Neural signatures of human psychological resilience driven by acute stress

Significance

Human resilience to stress extends beyond basic physiological responses, engaging higher-order processes such as self-confidence and positive attitudes. This study demonstrates that neural markers of resilience emerge more than 60 min after acute stress, indicating a delayed yet distinct recovery phase. Functional magnetic resonance imaging and electroencephalogram analyses reveal that resilient individuals show increased activity in the default mode network and posterior hippocampus; whereas less resilient individuals exhibit stronger salience network connectivity and elevated high-beta and gamma oscillations. Our findings suggest that human-specific resilience depends on unique, time-lagged neural dynamics rather than immediate stress responses. Identifying the patterns thereof offers promising avenues for interventions aimed at enhancing recovery from stress-related disorders, including posttraumatic stress disorder.

Abstract

Neurophysiological mechanisms underlying psychological resilience—the ability to overcome adversity—have been extensively studied in animals. However, compared to that in animals, human resilience is unique in that it is underpinned by higher-order cognitive functions, such as self-confidence, tenacity, and a positive attitude to challenges. Given these discrepancies, the neurophysiological mechanisms underlying human-specific resilience remain unclear. To address this issue, we aimed to record multimodal responses after acute stress exposure over 1.5 h using functional brain imaging and peripheral physiological measurements. We showed that the degree of individual resilience is indexed by multiple changes in neural dynamics 1 h after acute stress. Functional magnetic resonance imaging and electroencephalography show that activity in the cortical salience network and power in high-beta and gamma oscillations increase in less resilient individuals. Contrastingly, activity in the cortical default mode network and spontaneous activity in the posterior hippocampus increase in more resilient individuals. Machine learning analysis confirmed that, 1 h after stress exposure, the functional connectivity in the salience network was the most influential, followed by that in the default mode network, gamma power, high-beta power, and hippocampal activity. The neurophysiological dynamics for resilience do not occur as previously thought, but rather in a time-lagged manner against stress exposure. Our findings shed light on an approach to recovery from stress-induced deficits such as delayed neuromodulation after a stressful event.

Psychological resilience is the ability to adapt positively, expressed through dynamic neural and behavioral processes of coping and recovery in the face of acute stress, trauma, or chronic adversity (1, 2). Despite the critical role of psychological resilience in mental health, this ability exhibits marked interindividual variability (Fig. 1A). Those with lower resilience are at high risk of depression and other mental illnesses (3). Accumulating evidence has demonstrated that acute stress initiates successive responses in the brain and body, similar to waves (Fig. 1B), including rapid autonomic (4–13), time-lagged (peaking 20 min after stress exposure) hormonal (14–20), and immune and slow-corticosteroid responses (peaking more than 1 h after stress exposure) (21–26). These homeostatic responses may contribute to the initiation of the processes underlying resilience.

Fig. 1.

Hypothesis and experimental design. (A) Diversity of psychological resilience. Individuals with low adaptability to stressful environments are referred to as susceptible (SUS), whereas those with high adaptability are termed resilient (RES). (B) Three stress-related neurophysiological responses to acute stress identified in previous studies. The first wave represents a transient autonomic nervous system response during the stress experience, such as pupil dilation, increased heart and respiration (Resp) rates, and changes in brain state measured using functional magnetic resonance imaging (fMRI) and electroencephalography (EEG). The second wave is characterized by a hormone response, including increased salivary cortisol and brain state changes peaking 20 min after the stress event. The third wave corresponds to the immune and slow-cortisol response, lasting more than 1 h and affecting brain activity. (C) The two hypothesized models for resilience-related neurophysiological responses triggered by acute stress exposure. In the Upper model (human-specific resilience) the neurophysiological dynamics underlying resilience are presumed to correlate with subjective assessments of resilience, rather than with the magnitude of stress-related physiological responses. In the Lower model (conserved stress resilience) the neurophysiological dynamics are influenced by both the intensity of the physiological stress response (as illustrated in B) and the level of resilience reported through subjective assessments. (D) The current experiment examined acute stress-related neurophysiological responses over 1 h. Participants first completed resilience-related questionnaires. Six types of neurophysiological responses (fMRI, EEG, pupil response, heart and Resp rates, and salivary cortisol) and subjective stress ratings (Subj) were collected. After collecting prestress neurophysiological activity in the first session, all participants underwent a cold pressor test (CPT) as an acute stressor. The second session was conducted immediately after the stress, and the third session overlapped with the expected peak of salivary cortisol elevation. The fourth and fifth sessions were conducted 60 and 80 min after the exposure, respectively.

The neurophysiological aspects of resilience have been extensively studied using animal models, in which individual animals are typically classified as either SUS or RES based on their depressive behavior in response to stress (27–31). In contrast, human resilience is a more multifaceted construct; the mere observation of depressive symptoms poststress is insufficient for the comprehensive assessment thereof. Evaluating human resilience requires consideration of other factors, such as prior experiences and mental processes involving higher-order cognitive functions, including self-confidence, tenacity, and positive acceptance of challenges faced (3, 32).

Subjective questionnaires that assess individual levels of resilience offer a valuable means of capturing human-specific adaptability toward adversity. Accordingly, while certain components of resilience may be shared across species, it remains unclear how higher-order cognitive resilience in humans is represented in neurophysiological signatures.

In traditional animal studies, signatures of resilience have been documented as components of well-established stress-related neurophysiological responses (1, 31). These stress responses may also reflect resilience in humans. In this case, physiological stress sensitivity would be expected to correlate with the level of resilience assessed by questionnaires (conserved stress resilience; Fig. 1 C, Lower model). Alternatively, higher-order cognitive resilience in humans may involve distinct brain dynamics that are independent of known physiological stress responses (human-specific resilience; Fig. 1 C, Upper model). Under this framework, resilience-related responses triggered by stress exposure correlate with subjective resilience assessments, but not with the known physiological stress response. This latter type of response may have been overlooked in prior research, as it is not accessible through studies involving nonhuman animals.

Therefore, to examine these two types of resilience, we aimed to evaluate neurophysiological responses to acute stress using a multimodal approach including peripheral and neuroimaging measurements, and to explore the relationship between these responses and subjective assessments of psychological resilience.

Results

We collected multiple noninvasive brain and peripheral responses before and after acute stress exposure (Fig. 1D) from 102 participants (77 males and 25 females). Individual resilience levels were measured using the Connor-Davidson Resilience Scale (CD-RISC) (32), which is a standard approach for assessing levels of human resilience (SI Appendix, Fig. S1 and Texts S1). All participants underwent five brain imaging sessions and eight inter-scan-interval (ISI) sessions. At each imaging session, resting-state brain activity was recorded using fMRI and EEG to assess spatially and temporally resolved neural responses. Furthermore, measurements of pupil size (Pup), electrocardiogram (ECG), and Resp were recorded. Each ISI included saliva samples for cortisol density (Cort) and Subj. In addition to Pup, ECG, and Resp, a modified version of the CPT was administered as an acute stressor after the first brain imaging (baseline) session. All imaging data were evaluated as changes relative to the prestress baseline and labeled Ses2 to Ses5.

