkert, Bellingrath, Wüst, & Kudielka, Reference Zänkert, Bellingrath, Wüst and Kudielka2019). Subjects arrived at the laboratory 75 minutes before the stress induction followed by a relaxation phase after the first saliva sample was taken to measure cortisol. A total of ten saliva samples were collected $ \Big(-75,-15,-1,+15,+30,+50,+65,+80,+95,\mathrm{and}+110\hskip0.5em \min $ ) using ‘cortisol salivettes’ (Sarstedt, Nuembrecht, Germany); three before, one during, and further six after ScanSTRESS, with sample −1 referring to stress onset. The saliva samples were analyzed using a time-resolved fluorescence immunoassay with fluorometric endpoint detection (dissociation-enhanced lanthanide fluorescence immunoassay, DELFIA, Dressendörfer et al., Reference Dressendörfer, Kirschbaum, Rohde, Stahl and Strasburger1992) with an intra-assay coefficient of variation between $ 4.0\%\hskip0.5em \mathrm{and}\hskip0.5em 6.7\% $ and an inter-assay coefficient of variation between $ 7.1\%\hskip0.5em \mathrm{and}\hskip0.5em 9.0\% $ .

Figure 1. Experimental protocol. Illustration of ScanSTRESS and its block design, repeated collection of salivary cortisol, glucose administration, and acquisition of ScanSTRESS-responses as well as T1-weighted (T1_w-) images.

The ScanSTRESS paradigm is a block-designed protocol with alternating stress and control blocks presented in two runs. During control blocks, subjects perform simple figure and number matching tasks, while stress blocks involve rotation and arithmetic tasks under the supervision of a panel in professional attire, under time pressure, and with constant negative feedback. We used an improved version of the original protocol (Streit et al., Reference Streit, Haddad, Paul, Frank, Schafer, Nikitopoulos and Wüst2014) across all included studies to enhance the stress-induced responses without changing the paradigm (Henze et al., Reference Henze, Konzok, Kreuzpointner, Bärtl, Peter, Giglberger and Wüst2020). This improvement included: First, a 45-min relaxation phase was introduced to ensure low baseline cortisol levels, along with a detailed explanation of the scanning procedure to reduce pre-scan anxiety (McGlynn, Smitherman, Hammel, & Lazarte, Reference McGlynn, Smitherman, Hammel and Lazarte2007; Thorpe, Salkovskis, & Dittner, Reference Thorpe, Salkovskis and Dittner2008). Second, subjects received a sugary drink (75 g glucose in 200 ml herbal tea) to boost cortisol reactivity (Zänkert, Kudielka, & Wüst, Reference Zänkert, Kudielka and Wüst2020). Third, the interval between the end of the relaxation period and the onset of the stress task inside the MRI scanner was kept below 10 min, ensuring a more abrupt transition from relaxation to stress exposure (Henze et al., Reference Henze, Konzok, Kreuzpointner, Bärtl, Peter, Giglberger and Wüst2020). After ScanSTRESS exposure, study-specific acquisitions were conducted (i.e. a resting state sequence in all studies except one in which a moral decision paradigm was conducted, Speicher et al., Reference Speicher, Henze, Bärtl, Böhm, Sommer, Wüst and Kudielka2023), followed by an anatomical assessment for all subjects.

Data analysis

To determine potential relationships between acute cortisol stress responses and neural structural measures, we preprocessed both variables as follows. A cortisol increase was defined as the difference between the peak cortisol level $ \left(\mathrm{sample}\hskip0.52em +30,\hskip0.52em ,+50,\hskip0.52em ,+65\right) $ and the pre-stress cortisol level $ \left(\mathrm{sample}-1\right) $ , as marked in bold in Figure 1. We focused on cortisol increase as it captures acute stress exposure and HPA axis reactivity-latency (Dickerson & Kemeny, Reference Dickerson and Kemeny2004), while area under the curve measures includes pre-stress and recovery phases, reflecting total cortisol release.

