acceptodds
Under review as a conference paper at ICLR 2027

Learning Brain-State Dynamics for Objective Assessment of Stress Resilience via EEG

Abstract

Stress resilience is a relatively stable psychological trait that reflects an individual's capacity to maintain or rapidly recover adaptive functioning under adversity and sustained pressure. Although EEG signals offer a promising basis for objective resilience assessment beyond self-report measures, estimating stress resilience from EEG remains challenging because resilience is a long-term trait, whereas EEG predominantly reflects real-time, transient neural activity that is entangled with and obscures the slow trait signal. To address this, we propose KoMaNet, the Koopman Manifold Network. It suppresses nuisance at three levels to highlight the underlying trait. A self-supervised state network encodes each clip by imitating an expert descriptor, yielding a stable representation of the brain state. A Koopman dynamic network models the stable dynamics of brain states with a conditioned Koopman operator over the recording context. A manifold subject network classifies subjects by their dynamics on a manifold. We collect the RESIST dataset, comprising overnight-sleep and afternoon-nap paradigms, for stress resilience assessment. KoMaNet achieves AUCs of 0.740 and 0.700 on RESIST Night and RESIST Day, respectively. Model-attribution analyses identify discriminative patterns in occipital delta and frontal theta/alpha/beta bands. On RESIST Night, high-resilience subjects show smaller deviations from stationary dynamics than low-resilience subjects after sleep onset.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Stress resilience, Electroencephalography (EEG), Trait assessment, Koopman operator

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Pb5SOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes KoMaNet (Koopman Manifold Network) for EEG-based objective assessment of stress resilience, aiming to capture stable, long-term psychological traits from transient brain activity. The framework consists of three components: a Self-supervised State Network for learning stable brain-state representations, a Koopman Dynamic Network for modeling state evolution and deviations from stable dynamics, and a Manifold Subject Network for subject-level prediction through distributional representations. The authors also introduce RESIST, a self-collected dataset comprising 160 subjects and approximately 630 hours of EEG recordings across overnight-sleep and afternoon-nap paradigms. Experiments show that KoMaNet outperforms existing EEG and Koopman-based baselines, achieving AUCs of 0.740 and 0.700 on RESIST Night and RESIST Day, respectively. Ablation studies, visualization analyses, and additional experiments on the combined RESIST setting and external LEMON dataset further evaluate the effectiveness and generalizability of the framework.