Unsupported Conserved Stress Resilience.

Our experiment successfully induced acute stress-related peripheral responses in the Pup, ECG, Cort, and Subj (SI Appendix, Fig. S2). Stress-related neural activity was also increased in resting-state fMRI [stress-driven functional connectivity (SFC) and stress-related amygdala-insula connection (ScAI)] and resting-state EEG (SI Appendix, Figs. S3 A–D, S4 A–C and Texts S2). To validate the conserved stress resilience (Fig. 1 C, Lower model), correlations between the CD-RISC scores and stress-related neurophysiological changes from baseline were verified. We found no significant correlations in any peripheral responses throughout the experiments (SI Appendix, Fig. S2, Kendall’s correlation test, τ = −0.147 to 0.127; P_FDR = 0.322 to 0.987) and no correlations in brain responses (SI Appendix, Figs. S3 E and F and S4 D and E, Pearson’s correlation test, r = −0.143 to 0.165, P_FDR = 0.121 to 0.958). These results indicated that our data did not support the conserved stress resilience model. We further confirmed the strength of these null results using the Bayes factor (BF₀₁) for each correlation analysis. With regard to the correlation between CD-RISC scores and peripheral responses, 46 of 47 correlations supported the null hypothesis (median BF₀₁: 6.183, interquartile range of BF₀₁: 5.292 to 6.741). Regarding the correlation between CD-RISC scores and brain responses, all eight correlations supported the null hypothesis (median BF₀₁: 4.609, interquartile range of BF₀₁: 3.961 to 4.766).

Spatial Features Support Human-Specific Resilience.

We explored whole-brain activities associated with individual differences in resilience (Fig. 1 C, Upper model). First, we evaluated the functional connectivity (FC) network changes that correlated with individual CD-RISC scores in four poststress sessions (Ses2 to Ses5). The number of resilience-related functional connections (RFC) was increased from Ses2 to Ses4 and then decreased at Ses5 (Fig. 2 A and B and SI Appendix, Fig. S5 for an evaluation of the global significance of the correlation). In Ses4, the two networks were significantly correlated with the CD-RISC scores (Fig. 2C; cluster-level P_FDR < 0.05), with one network showing negative correlation (RFCn: F test, F_(3,84) = 9.25, cluster-level P_FDR = 0.041, η² = 0.388) and the other a positive correlation (RFCp: F_(3,84) = 8.50, cluster-level P_FDR = 0.046, η² = 0.233) with the resilience score. The RFCn includes the bilateral insula, bilateral dorsal anterior cingulate cortex (dACC), right superior medial frontal cortex (SMF), and right middle frontal gyrus (MFG), which overlap with the cortical salience network (SaN) (33). RFCp includes the right inferior frontal gyrus (IFG), bilateral posterior cingulate cortex (PCC), and precuneus (Prec), which overlap with the cortical default mode network (DMN) (33) (SI Appendix, Fig. S6). The strength of correlations in RFCn and RFCp gradually increased from Ses2 to Ses4 and attenuated in Ses5 (Fig. 2D): the coefficient in Ses4 was significantly higher than that in Ses2 (test for the difference between correlation coefficients: T₍₈₇₎ = −2.787, P_FDR = 0.020, d = −0.597) and Ses5 (T₍₈₇₎ = 3.040, P_FDR = 0.019, d = 0.652) in RFCp. Similar tendencies were also observed in RFCn, but did not reach statistical significance (test for the difference between correlation coefficients: Ses2 and Ses4: T₍₈₇₎ = 2.095, P_FDR = 0.10, d = 0.219; Ses4 and Ses5: T₍₈₇₎ = −2.486, P_FDR = 0.089, d = −0.534) [see SI Appendix, Fig. S7 for another resilience-related functional connections, identified in Ses2 (RFCs2)]. We also tested the correlation between individual CD-RISC scores and the mean FC strength (averaged across Ses2 to Ses5 relative to baseline). No statistically significant clusters were detected at the same cluster threshold (F test, cluster level P_FDR < 0.05). These findings suggest that the RFCn and RFCp patterns are specific to Ses4.

Fig. 2.

Correlated FC in fMRI with resilience peaked 60 min poststress exposure. (A) The connectome ring depicts the number of CD-RISC-related FC from Ses2 to Ses5. The statistical threshold for the ring was set at P < 0.05, uncorrected, for display purposes. The number of FCs increased as a function of time, peaking at Ses4 (approximately 60 to 70 min after acute stress) and then decreasing at Ses5 (approximately 75 to 85 min after acute stress). (B) The number of significant FCs was greatest in Ses4, regardless of the statistical thresholds (both uncorrected and corrected). (C) Two significant FCs were identified at Ses4 with a cluster-level threshold P_FDR of <0.05. FC negatively correlated with CD-RISC scores (RFCn; shown as blue lines) in the bilateral insula, bilateral dACC, right SMF, and right MFG. Another FC positively correlated with CD-RISC scores (RFCp; shown in red lines), including the right IFG, bilateral PCC and Prec. The specific ROI names and locations in the brain are shown in SI Appendix, Fig. S6. (D) Temporal dynamics of the correlation between CD-RISC and FCs in RFCn and RFCp. The correlation coefficient (r) of RFCp was significantly higher in Ses4 than in Ses2 and Ses5. *P_FDR < 0.05, corrected for the test of equality of two dependent correlations. Insets, scatter plots of individual CD-RISC scores and FC strengths in RFCn (r = −0.517, P_FDR = 1.84E−7) and RFCp (r = 0.447, P_FDR = 1.01E−5) at Ses4. (E) Group-wise analysis of temporal dynamics of FC strength in RFCn and RFCp. Participants were classified as RES or SUS based on the median CD-RISC score. ^†P_FDR < 0.05 for one-sample t tests against zero.