Neuroimaging data were collected using a Siemens MAGNETOM Prisma $ 3T $ MRI (Siemens Healthcare, Erlangen, Germany) with a 64-channel head coil. Two series of planar BOLD echo sequences (ScanSTRESS), multiband resting state (results to be published), and anatomical sequences were performed. Anatomical measurements included a 3D structural T1_w-image $ \hskip0.5em \left(\mathrm{T}1/\mathrm{TR}/\mathrm{TE}=1200/2400/2.18\hskip0.42em \mathrm{ms},\mathrm{flip}\ \mathrm{angle}=8,\mathrm{distance}\ \mathrm{factor}=50\%\right) $ . Anatomical images were initially segmented using SynthSeg, a convolutional neural network-based technique robust to variations in contrast and resolution (Billot et al., Reference Billot, Greve, Puonti, Thielscher, Van Leemput, Fischl and Iglesias2023). Further preprocessing was conducted using ENIGMA HALFpipe (Waller et al., Reference Waller, Erk, Pozzi, Toenders, Haswell, Büttner and Veer2022), which integrates fMRIprep, FSL, and FreeSurfer. FreeSurfer was used for cortical parcellation and subcortical segmentation of the T1_w-volume, including estimation of TBV (‘BrainSegVolNotVent’). Preprocessing steps included non-brain tissue removal, automatic Talairach transformation, segmentation, intensity normalization, gray/white matter boundary tessellation, topology correction, and surface shaping. Additionally, the structural output of each subject was visually inspected before extracting subcortical volume, cortical thickness, and surface area measurements of pre-limbic structures for each hemisphere using asegstats2table and aparcstats2table.

In confirmatory analysis, PALM (Winkler, Ridgway, Webster, Smith, & Nichols, Reference Winkler, Ridgway, Webster, Smith and Nichols2014) was used to investigate potential relationships between individual cortisol stress responses and structural measures of the limbic system. As described in our previous study (Henze et al., Reference Henze, Konzok, Kudielka, Wüst, Nichols and Kreuzpointner2023), the following ROIs were selected based on the current literature on the interaction between neural substrates and acute HPA axis responses (Harrewijn et al., Reference Harrewijn, Vidal-Ribas, Clore-Gronenborn, Jackson, Pisano, Pine and Stringaris2020; Henze et al., Reference Henze, Giglberger, Bärtl, Konzok, Neidhart, Krause, Serin, Waller, Peter, Kreuzpointner, Speicher, Streit, Veer, Kirsch, Nichols, Kudielka, Wüst, Erk and Walter2025; Van Oort et al., Reference Van Oort, Tendolkar, Hermans, Mulders, Beckmann, Schene and van Eijndhoven2017): volumes were extracted for the thalamus, striatum (ncl. caudatus, ncl. accumbens, and putamen), hippocampus, and amygdala. Cortical thickness and surface area were measured for the cingulate cortex (CC: rostral anterior cingulate cortex, rACC; caudal anterior cingulate cortex, cACC; and posterior cingulate cortex, PCC), parahippocampus, the orbitofrontal cortex (OFC: lateral OFC, lOFC; and medial OFC, mOFC), insula, and precuneus. Note that, we use the terminology based on Destrieux- and Desikan-Killiany atlases (Desikan et al., Reference Desikan, Ségonne, Fischl, Quinn, Dickerson, Blacker and Killiany2006; Destrieux, Fischl, Dale, & Halgren, Reference Destrieux, Fischl, Dale and Halgren2010). However, in other literature (Shackman et al., Reference Shackman, Salomons, Slagter, Fox, Winter and Davidson2011), the rACC is often split into subgenual and pACC, the cACC is called anterior mid-CC (aMCC) or commonly dACC, and the PCC is referred to as posterior mid-CC (pMCC). Six models were tested, where lateralized structural measures (volume, thickness, and surface area) of the left and right hemispheres were separately used as dependent variables, given the assumption that HPA axis regulation may be lateralized (Cerqueira, Almeida, & Sousa, Reference Cerqueira, Almeida and Sousa2008) and following our preliminary results (Henze et al., Reference Henze, Konzok, Kudielka, Wüst, Nichols and Kreuzpointner2023). Since sex, age, and TBV can influence structural brain measures (Giedd et al., Reference Giedd, Blumenthal, Jeffries, Castellanos, Liu, Zijdenbos and Rapoport1999; Moog et al., Reference Moog, Nolvi, Kleih, Styner, Gilmore, Rasmussen and Buss2021), these variables were included as covariates, with TBV considered only in volumetric models. In addition, menstrual cycle phase, hormonal contraceptive use, and reproductive status also influence brain structure (Brønnick, Økland, Graugaard, & Brønnick, Reference Brønnick, Økland, Graugaard and Brønnick2020; B. Pletzer, Winkler-Crepaz, & Hillerer, Reference Pletzer, Winkler-Crepaz and Hillerer2023; Song et al., Reference Song, Patton, Friedman, Mahajan, Nordlicht, Sayed and Lipton2023). Therefore, each of the six models with lateralized structural measures as dependent variables included cycle (male, luteal phase, contraception, post-menopause; dummy-coded with male as reference), age (continuous), and TBV (continuous, where appropriate) as control variables, along with individual (sex-specific) cortisol increases (continuous) as predictors. This grouping variable was implemented as follows, for instance, male cortisol increase was coded to contain the actual cortisol increase value for male subjects and 0 for female subjects; the same procedure was applied for female cortisol increase, but in a reverse order. This approach allowed us to correct for and investigate the possible influence of sex-specific cortisol responses. To test for the general relationship between cortisol increase and brain structure (independent of sex), the influence of both variables, male and female cortisol increase, was tested simultaneously (see Supplementary Methods for tested statistical model, Henze et al., Reference Henze, Konzok, Kudielka, Wüst, Nichols and Kreuzpointner2023). We did not create a variable to differentiate cortisol increases based on the cycle categories, even though such differences were reported (Zänkert et al., Reference Zänkert, Bellingrath, Wüst and Kudielka2019), because although there was significant variation among groups $ F\left(3,287\right)=13.70,p<.000\Big) $ , this was driven by significant differences between males compared to two female categories (luteal phase, $ p<.000 $ ; contraceptive users, $ p<.000\Big) $ , but not among the female categories $ \left(p>.637\right) $ (Supplementary Table S1). Multiple comparisons were corrected for each modality (e.g. left hemisphere cortical thickness) across all ROIs using the False Discovery Rate (FDR) -fdr option.