Critical strengths
  1. Clear motivation and reasonable methodological design. The paper addresses an interesting problem: predicting a relatively stable psychological trait from EEG signals that primarily reflect transient brain activity. The three components of KoMaNet are designed to address different aspects of this challenge, from learning stable state representations to modeling temporal dynamics and aggregating subject-level information. The ablation results show that each component contributes to the final performance, providing reasonable support for the proposed design.
  2. A new dataset and encouraging experimental results. The authors collect RESIST, containing 160 subjects and approximately 630 hours of EEG recordings, which is valuable given the limited availability of EEG datasets with stress resilience annotations. KoMaNet achieves AUCs of 0.740 and 0.700 on RESIST Night and RESIST Day, outperforming the strongest baselines by 0.070 and 0.104, respectively. The additional ablation and visualization experiments also help support the effectiveness of the framework.
Critical weaknesses
  1. Concerns about the robustness of KoMaNet. KoMaNet achieves AUCs of 0.740 and 0.700 on RESIST Night and RESIST Day, respectively, but its performance drops to 0.577 in the combined RESIST experiment (Appendix B), only slightly higher than the best baseline (0.569). The combined setting uses six shared electrodes and 30-minute recordings, so the reduced recording duration and differences between cohorts may partly explain this degradation. Nevertheless, this result raises concerns about the robustness of KoMaNet under different recording configurations. Further analysis would help clarify the factors behind this performance drop, particularly given the paper's emphasis on its applicability to both overnight sleep and afternoon naps.
  2. Impact of class imbalance on evaluation reliability. The RESIST dataset exhibits substantial class imbalance, with only 19 low-resilience subjects out of 100 in RESIST Night and 11 out of 60 in RESIST Day. Under five-fold cross-validation, each test fold therefore contains very few low-resilience subjects, making the evaluation potentially sensitive to individual predictions. Although the authors report balanced accuracy and AUC over multiple random seeds, it remains unclear how consistently KoMaNet performs on the minority class. Additional analysis of class-specific performance and variability across folds would help establish whether the reported improvements are reliable, particularly for identifying low-resilience subjects.
Questions for authors
  1. In Appendix B, KoMaNet's AUC drops to 0.577 in the combined RESIST setting, compared with 0.740 and 0.700 on RESIST Night and RESIST Day separately. What are the main factors behind this performance drop? Could it be related to the shorter recording duration, reduced electrode coverage, differences in the pretrained encoders, or differences between the two cohorts?
  2. Given the substantial class imbalance in RESIST Night (19 low-resilience vs. 81 high-resilience subjects) and RESIST Day (11 vs. 49), could the authors provide class-specific performance, such as sensitivity and specificity, along with the variability across cross-validation folds? This would help assess how reliably KoMaNet identifies low-resilience subjects.
  3. How was the CD-RISC-10 threshold of 20 selected for binary classification? Was it chosen as the midpoint of the scale, based on established criteria, or according to the score distributions? Have the authors examined whether the classification results are sensitive to this threshold?
  4. Appendix A states that the control states are determined using sleep stages, phase bins, or k-means clustering. Could the authors clarify how the sleep-stage information is obtained for RESIST Night and RESIST Day, and whether external sleep-stage annotations are required during inference?
  5. Figure 5 demonstrates cross-session stability using six EEG recordings from a single subject over approximately 2.5 months. Could the authors clarify whether these predictions follow the same subject-wise held-out evaluation protocol described in Section 4.2? Are repeated-session recordings available for other subjects?
  6. Since RESIST is presented as one of the main contributions, do the authors plan to release the dataset, including EEG recordings and resilience annotations? If public release is restricted by privacy or ethical considerations, would controlled access be possible?
5 comments
AuthorsOpenReview ↗

R1. Performance on the combined RESIST cohort (Weakness 1, Question 1).

Thanks for your constructive comments. On RESIST Night, KoMaNet predicts resilience from six hours of overnight EEG using only six electrodes (AUC 0.740). On RESIST Day, it uses 128 electrodes but requires only 30 minutes of afternoon-nap EEG (AUC 0.700).

Appendix B asks whether a more portable setup can combine the low electrode count of the overnight setting with the short recording time of the daytime setting. We evaluated a single model across both cohorts using 30 minutes of EEG from the six shared electrodes. The model achieved an AUC of 0.577, indicating that prediction remained feasible under these more restrictive conditions, although discrimination was modest and performance declined relative to both main-text settings.

AuthorsOpenReview ↗

R2. Class-specific performance and fold variability (Weakness 2, Question 2).

Thanks for your constructive comments. Following your suggestion, we report sensitivity and specificity with low resilience as the positive class. Sensitivity and specificity were 0.412 ± 0.091 and 0.858 ± 0.109 on RESIST Day, and 0.557 ± 0.252 and 0.891 ± 0.033 on RESIST Night, respectively.

At these thresholds, the model missed roughly half of the low-resilience participants on average across folds and identified the larger high-resilience group more accurately in both cohorts. The small number of low-resilience participants limits how precisely we can estimate sensitivity. In future work, we plan to collect EEG recordings from more participants with low resilience.

AuthorsOpenReview ↗

R3. Choice of the CD-RISC-10 threshold (Question 3).

Thanks for your constructive comments. We chose 20 because it is the midpoint of the CD-RISC-10 score range (0–40). Higher scores indicate greater self-reported resilience. For our binary prediction task, we assigned scores of 20 or below to the lower-resilience group and scores above 20 to the higher-resilience group. At this cutoff, KoMaNet achieved an AUC of 0.740 on RESIST Night and 0.700 on RESIST Day (Table 1).

We also predicted continuous CD-RISC-10 scores without applying a cutoff. We evaluated these predictions using Gaussian negative log-likelihood (NLL) and empirical coverage of the central 50% prediction interval (Cov@50). NLL was 3.25 on Night and 3.36 on Day. The corresponding intervals covered 58% and 55% of observed scores, respectively (Table 1).