Next, to determine whether this overall trend was primarily driven by the more RES or the more SUS participants, we split the sample at the median CD-RISC score (60 points) and examined group-level trajectories in brain dynamics (Fig. 2E). This analysis revealed that the correlation patterns were largely driven by those in the RES group. For RFCn (Fig. 2 E, Left), FC strength in the RES group progressively decreased from Ses2 to Ses4. At Ses4, FC was significantly reduced relative to baseline (one-sample t test against zero, RES, T₍₄₃₎ = −3.817, P_FDR = 0.003, d = 1.164) and tended to return toward baseline at Ses5 (T₍₄₃₎ = −2.232, P_FDR = 0.078, d = 0.680). Such dynamics were not observed in the SUS group (T₍₄₅₎ = 1.331 to 1.628, P_FDR = 0.176 to 0.190, d = 0.398 to 0.486). For RFCp (Fig. 2 E, Right), the RES group exhibited a similar temporal profile, albeit with smaller effects (one-sample t test against zero, Ses4: T₍₄₃₎ = 2.727, P_FDR = 0.072, d = 0.832; Ses5: T₍₄₃₎ = 0.717, P_FDR = 0.711, d = 0.219). The SUS group did not show any significant deviations from baseline (T₍₄₅₎ = −2.427 to 0.187, P_FDR = 0.076 to 0.813, d = 0.187 to 0.723).

Previous nonhuman animal studies have primarily highlighted the contribution of subcortical regions (34). To evaluate the contribution of the subcortex to psychological resilience, we calculated the spontaneous neural activity using the fractional amplitude of low-frequency fluctuations (fALFF). Based on a review of prior animal resilience studies (34), seven subcortical regions were selected as the regions of interest (ROIs) (SI Appendix, Fig. S8). In Ses4, fALFF magnitudes in the bilateral posterior hippocampus (HIPp) positively correlated with individual CD-RISC scores (SI Appendix, Fig. S8 A and B: Pearson’s correlation test, r₍₉₀₎ = 0.296, P_FDR = 0.033). In the other six ROIs, this analysis failed to detect significant correlations with CD-RISC (SI Appendix, Fig. S8C).

Temporal Features Support Human-Specific Resilience.

In animal models, individual differences in resilience are represented by differences in brain rhythms, measured using local field potentials (30, 35). These differences in brain rhythms may be influenced by resilience level in humans. We, therefore, evaluated the spontaneous human brain rhythms using EEG (SI Appendix, Fig. S9). In Ses4, as shown in Fig. 3A, we identified two frequency components that peaked at 26.5 Hz (E26.5 in high-beta band) and 43.0 Hz (E43.0 in gamma band), which were negatively correlated with individual CD-RISC scores (Pearson’s correlation test, P_FDR < 0.05). Both components were spatially distributed from the right frontal region to the bilateral parieto-occipital region in Ses4 (Fig. 3B; E26.5: r = −0.436 to −0.270, P_FDR = 7.65E-04 to 0.026; E43.0: r = −0.413 to −0.259, P_FDR = 0.002 to 0.043). This negative correlation partially remained during Ses5, but only the channels at E26.5 (r = −0.359 to −0.292, P_FDR = 0.019 to 0.026). These results were replicated when using a cluster-based nonparametric permutation test (36) (SI Appendix, Fig. S10). The temporal correlation changes highlighted that the correlation was not observed during Ses2 (Fig. 3C; E26.5: r₍₈₈₎= −0.091, P_FDR = 0.535, E43.0: r₍₈₈₎= −0.116, P_FDR = 0.374) or Ses3 (E26.5: r₍₈₈₎= 0.007, P_FDR = 0.951, E43.0: r₍₈₈₎= −0.060, P_FDR = 0.578), but instead appeared >60 min later in Ses4 (E26.5: r₍₈₈₎= −0.395, P_FDR = 0.001, E43.0: r₍₈₈₎= −0.369, P_FDR = 0.002) and were sustained till Ses5 (E26.5: r₍₈₈₎= −0.304, P_FDR = 0.008, E43.0: r₍₈₈₎= −0.255, P_FDR = 0.033). The correlation coefficient was stronger in Ses4 than in Ses2 (test for the difference between correlation coefficients, E26.5: T₍₈₇₎ = 2.942, P_FDR = 0.013, d = 0.631; E43.0: T₍₈₇₎ = 2.745, P_FDR = 0.022, d = 0.588) and Ses3 (E26.5: T₍₈₇₎ = 3.698, P_FDR = 0.002, d = 0.794; E43.0: T₍₈₇₎ = 3.446, P_FDR = 0.005, d = 0.740), while the coefficient in Ses5 was stronger than in Ses2 (T₍₈₈₎ = 2.760, P_FDR = 0.014, d = 0.588) only in E26.5.

Fig. 3.

Spectral EEG features correlate with resilience 60 min after the acute stress experience. (A) Top: Pearson’s correlation coefficients between the individual CD-RISC scores and the power spectral density (PSD) for each frequency band at Ses4. The color of each line corresponds to the 31 channels located on the scalp. Correlation peaks were identified at 26.5 Hz (high-beta) and 43.0 Hz (gamma). Bottom: The statistically significant frequencies in each channel. Large and small dots indicate a significant channel and frequency at the threshold of P_FDR < 0.05 corrected and P_unc. < 0.05, respectively. Frequency bands shown in gray masks are not analyzed due to noise from the fMRI scanner. See SI Appendix, Fig. S9 for correlations between CD-RISC and EEG in other sessions. (B) Temporal changes of topographic maps from Ses2 to Ses5 at 26.5 and 43.0 Hz. White asterisks indicate statistically significant EEG channels with P_FDR < 0.05, corrected. (C) Temporal dynamics of the correlation between CD-RISC and PSD at 26.5 and 43.0 Hz. The correlation coefficient (r) at 26.5 Hz was significantly lower in Ses4 than in Ses2 and Ses3, and lower in Ses5 than in Ses3. The coefficient at 43.0 Hz was significantly lower in Ses4 than in Ses2 and Ses3. *P_FDR < 0.05, **P_FDR < 0.01, corrected for the test of equality of two correlations. Insets are scatter plots of individual CD-RISC scores and means of PSDs taken from statistically significant channels at 26.5 Hz (r = −0.395, P_FDR = 1.55E−4) and 43.0 Hz (r = −0.304, P_FDR = 0.004) in Ses4. (D) Group-wise analysis of the temporal dynamics of PSD at 26.5 and 43.0 Hz. Participants were classified as RES or SUS based on the median CD-RISC score. *P_FDR < 0.05 for paired t test, ^†P_FDR < 0.05 for one-sample t test against zero.

As shown in Fig. 3D, the RES group exhibited significant reductions between Ses3 and Ses4 (paired t test, E26.5: T₍₄₁₎ = 2.860, P_FDR = 0.040, d = 0.894; E43.0: T₍₄₁₎ = 3.193, P_FDR = 0.016, d = 0.997). This significant decline first appeared at Ses4 (one-sample t test against zero, E26.5: T₍₄₁₎ = −3.023, P_FDR = 0.016, d = 0.944; E43.0: T₍₄₁₎ = −3.738, P_FDR = 0.004, d = 1.167) and persisted at Ses5 (E26.5: T₍₄₁₎ = −3.177, P_FDR = 0.016, d = 0.991; E43.0: T₍₄₁₎ = −3.591, P_FDR = 0.004, d = 1.121). Compared with the more gradual session-to-session changes observed in fMRI FC (Fig. 2E), EEG exhibited sharp reductions beginning approximately 60 min poststress and continuing through Ses5 (Fig. 3D).