In the first stage of exploratory analyses, FreeSurfer’s (version 7.4.1) vertex-wise general linear model (GLM)-based group analyses were performed to investigate the relationship between sex-specific cortisol responses and structural measures of the entire cortex without prior assumptions about regions. Subjects’ cortical surface was first mapped onto the fsaverage template, smoothed with a $ 10\hskip0.5em \mathrm{mm} $ full width at half-maximum kernel and concatenated into a single data set for each hemisphere (mris_preproc). In the GLM (mri_glmfit, Different Offset, Different Slope), volume, thickness, and surface area were dependent variables, while sex-specific cortisol increase was the predictor, and the confounding effects of cycle and age were controlled. As in PALM analyses, sex-independent and sex-specific cortisol increase were evaluated separately using corresponding contrast matrices in vertex-based analyses. Monte Carlo Simulation cluster analyses at a significance level of $ p=.050 $ were used to correct for multiple comparisons (mri_glmfit-sim –2spaces). The cluster-forming threshold was set to 3.0 to prevent inflated false positives (Greve & Fischl, Reference Greve and Fischl2018).

As further exploratory analyses, a ML approach, ridge regression, was employed to predict subjects’ cortisol stress response from their volume, thickness, and surface area measures (158 features in total). Although the GLM is a standard method for localizing the relationship between cortisol increase and structural brain measures, evaluating the generalizability of these associations is not straightforward. However, ML inherently enables out-of-sample evaluation, which is essential for evaluating generalizability (Kriegeskorte, Simmons, Bellgowan, & Baker, Reference Kriegeskorte, Simmons, Bellgowan and Baker2009). Additionally, as the generative process is reversed in ML compared to GLM, it allows for the extraction of complex multivariate patterns in brain measures to predict cortisol response, thereby enabling the identification of the physiological stress state at the subject level. Briefly, ridge regression was fit and evaluated within a ten-repeated nested cross-validation framework: ten outer folds for model evaluation and five inner folds for hyperparameter optimization. Ridge regression is a type of multivariate linear regression that incorporates regularization, which shrinks the coefficients of uninformative features, thereby making it more robust to multicollinearity (Hoerl & Kennard, Reference Hoerl and Kennard1970). Cross-validation is a standard approach for evaluating the predictive performance of a model by splitting the data into training and test sets recursively. As cross-validation is a stochastic approach (i.e. the data are split randomly), we repeated the cross-validation ten times to reduce arbitrariness (Varoquaux et al., Reference Varoquaux, Raamana, Engemann, Hoyos-Idrobo, Schwartz and Thirion2017). Additionally, the potential confounding effect of cycle was corrected using a cross-validated confound-correction technique to avoid any possible data leakage (Snoek, Miletić, & Scholte, Reference Snoek, Miletić and Scholte2019). Age was not corrected as it was not found to be associated with cortisol increase $ \left(F\left(289,1\right)=.819,\hskip0.5em p=.366\right) $ . The model performance was evaluated using correlation coefficient (r) and tested against a null distribution using permutation testing (Ojala & Garriga, Reference Ojala and Garriga2009) with 1000 permutations for significance. Shapley additive explanations (SHAP) were used to evaluate the contributions of brain measures to the prediction performance (Lundberg & Lee, Reference Lundberg and Lee2017). Details of model interpretation are given in Supplementary Methods.