Following your suggestion, we evaluated two additional CD-RISC-10 cutoffs. In each fold, we calculated the mean and median questionnaire scores using only the training participants.

CD-RISC-10 cutoffNight BAccNight AUCDay BAccDay AUC
Scale midpoint (20)0.7240.7400.6350.700
Training-fold mean0.6360.7100.5590.638
Training-fold median0.6280.7050.5450.621

The 20-point cutoff yielded the highest BAcc and AUC on both datasets. Using the training-fold mean or median can place more participants with similar questionnaire scores on opposite sides of the cutoff, which may make the groups harder to distinguish. We use 20 as a scale-based task definition.

Reviewer_D5jqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes KoMaNet for predicting stress resilience from sleep EEG. It combines self-supervised representation learning, a state-conditioned temporal predictor, and a distribution-based subject classifier. The authors introduce the RESIST dataset and report improvements over several EEG models. However, the paper provides limited insight into what this combination contributes beyond existing methods.

Critical strengths

EEG-based assessment of psychological traits like stress is a meaningful application. Collecting the RESIST dataset requires substantial effort and could support further research. As the authors point out, turning correlates into a robust predictive model is challenging due to lack of publicly available EEG datasets. RESIST has direct stress resilience annotations. Experiments on subject-wise cross validation and ablation studies have been conducted.

Critical weaknesses

I am afraid the technical contribution is minor. The method combines components from existing methods without clearly identifying a new learning principle or explaining why this particular combination is necessary. The motivation seems quite specific and restricted to the application, with limited machine learning insight. The presentation is not good, with most of the equations explaining standard operations in deep learning. Little explanations on insights are provided.

Empirically, the experiments do not resolve the concerns above. Comparison against the baselines seem unfair since they aggregate predictions by voting while KoMaNet learns subject-level aggregation. It is unclear where the improvement comes from. The ablation studies remove components but do not compare them with simpler alternatives. Finally, predicting questionnaire scores does not establish that the model measures objective resilience or recovery from stress.

Questions for authors

Please refer to the weaknesses

4 comments
AuthorsOpenReview ↗

R1. Technical contribution (Weakness 1).

Thanks for your comments. Rather than proposing an isolated optimization rule, we build on the established Koopman principle of modeling nonlinear dynamics [1] (Lines 129–133) to address a fundamental timescale mismatch in long-horizon EEG modeling. Existing EEG models excel at recognizing short-lived patterns associated with particular stimuli or brain states, whereas stress resilience is a relatively stable trait whose associated dynamics may be obscured by transient activity (Lines 077–085).

KoMaNet comprises the Self-supervised State Network, the Koopman Dynamic Network, and the Manifold Subject Network (Lines 087–098). First, the Self-supervised State Network learns to predict the average physiological descriptor of neighboring clips from a single clip, capturing locally sustained brain states while reducing the influence of isolated transients (Lines 162–181). The Koopman Dynamic Network then models brain-state transitions across neighboring clips, learning the typical evolution under each state condition and using the residual to quantify how far each observed transition departs from that evolution (Lines 089–094, 198–219). Finally, the Manifold Subject Network aggregates state embeddings and transition residuals across the entire recording into a joint empirical distribution for each subject and compares these distributions to predict stress resilience (Lines 094–098, 218–219, 229–252).

The ablations in Section 4.3 (Fig. 3) show that ablating any of the three modules reduces both balanced accuracy and AUC, supporting their necessity and complementary roles (Lines 344–354). Section 4.4 (Fig. 6) provides an interpretation of the model on RESIST Night. Relative to each group’s wake baseline, the mean deviation curves are close before sleep onset, but after sleep consolidates into NREM, the low-resilience group consistently shows greater deviation than the high-resilience group (Lines 408–414). This group difference persists across successive state transitions, supporting our approach of aggregating local states and their deviations over the full recording for subject-level trait prediction (Lines 094–098, 218–219, 229–252, 408–414).