Note that, at Ses4, source power at 26.5 Hz showed a significant negative correlation with CD-RISC scores in the left dACC (SI Appendix, Fig. S11: Pearson’s correlation test, r = −0.296, P_FDR = 0.048). We also confirmed that the identified brain activations in fMRI and EEG were not correlated with other stress-related measures (SI Appendix, Fig. S12), but were only correlated with the CD-RISC.

Impact Level of Identified Components.

We tested 13 neural and peripheral components at four time points (13 × 4 = 52) that were potentially related to individual differences in resilience. However, the critical question remains: which neurophysiological components are best at predicting the levels of resilience measured by the CD-RISC? Thus, the Elastic-Net was used to calculate the optimal combination of components and their impact on psychological resilience. The Elastic-Net optimization process from the 52 factors revealed a model that included six factors (Fig. 4A, Model 1: adjusted R² = 0.322, P = 1.07E−05). Based on this optimized model, we found the factor that contributed most was RFCn [standardized partial regression coefficient β = −0.254, contribution level (CL) = 40.60%]; the second factor was RFCp in Ses4 (β = 0.204, CL = 32.57%); and E43.0 (β = −0.056, CL = 8.89%), HIPp (β = 0.052, CL = 8.31%), E26.5 (β = −0.047, CL = 7.48%) in Ses4 in descending order (Fig. 4B). This analysis showed that the stress-driven brain or peripheral responses reported in previous studies showed negligible contribution to predict individual psychological resilience in humans (see SI Appendix, Fig. S13 for the robustness of our results).

Fig. 4.

Key neural components associated with psychological resilience factors and a proposed neural model of human resilience. (A) The heat map shows the R² values during the process of optimizing the alpha and lambda parameters for regularization term using the Elastic-Net. The star in the map indicates the peak value. This optimization yielded the best six factors to predict individual resilience (Model 1: α = 1.00, λ = 0.14, adjusted R² = 0.322, P = 1.07E-05) out of 52 (13 measurements × 4 sessions) factors. (B) The weights of the six selected components suggest that RFCn, which overlaps with SaN, and RFCp, which overlaps with DMN at Ses4, are the dominant contributors to the prediction of individual levels. The circled numbers indicate the ascending order of the absolute value of the optimized weights. (C) Model of the neural mechanism of human resilience based on the current findings. In this model, sustained activation of the dACC and MFG, both of which are part of the SaN, suppresses activation of the medial parietal lobe, part of the DMN, through high-beta oscillations. This oversuppression occurs approximately 60 min after stress, forming SUS individuals, and the disinhibition of the DMN contributes to the development of RES individuals.

Discussion

We applied a multimodal approach to elucidate the neurophysiological dynamics related to human resilience and observed that their individual differences peaked 60 min after stress exposure. The identified neural characteristics are independent of neurophysiological responses focused on known stress responses that end within 60 min, and are potential therapeutic targets that effectively leverage individual resilience.

We identified six factors correlated with individual resilience. Furthermore, the machine learning approach revealed that the most influential factors were RFCn and RFCp, both from the FC in Ses4. This result indicates that human resilience is supported by the SaN and DMN in the neocortex, rather than the subcortex system, which has been the focus for many years. A further discussion of each identified factor is provided in SI Appendix, Texts S3.

The combination of fMRI and EEG revealed that resilience-related EEG changes in the prefrontal area spatially and temporally overlapped with FC increases observed in fMRI. Increases in RFCn (SaN) and E26.5 (high-beta) in individuals with lower resilience may reflect the same underlying neural mechanisms. Indeed, a study investigating the resting-state natural frequency reported that high-beta activity (23.3 to 27.1 Hz) is dominant in the MFG (37).

Taken together, a plausible neural mechanism in individuals with lower resilience is that high-beta activity within the SaN, originating in the left dACC, suppresses DMN activity in the parietal and occipital regions (Fig. 4C, “Susceptible”). In contrast, in RES individuals, a reduction in high-beta activity may facilitate increased DMN activity (Fig. 4C, “Resilient”). Note that a similar phenomenon of reduced high-beta power and increased DMN activity has been reported (38). To causally test the validity of this model, it is necessary to artificially suppress high-beta activity after stress. The next research step would be to test the neuromodulation effects: whether the suppression of high-beta activity (1) reduces the activation of the SaN in the prefrontal cortex while enhancing that of the DMN in the parietal cortex, and (2) contributes to individual resilience levels in real life.

It is noteworthy that our group-wise analysis revealed divergent temporal dynamics between fMRI and EEG in individuals with higher resilience (Figs. 2 E and 3 D). One possible interpretation of these findings is that high-resilience individuals exhibit rapid electrophysiological changes specifically between 30 and 60 min after stress exposure, whereas hemodynamic alterations seem to emerge as early as 30 min poststress.

Although previous human studies have examined the correlation between individual resilience scores and resting-state brain activity (34, 39), we provide several advancements beyond the understanding of resilience from prior animal models and existing imaging studies. First, most animal studies have focused on individual differences in the responses to stress exposure, whereas most human studies have not experimentally introduced stress exposure to participants. Conversely, we directly observed the neurophysiological responses to actual stress to verify the initiation of the resilience process. Second, previous studies with acute stressors have not focused on responses longer than 1 h because autonomic and hormonal responses apparently subside within that period. However, physiological responses persist beyond 1 h after stress exposure (Fig. 1B) (21–23), and the neural states more than 1 h after stress are significantly different from the prestressed state (24–26). We, therefore, examined neurophysiological dynamics for more than 60 min, which has been overlooked in conventional resilience studies.

While our fMRI and EEG findings are consistent with the human resilience model, our results do not align with the conserved stress resilience model (SI Appendix, Fig. S2 and Fig. 1 B and C for the hypothesized two models). Although there is no definitive evidence at present, given the time course following stress exposure, the identified changes in fMRI and EEG responses may be driven by immune system processes (21, 22, 29) and/or slow cortisol dynamics (23, 24). Uncovering the physiological underpinnings of these resilience-related brain responses will be instrumental in advancing the field of resilience research.

This study has several technical limitations. First, psychological resilience was operationalized solely using the CD-RISC scores. The adequacy of assessing psychological resilience through subjective, questionnaire-based measures alone remains a matter of debate (40). Although the CD-RISC demonstrates strong reliability and validity among existing resilience-related questionnaires (41–47), several meta-analyses and reviews argue that such instruments may not fully reflect adaptability to future adversity (48, 49). To rigorously determine whether the brain networks identified in the present study can serve as predictive biomarkers for stress-related mental health risk, future research will need to incorporate integrated assessments, including behavioral, ecological, and longitudinal outcomes, alongside questionnaire-based measures.