[1] Bethany Lusch, J. Nathan Kutz, and Steven L. Brunton. Deep learning for universal linear embeddings of nonlinear dynamics. Nature Communications, 9(1):4950, 2018.

AuthorsOpenReview ↗

R2. Machine learning insight and interpretation of the equations (Weakness 1).

Thanks for your comments. Assessing stress resilience has practical relevance for human resource management, particularly in roles such as air traffic control and emergency medicine that require people to function under sustained pressure (Lines 034–041). We introduce KoMaNet, the first model for predicting individual stress resilience from EEG (Lines 099–104). We formulate the task as learning a representation of each subject’s EEG sequence that predicts a long-term trait (Lines 077–098, 187–190). We also evaluated KoMaNet on STAI anxiety prediction using resting-state EEG from the public LEMON dataset. It achieved a balanced accuracy of 0.624 and an AUC of 0.617, exceeding all baselines in Table 3 (Appendix D, Lines 782–827).

Eqs. (3)–(17) connect the two timescales by pairing a training target that favors activity persisting across nearby clips with a subject-level readout that summarizes state transitions across the recording. Eqs. (3)–(5) show how the encoder learns features of locally sustained brain activity. From a single EEG segment, the encoder predicts the mean physiological descriptor across neighboring segments. Averaging the target reduces the influence of isolated fluctuations and favors patterns that persist across segments (Lines 162–181). Eqs. (7)–(10) describe how we model brain-state transitions. The model summarizes preceding segments and uses a spectrally constrained Koopman operator to capture the typical evolution of a given state across training subjects. Subtracting the expected embedding from the observed embedding gives a residual that measures how far the transition deviates from this reference (Lines 198–216). Eqs. (11)–(12) specify how we learn this reference. The model fits one-step transitions using the observed embedding of the next segment as the target and the Huber loss as the training objective (Lines 213–225). Eqs. (13)–(17) show how local transitions yield a subject-level prediction. The model represents each subject by the joint empirical distribution of state embeddings and their corresponding residuals across the recording. It uses a learned projection and maximum mean discrepancy (MMD) to compare this distribution with those of the training subjects. The resulting similarities are used to predict stress resilience, linking short-term state changes to a long-term subject-level trait (Lines 094–098, 218–219, 229–256).

AuthorsOpenReview ↗

R3. Baseline comparisons and subject-level aggregation (Weakness 2).

Thanks for your comments. Table 1 compares KoMaNet with 19 methods across three families. Counting from TCNet as the first model row, rows 1–7 cover task-specific EEG models, rows 8–12 cover EEG foundation models including the recent CBraMod and CSBrain, and rows 13–19 cover Koopman-based models including ResKoopNet and SKOLR. Row 20 reports KoMaNet (Lines 289–312, 328–338).

To compare KoMaNet with baselines trained for subject-level prediction, we use a multiple instance learning (MIL) design. We select the three models in each family with the highest mean AUC across RESIST Night and Day in Table 1. For each model, the clip embeddings from one subject form a bag with one subject-level label. The mean pooling classifier (ME) averages the clip embeddings and passes the result through two linear layers with a ReLU between them. The max pooling classifier (MA) takes the elementwise maximum across clip embeddings and uses the same classifier head. Following [2], the gated attention classifier (GA) learns normalized clip weights and passes the weighted sum of the embeddings to the same classifier head.

ModelRESIST Night (BAcc.)RESIST Night (AUC)RESIST Day (BAcc.)RESIST Day (AUC)
EEGNet-ME0.6600.6540.5760.579
EEGNet-MA0.6490.6530.5710.567
EEGNet-GA0.6500.6540.5980.609
CSPNet-ME0.6560.6460.6050.616
CSPNet-MA0.6430.6200.6070.620
CSPNet-GA0.6520.6340.5910.606
TSCeption-ME0.6500.6420.5970.570
TSCeption-MA0.6670.6430.6000.578
TSCeption-GA0.6670.6520.5750.557
EEGPT-ME0.5860.6340.5290.527
EEGPT-MA0.6000.6490.4970.512
EEGPT-GA0.5950.6360.5010.519
CBraMod-ME0.6060.6440.5270.529
CBraMod-MA0.5900.6240.5330.539
CBraMod-GA0.5810.6160.5330.528
LaBraM-ME0.5730.6000.5250.524
LaBraM-MA0.5930.6030.5030.506
LaBraM-GA0.5860.6040.5030.495
Koopa-ME0.6160.6770.5730.585
Koopa-MA0.6150.6660.5380.546
Koopa-GA0.6420.6900.5430.565
ResKoopNet-ME0.6180.6540.5520.556
ResKoopNet-MA0.6030.6350.5210.531
ResKoopNet-GA0.6200.6550.5520.552
DeepKoopman-ME0.6070.6360.5470.537
DeepKoopman-MA0.5980.6170.5480.539
DeepKoopman-GA0.6130.6300.5710.559
KoMaNet (ours)0.7240.7400.6350.700

The three pooling methods have similar expected scores within each baseline model, and KoMaNet scores higher than all 27 MIL variants on both balanced accuracy and AUC across RESIST Night and Day.

Finally, Section 4.3 follows standard ablation practice [3] and explicitly compares each removed component against a simpler alternative (Lines 344–354). Specifically, we replace the Self-supervised State Network with supervised chunk classification, replace the Koopman Dynamic Network with raw state embeddings, and replace the Manifold Subject Network with mean pooling and a linear classifier (Lines 344–349). In all cases, replacing our modules with these simpler alternatives degrades both BAcc and AUC (Fig. 3).

[2] Maximilian Ilse, Jakub Tomczak, and Max Welling. Attention-based deep multiple instance learning. In Proceedings of the 35th International Conference on Machine Learning, pp. 2127–2136, 2018.
[3] Richard Meyes, Melanie Lu, Constantin Waubert de Puiseau, and Tobias Meisen. Ablation studies in artificial neural networks. arXiv preprint arXiv:1901.08644, 2019.

AuthorsOpenReview ↗

R4. Validity of the resilience assessment (Weakness 2).

Thanks for your comments. We used the 10-item Connor–Davidson Resilience Scale (CD-RISC-10), a widely used self-report measure of psychological resilience (Lines 042–045). Its original validation study found good internal consistency and provided evidence of construct validity [4]. In RESIST, participants completed the questionnaire once in a controlled research setting. Their scores carried no selection consequences that would give them an incentive to exaggerate their stress tolerance, and a single administration avoided familiarity effects from repeated testing (Lines 046–057, 068–071).

Figure 5 examines within-person consistency. For one participant recorded during six afternoon naps over approximately 2.5 months, the session-level EEG predictions remained close to the CD-RISC-10 score of 19 obtained after the final session (Lines 389–406). This case provides preliminary evidence of repeatability across recordings.

For future work, we plan to validate the learned assessment against behavioral stress-reactivity measures as an external criterion (Lines 466–469).

[4] Laura Campbell-Sills and Murray B. Stein. Psychometric analysis and refinement of the Connor-Davidson Resilience Scale (CD-RISC): Validation of a 10-item measure of resilience. Journal of Traumatic Stress, 20(6):1019–1028, 2007.

Reviewer_BoEdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces KoMaNet for predicting questionnaire-assessed stress resilience from sleep EEG. A self-supervised encoder learns brain-state representations from physiological descriptors, a state-conditioned temporal model predicts their evolution, and a subject-level network aggregates distributions of representations and prediction residuals. The proposed interpretation is that these features capture deviations from stable dynamics relevant to resilience. The evaluation includes classification and regression on the RESIST Night and RESIST Day cohorts, component ablations, analyses of temporal and spatial patterns, and an additional anxiety-prediction task on LEMON.

Critical strengths
  1. The data collection addresses a useful and difficult prediction problem. Pairing extended sleep EEG with resilience questionnaires in two recording paradigms provides a valuable setting for studying individual differences. The subject-wise evaluation and comparisons with several model families support the practical value of the complete approach.

  2. The subject-level representation is a sensible use of long recordings. Aggregating distributions of state embeddings and prediction residuals allows the model to use both typical states and departures from expected evolution. The ablations support the usefulness of this combination, which is a more consequential contribution than any single performance comparison.