Second, considering that the physical stress used in the present study represents only one aspect of the wide range of stressors we encounter, future research should explore responses to other types of stressors during this time window.

Our findings highlight the time window in which the neurophysiological resilience process is initiated by acute stress. These neurophysiological dynamics, captured by both fMRI and EEG, emerged more than 1 h after stress exposure—a time window that has received little attention in previous studies. Targeting interventions within this specific window could maximize the effectiveness of protection and treatment for acute stress-related disorders such as posttraumatic stress disorder.

Materials and Methods

Participants.

All experiments were conducted in accordance with the principles of the Declaration of Helsinki and were approved by the ethical review board (approval number 177) of the Kochi University of Technology. Participants were recruited using convenience sampling (via campus flyers and emails to the university’s participant pool). Prior to the main experiment, presurvey questionnaire data were collected in face-to-face sessions between August 8, 2019, and April 28, 2020 (N = 265; 166 males; mean age 20.02 y, range 18 to 24). Each participant in the presurvey received monetary compensation of 1,000 Japanese yen. In the main experiment, 102 participants, who were recruited independently of the presurvey, were enrolled; all had normal or corrected-to-normal vision (77 males; mean age 20.11 y, range 18 to 25). Eight participants were left-handed. Each participant was compensated at a rate of 1,000 yen per hour. Data were collected between November 13, 2019, and April 19, 2021.

The sample size was predetermined based on recent psychological resilience studies employing resting-state fMRI data (50–52). It is noted that the adequacy of sample sizes in brain-wide association studies (BWAS) to ensure reproducible findings remains under debate (53–56). Marek et al. cautioned that small sample sizes in BWAS often lead to inflated effect sizes and poor replicability due to low statistical power (53). Conversely, Bandettini et al. and Gratton et al. have highlighted the utility of longitudinal designs, noting that tracking within-subject changes over time or across interventions can circumvent the pitfalls associated with between-subject variability (54, 55). Indeed, in a recent paper by Kang and colleagues, the authors demonstrated that the longitudinal paradigm increased both effect sizes and replicability (56). Our study design includes a within-subject comparison of changes in brain activity over time following the stressor intervention relative to baseline, thereby meeting these proposed recommendations to some extent.

All participants provided informed consent prior to the experiments and completed a resilience-related personality trait test twice on different days. Four individuals with unstable scores (change >20 points) on the resilience questionnaire were excluded from the analysis. Two participants dropped out of the study.

Data from the remaining 96 participants were used to analyze the physiological data (72 males; mean age, 20.13 y). For the brain imaging analyses, we excluded three individuals with brain morphological abnormalities and two individuals who were prescribed psychiatric medications. During the experiment, we failed to collect one fMRI and two EEG datasets due to machine malfunctions. Therefore, we analyzed data from the remaining 90 participants for fMRI (68 males; mean age, 20.13 y) and 88 participants for EEG (67 males; mean age, 20.17 y).

Experimental Procedures.

Participant experiences.

All neurophysiological experiments were conducted between 12:00 and 20:00 to minimize the diurnal effects of cortisol levels. After completing the four questionnaires, an ECG, a respiratory belt transducer (Resp), and an electroencephalogram (EEG) were attached to the participant. The experiment consisted of five brain imaging sessions with eight physiological measurements between sessions (Fig. 1D). During each brain imaging session, resting-state fMRI and EEG data were collected concurrently for 10 min. The participants were required to open their eyes and fixate on a cross displayed in the center of the screen through a mirror in the scanner. The Pup, ECG, and Resp data were collected during each session. During the ISI, saliva samples were collected, and participants reported their current feelings by moving a slider displayed on the screen. Pup, ECG, and Resp information were collected during this interval. Each interval lasted for 2 min.

Between brain imaging sessions 1 and 2, all participants underwent a 2-min CPT as an acute stressor in the MRI scanner. Between sessions 3 and 4, the participants took breaks outside the scanner, during which they were informed that no further stress would be applied during the experiment.

Questionnaires.

All questionnaires were administered in Japanese using publisher-authorized translations or previously validated Japanese versions. Due to licensing agreements, the full wording and order of the questions cannot be disclosed. All participants completed the battery of four questionnaires using their smartphones or personal computers. The main questionnaire used in this study was the CD-RISC, which assesses individual levels of resilience to stress (32). This scale is considered the most dominant for assessing individual resilience level, as it has the best psychometric properties (41). It comprises 25 items with a total score ranging from 0 to 100. Individuals with a higher score are more RES to upcoming challenges. In addition, we used the Stress Mindset Measure (SMM: eight items, scored from 0 to 32) (57) to assess individual “stress-is-enhancing” or “stress-is-debilitating” mindsets, the Perceived Stress Scale (PSS: 14 items, scored from 0 to 56) (58) to assess chronic daily stress levels, and the Beck Depression Inventory-II (BDI-II: 21 items, scored from 0 to 63) (59) to assess individual depression levels. The presentation order was fixed as follows: 1) CD-RISC, 2) SMM, 3) PSS, 4) BDI-II. No randomization or skip routing was employed for any of the questionnaires.

The presurvey was designed to estimate means and distributions within the population of Japanese university students, rather than to provide nationally representative estimates (SI Appendix, Fig. S1). No sampling weights or model-based adjustments were applied; all analyses were conducted on unweighted data, including all observations. In the main experiments, participants completed the four questionnaires twice: once a couple of days before the experiment and again on the day of the experiment, approximately 40 min before entering the MRI scanner. To ensure data reliability, individuals with unstable CD-RISC scores (defined as a change greater than 20 points between sessions) were excluded from the analyses.

Acute stress manipulation.

Participants underwent a modified version of the CPT, essentially mimicking previous studies (15, 16), but used an ice glove instead of ice water as the stressor to avoid the risk of spilling water in the scanner. The gloves were chilled to approximately −20 °C to match the subjective stress level of placing the hands in ice water. After providing instructions regarding the procedure for stress exposure, the experimenter placed an ice glove on the participant’s left hand for 2 min to induce a physiological stress response. Throughout the procedure, participants maintained a supine position in the scanner and fixated on a central cross displayed on the screen. Participants were able to stop the stress induction procedure in the middle of the test by pressing an emergency bulb in their right hand. The participants were informed that their faces would be recorded using a video camera. The 2-min duration of our procedure differed from the 3-min duration typically used for CPT (15, 16). The decision to use a shorter duration was based on previous studies showing that, in some experiments, more than 20% of participants are unable to keep their hands in cold water beyond the 2-min mark (60, 61). Despite a shorter exposure duration, this procedure was effective (SI Appendix, Fig. S2D). EEG, Pup, ECG, Resp, and saliva samples were collected during the CPT. Participants also reported their subjective feelings immediately after the test.

Data Acquisitions.

Software.

We used MATLAB R2019a (The MathWorks GK, Natick, MA) with Psychtoolbox 3.0.12 for stimulus presentation and data collection, and Google Forms for the collection of questionnaires. The operating system used was Windows 10 Pro.

fMRI.

A 3 T MRI scanner (Siemens Prisma, Munich, Germany) with a 64-channel head coil was used. Functional images were acquired using an echo-planar imaging (EPI) sequence [repetition time (TR): 2.0 s; echo time (TE): 27 ms; flip angle (FA): 90°; 36 slices; slice thickness: 3 mm; in-plane resolution: 2 × 2 mm; descending]. We did not use multiband accelerated imaging because the guidelines for simultaneous fMRI-EEG recordings provided by Brain EEG Products (Germany) did not permit it. Each functional scan comprised 300 scans. High-resolution anatomical images were acquired using an MP-RAGE T1-weighted sequence (TR: 2,400 ms; TE: 2.32 ms; FA: 8°; 192 slices; slice thickness: 0.9 mm; in-plane resolution: 0.93 × 0.93 mm). The anatomical image was acquired twice, before sessions 1 and 5, to account for changes in the head position after the break outside the scanner.

EEG.

EEG data were acquired from MRI-compatible 32-sintered Ag/AgCL electrodes with extended scalp coverage, including 31 scalp recording electrodes and one electrode on the upper back for ECG (BrainCap-MR; Brain Products, Bayern, Germany). Electrodes AFz and FCz served as ground and online references, respectively. All signals were recorded with a bandpass of 0.016 to 250 Hz and were digitized at 5,000 Hz (BrainAmp MR Plus, Brain Products). Electrode impedances were lowered to <10 kΩ prior to recording and monitored throughout the experiment.

Peripheral physiological measurements.

Pup.

Eye movements and Pup were recorded using EyeLink 1,000 PLUS (SR Research Ltd., Ottawa Canada). The pupil diameter was tracked from the right eye at a sampling rate of 500 Hz, measured in centroid mode, to reduce noise in the pupil data. This system was also used to confirm whether the participants were awake.

Heart rate and resp.

An MRI-compatible three-lead ECG and contractible belt (BIOPAC Systems Inc. Goleta, CA) were attached to the chest and abdomen, respectively, to detect the timing of the R peaks and respiratory rhythm. The sampling rate was 10,000 Hz.

Saliva.

To obtain cortisol data, saliva samples were collected using a SalivaBio Oral Swab (Salimetrics, LLC, Carlsbad, CA). Participants were required to place the swab under their tongue for 2 min. The collected swab was placed in a single centrifuge tube and frozen and stored at −50 °C.

Subjective ratings.

The participants rated their momentary stress levels eight times during the experiment, in the scanner (seven times) or outside the scanner (once during the break). For each rating, participants answered the question, “How stressed are you feeling right now?,” which was displayed on the screen. Responses were provided on a −50 to 50 scale moving the slider from 0. In addition, immediately after the CPT, participants rated their subjective feelings of unpleasantness and pain levels following the stress rating.

Statistical Analyses.

Software.

We used MATLAB R2019a, Microsoft Excel ver. 2016 and Microsoft 365 (Microsoft Corporation, Redmond, WA), SPSS ver. 23 (IBM Corp., Armonk, NY), and JASP ver. 0.18.3 (https://jasp-stats.org/). RStudio ver. 2022.07.2 was used for statistical tests and machine learning approaches. The operating systems used were Windows 11 Pro and Linux Ubuntu 18.04.5 LTS.

fMRI.

Resting-state fMRI.

Preprocessing and denoising procedures as well as statistical analyses were performed using the CONN FC Toolbox (ver. 20.b: www.nitrc.org/projects/conn) and SPM12 (ver. 7771: www.fil.ion.ucl.ac.uk/spm/).

The default preprocessing pipeline for volume-based analysis provided by CONN was used. For each participant, the EPI images were temporally realigned across volumes and sessions, and each slice was time-shifted and resampled using interpolation to match the middle of each acquisition time (slice-timing correction). An artifact detection tool (ART)-based scrubbing was performed to detect outliers. The outlier detection threshold was set at the 95th percentile for the normative samples. Functional and anatomical data were normalized to the standard Montreal Neurological Institute space and segmented into gray matter, white matter, and cerebrospinal fluid (CSF) tissue classes using the SPM12 unified segmentation and normalization procedure. Finally, the images were smoothed using a Gaussian kernel of 8 mm full width at half maximum.

Then, in denoising, linear regression of potential confounding effects in the blood-oxygen-level dependent (BOLD) signal and temporal band-pass filtering were conducted (62–64). Noise components were modeled for each participant and each session: white matter (five components), CSF (five components), estimated subject motion parameters [translational displacement (x, y, z) + rotational displacement (pitch, roll, yaw) and their first derivatives (velocity) = [(3 + 3) × 2] = 12 components], estimated outlier scans, and session effects were included as nuisance regressor sets. To remove peripheral cardiac and respiratory artifacts, the collected ECG and Resp data were modeled using RETROICOR with the PhysIO toolbox (65) and applied as additional noise regressors (20 components).

In addition, linear trends, low-frequency drift, and high-frequency noise in the BOLD signal were removed using a temporal bandpass filter (0.008 to 0.09 Hz) to focus on low-frequency fluctuations.

The quality of the preprocessing and denoising was evaluated using the “QC–FC %” (percent match in Quality Control–FC correlations), a quality control metric implemented in the CONN toolbox. We confirmed that the QC–FC % improved from 0.12 ± 0.11 (55.3% overlap with the null distribution) in the raw data to 0.02 ± 0.12 (90.7% overlap) after preprocessing and denoising.

ROI-to-ROI analysis.

We used an original set of ROIs, with 360 regions in the cerebral cortex defined based on Glasser’s template (66), 54 ROIs in the subcortex defined based on Tian’s template (67), and 12 ROIs in the brainstem defined based on the AAL3 (68). In total, 426 ROIs were used for the FC analysis. The cerebellum was excluded from the ROI-based analysis. Individual correlation maps were generated by extracting the mean BOLD time course from each ROI and calculating the correlation coefficients between two ROIs.

Cluster-level inferences were performed using the standard procedures implemented in the CONN toolbox to identify significant networks (69). Briefly, we applied a hierarchical clustering procedure [complete-linkage clustering (70)] based on ROI-to-ROI anatomical proximity and functional similarity metrics. The linkage order was determined using data from the baseline session, with sex and age effects regressed out. By separating the session used to define ROI ordering from those used for the primary objective of detecting networks whose dynamics correlated with CD-RISC scores in sessions 2 to 5, we minimized the risk of statistical circularity.

Based on the linkage order and cluster sizes, we identified a total of 59 clusters from which all reported FC networks were derived. This clustering procedure reduced the dimensionality of statistical testing from 90,525 ROI pairs (426 × 425/2) to 1,711 cluster pairs (59 × 58/2). A cluster-level FDR-corrected P-value of <0.05 was applied to determine the significant FC networks.

Permutation-based analyses using the L₂ norm.

To evaluate the overall distribution of correlations between FCs and CD-RISC scores for each poststress session relative to baseline, we conducted a permutation-based analysis using the L₂ norm as a summary index. Specifically, from a matrix containing the Pearson correlation coefficients between FC and CD-RISC scores across participants for all ROI pairs (426 × 425/2), we calculated the L₂ norm, defined as the square root of the sum of the squared values across all pairs. Statistical significance was assessed via a label-permutation test (n = 1,000), in which CD-RISC scores were randomly shuffled across participants. We report the 90th and 95th percentiles of the resulting permutation distribution (SI Appendix, Fig. S5). Note that by collapsing the matrix into a single index, this approach provides a robust statistical test that mitigates the multiple comparisons problem, avoiding the high false-positive rate inherent in testing thousands of individual ROI pairs.

Seed-to-voxel analysis.

To confirm consistency with previous stress-induced FC changes reported by van Marle et al. (17), we applied seed-based correlation analysis (SI Appendix, Fig. S3D). The bilateral amygdala was used as the seed region. In accordance with the previous report, the statistical threshold was set at P < 0.001, uncorrected.

fALFF.

fALFF maps represent a relative measure of BOLD signal power within the frequency band of interest (e.g., 0.01 to 0.10 Hz) compared to that over the entire frequency spectrum. We used these to assess the BOLD changes in specific subcortical regions reported as resilience-related systems (34). Correlations between individual CD-RISC scores and the bilateral anterior hippocampus, HIPp, amygdala, nucleus accumbens, ventral tegmental area, locus coeruleus, and dorsal raphe nucleus were evaluated (SI Appendix, Fig. S8). The amplitudes of the left and right ROIs were averaged. The significance of the correlation was corrected using an FDR P < 0.05.

Data collection from neurosynth datasets.

Neurosynth is a meta-analysis tool that displays brain activity maps corresponding to cognitive functions by analyzing large amounts of previously accumulated fMRI data (71). This tool is useful to avoid inappropriate inferences regarding the relationship between brain activity and cognitive function. As shown in SI Appendix, Fig. S3C, this tool was used to examine whether the seeds of the stress-related functional networks identified in our analysis overlapped with previously reported stress-related brain activity. The word “stress” was used to search for brain activity, and a brain activity map (uniformity test, P < 0.01, FDR-corrected) based on 321 studies is shown.

EEG.

Resting state.

Preprocessing procedures and statistical analyses were performed using a Brain Vision Analyzer 2 (BVA2; Brain Products) and EEGLAB (ver. 2021.1, https://sccn.ucsd.edu/). First, in BVA2, the MR gradient artifact was removed using an average template subtraction method (72), and the resulting EEG data were downsampled to 500 Hz. Second, the EEG data were bandpass-filtered between 1 and 200 Hz. Third, cardiovascular artifacts were removed using data from one ECG channel. Denoising was performed using a semiautomatic template-matching procedure. Fourth, eyeblink correction was performed using independent component analysis (ICA). Because specific sensors for vertical/horizontal electrooculography (EOG) were not attached to the scanner for safety reasons, we used the EEG data obtained from electrode FP1 as a proxy for the vertical activity of the EOG channel.

Further denoising and analysis were performed using custom scripts and EEGLAB, as follows. The 10 min continuous EEG data were segmented into 300 epochs with a duration of 2.0 s, time-locked to the TR of the fMRI. To detect abnormally noisy channels due to insufficient adhesion between the scalp and electrode, the power spectrum density (PSD) distribution of the channel with the highest signal intensity was compared with the average PSD distribution of the other top ten PSD intensity channels. If the PSD distribution in the given channels was significantly different from the average of the other top 10 channels (Wilcoxon rank-sum test), the channel was excluded, and the data were interpolated by the corresponding channel by replacing it with the averaged EEG data from the neighboring channels. Thereafter, the noise in each epoch was inspected and rejected with the following parameters: abnormal values, over ±120 μV; abnormal trends, 50 μV/epoch; R² limit, 0.3; improbable data, 5 std.; abnormal distribution, 5 std.; abnormal spectra (slow), over −60 to 40 dB, 0 to 40 Hz. On average, 12.10 ± 9.65% of the epochs were excluded with these criteria. Finally, we applied ICA-based denoising techniques implemented in EEGLAB (“runica” function). Each independent component was labeled using the ICLabel plugin (https://github.com/sccn/ICLabel). The independent components with <95% probability of the brain signal were flagged as IC artifacts and rejected. For EEG frequency analysis, each epoch’s data were Fourier transformed to PSD with a frequency resolution of 0.5 Hz, and deleted epochs were filled using the spline interpolation method. We excluded 0 to 0.5, 17.5 to 18.5, and 36.0 to 37.0 Hz from the PSD data because TR-driven fMRI noise components remained dominant even after several noise reduction procedures. The remaining 84 frequency bins were labeled as follows: delta (0.5 to 3.5 Hz), theta (4.0 to 7.5 Hz), low alpha (8.0 to 10.0 Hz), high alpha (10.5 to 12.5 Hz), low beta (13.0 to 21.5 Hz), high beta (21.5 to 29.5 Hz), and gamma (30.0 to 49.5 Hz).

To identify frequency bands that correlated with the CD-RISC score, we calculated Pearson’s correlation coefficients (r) for each frequency bin at each electrode. The obtained P-values were FDR-corrected for multiple comparisons across electrodes.

Cluster-based nonparametric permutation test.

To further validate our findings, we performed a cluster-based nonparametric permutation test (36) (SI Appendix, Fig. S10), implemented in the FieldTrip toolbox (https://www.fieldtriptoolbox.org/). Pearson’s correlation coefficients (r) between CD-RISC scores and PSD were computed across 84 frequency bins and 31 EEG channels. Cluster candidates were identified based on adjacency in frequency using a cluster-forming threshold of P_unc. < 0.01. For each cluster candidate, the sum of t-values (SumT) was used as the cluster-level statistic. The distribution of the maximum cluster mass was estimated using 5,000 permutations, and clusters that survived the family-wise error-corrected threshold of P_FWE < 0.05 were considered significant.

EEG source localization analysis.

To estimate the intracortical distribution of electrical activity from surface EEG data, we used eLORETA (73), implemented in the FieldTrip toolbox. For source modeling, the brain volume was defined using the Montreal Neurological Institute ICBM152 (MNI) template with Boundary Element Method (BEM) model and the 10 to 5 electrode coordinate system. The cortical space was segmented at an 8 × 8 × 8 mm resolution, yielding a source model of 3,274 voxels. In the first-level analysis, intracranial spectral density at 26.5 Hz (high-beta) and 43.0 Hz (gamma) was estimated for each participant using a 1.0 Hz bandwidth and an initial regularization parameter λ = 0.05. Whole-brain probability maps were then generated and resliced into 1 × 1 × 1 mm for subsequent ROI-based statistical analysis (74). Thirty-one seeds from RFCn and RFCp were grouped by spatial proximity within the hemisphere into nine ROIs: left/right dACC, left/right insula, right SMF, right MFG, right IFG, and left/right PCC, including the Prec. Pearson’s correlations between estimated source power and CD-RISC scores were assessed within these ROIs (SI Appendix, Fig. S11).

Group-wise analyses in fMRI and EEG data.

To examine whether the observed temporal changes in fMRI and EEG correlations were driven by more RES or more SUS individuals, participants were divided into two groups based on the median CD-RISC score (60 points). For fMRI analyses, the RES group (RES) comprised 44 participants, while the SUS group (SUS) included 46 participants. For EEG analyses, the RES group included 42 participants, and the SUS group included 46 participants.

Peripheral physiological data.

Pup.

The mean diameter (µm) during each brain imaging session (10 min), ISI (2 min), and CPT (2 min) was used as a representative value for each participant. Some data were lost because of partially open eyes or eyes covered with eyelashes. Data with <70% acquisition per session were excluded. The probability of the remaining valid data was 89.0% for the total sample. Kendall’s correlation coefficients were calculated between the pupil diameter and individual CD-RISC scores at each time point.

ECG analysis.

Beats per minute (BPM) were calculated from R-peak information. The data were downsampled to 1,000 Hz for the analysis. The ECG data contained noise owing to electromagnetic induction derived from rapid magnetic field changes during the MRI scan. A customized Pan–Tompkins ECG QRS detector (75) was used to detect the R–R intervals from the data obtained during scanning. Almost all R peaks were detected automatically by this algorithm; however, some data were manually detected by the experimenter. The probability of the remaining valid data was 98.3% of the total sample. Kendall’s correlation coefficients between the BPM and individual CD-RISC scores were calculated at each time point. In addition, although we attempted to calculate heart rate variability from the R peaks, this was abandoned because it is often difficult to detect accurate R-peak positions owing to MRI-derived noise.

Respiratory rate.

In contrast to the ECG data, the respiratory data were free of MRI-derived noise because they measured changes in air pressure in a belt worn over the abdomen. Data were downsampled to 1,000 Hz for analysis. Exhalation and inhalation peaks were detected, and respirations per minute (RPM) were calculated and averaged for each session. The probability of valid data was 99.6% for all the samples. Kendall’s correlation coefficients were calculated between the RPM and individual CD-RISC scores at each time point.

Cortisol concentration.

Saliva samples were analyzed in the laboratory using a salivary cortisol enzyme-linked immunosorbent assay kit (Salimetrics). The cortisol concentrations (µg/dL) during each ISI (2 min), CPT (2 min), and break outside the scanner (2 min for sampling) were evaluated. Although almost all saliva samples were analyzed in duplicate and the average optical density was calculated to reduce measurement error, some collected samples were tested once, and the other samples were not tested owing to insufficient sample quantity. The probability of the remaining data being valid was 83.9% for the total sample. Kendall’s correlation coefficients were calculated for Cort and individual CD-RISC scores at each time point.

Statistical Tests.

General.

All statistical tests were two-tailed. Correlations between CD-RISC scores and fMRI FC or EEG PSD at each time point were computed using Pearson’s correlation coefficient (r) (Figs. 2 A–D and 3A–C and SI Appendix, Figs. S3 E and F, S4 D and E, S5–S7, S8 B and C, S9 B and C and S10–S12). Comparisons of correlation coefficients between two time points were performed using the test for equality of two dependent correlations (76) (Figs. 2D and 3C). In the fMRI and EEG data, paired t-tests were used to assess temporal differences within participants (Fig. 3D and SI Appendix, Fig. S4B). One-sample t tests against zero were used to compare values from baseline (Session 1) within participants (Figs. 2E and 3D). Effect sizes were calculated using Cohen’s d. F tests were also used in fMRI-FC analyses, and effect sizes were reported using η² (Fig. 2C).

Bayes Factors for nonsignificant results.

To confirm that the conserved stress resilience model was not supported, we assessed the strength of evidence for the null hypothesis using the Bayes Factor (BF). The stretched beta prior was set at (1,1) and the range of coefficients was set at −1 to 1 (noninformative prior). Tests were performed to assess the BF in favor of the H₀ as the null hypothesis over H₁ as the alternative hypothesis (BF₀₁). For peripheral physiological responses, a total of 47 factors were assessed (11 time points for Pup, heart rate, and respiratory rate, and seven time points for Cort and subjective stress rating; SI Appendix, Fig. S2). For brain responses, a total of eight factors were assessed (two factors for fMRI and six factors for EEG data; SI Appendix, Figs. S3 and S4).

Machine learning.

We used linear regression models with a machine learning approach to evaluate the best combinations to predict individual resilience levels from the observed neurophysiological data. The best parameters were selected using the Elastic-Net, which regularized the L₁ and L₂ norms. We validated and reported eight neural factors (SFC, ScAI, RFCn, RFCp, RFCs2, HIPp, E23.5, and E43.0), four peripheral physiological factors (Pup, BPM, Resp, and Cort), and one subjective rating (Subj) potentially related to individual stress or resilience responses. Thirteen modalities at four time points (52 factors) were selected as explanatory factors. All values were subtracted from the value of the baseline session and z-normalized before estimation.

Missing data for explanatory variables were interpolated using the mean values of the remaining participants. In the main model, the individual CD-RISC scores (n = 90) were calculated using 52 parameters. In the second model, 13 modalities were used, but were optimized per session (Ses2 to Ses5). In the third model, only six neural modalities (RFCn, RFCp, RFCs2, HIPp, E23.5, and E43.0) showing significant correlations were used as explanatory factors. Alpha and lambda hyperparameters were optimized using ninefold cross-validation (CV). The search range of alpha was 0 to 1.0 per 0.02, and that of lambda was 0 to 1.5 per 0.02. To account for sampling error during cross-validation, repeated CV is conducted for lambda by repeating the nine-fold CV procedure 100 times. Specifically, the optimal values for the alpha and lambda hyperparameters were determined by identifying the combination that minimized the Mean Squared Error averaged across all 100 repetitions (77, 78).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The data and code that support the findings of this study are available on the Open Science Framework (OSF) at: https://osf.io/kvnh9/ (79).