Critical weaknesses
  1. The recovery interpretation is not established by the quantity being measured. The introduction motivates the method through deviations from, and returns to, stable brain-state dynamics. Equation 10 instead defines a conditional one-step prediction error. Such an error measures predictability, but does not by itself identify recovery after a disturbance or explain why greater predictability indicates greater adaptive capacity. Questionnaire supervision can support a useful objective proxy; it does not establish this particular dynamic interpretation. The limitations acknowledge the distinction between perceived and actual resilience, but the central deviation–recovery story still goes beyond what the reported residual analyses establish. This matters because that interpretation is the paper’s main distinction from a general EEG phenotype predictor.

  2. The scientific role of the Koopman structure remains insufficiently connected to the represented dynamics. Equation 9 constrains the state-specific operator, while the actual predictor also uses a context summary, normalization, and decoding. The paper needs a clearer account of which observable evolves under the operator and how its modes relate to persistence or stability in the representation used to compute the residual. The dynamic-module ablation removes the operators and residual pathway together, so it supports the module’s usefulness without isolating the role of its spectral structure. The operator constraint is valid as stated, but its broader interpretation as stable brain-state dynamics is not yet sufficiently established.

Questions for authors
  1. What operational definition and evidence support the recovery interpretation? Equation 10 defines a one-step residual relative to the conditional prediction. Please explain why this prediction represents a reference for stable dynamics and what subsequent change constitutes a return toward that reference. Which quantity measures this return and over what time scale—for example, a return time or a decay rate, if either is part of the intended interpretation? Please identify which reported analysis establishes this behavior: the smaller residuals and group separation in Figures 4 and 6 do not by themselves distinguish recovery from greater one-step predictability. What assumptions or evidence connect the proposed quantity to adaptive recovery under adversity, given supervision by CD-RISC-10? If the evidence supports only predictability associated with questionnaire scores, please state that scope explicitly and qualify the recovery claim.

  2. What dynamics does the Koopman operator represent, and what does its spectral constraint establish? Equation 7 constructs a history-dependent context state hn,t−1h_{n,t-1}, while Equation 10 compares Kcn,thn,t−1K_{c_{n,t}}h_{n,t-1} with the current embedding ω(en,t)\omega(e_{n,t}). Please write the evolution relation that the operator is intended to approximate, identifying the observable and whether the relation supports repeated evolution or only context-to-next-embedding prediction. Then distinguish what Equation 9 guarantees for a fixed KcK_c from the claimed persistence or stability of brain-state trajectories when contexts and states change. What derivation or targeted analysis connects its near-unit modes to those trajectories? For example, a comparison with an unconstrained linear predictor retaining the same context construction, downstream mapping, loss, and residual aggregation could isolate the spectral constraint's contribution; an analytical explanation could instead establish the relevant property. The current whole-module ablation removes the predictor and residual together and therefore does not isolate this role. Please clarify which interpretation the available evidence supports.

5 comments
AuthorsOpenReview ↗

R1. Interpretation of the Koopman residual (Weakness 1).

Thanks for your helpful comments. Our goal is to learn stable patterns of brain-state evolution shared across subjects within each state condition. The encoder first extracts locally persistent brain-state information from EEG clips (Lines 162–181). The model then pools transitions from training subjects in the same state condition to learn the corresponding Koopman operator. The spectral constraint prevents each linear step from amplifying differences between input representations. Under repeated application of the same operator, these differences therefore cannot grow over time. Given an individual’s recent trajectory and state condition, the model uses this shared reference to predict the next clip embedding. Equation (10) subtracts the predicted embedding from the observed embedding. A larger residual magnitude indicates that the individual transition departs further from the group’s stable pattern of evolution under the same condition (Lines 198–219).

Figure 6 shows how this distinction relates to resilience. Relative to their respective waking baselines, the two groups have similar mean residuals before sleep onset. During stable NREM sleep, the low-resilience group shows persistently larger deviations than the high-resilience group (Lines 408–414). The dynamics ablation also lowers AUC on both RESIST Night and RESIST Day (Fig. 3, Lines 344–354). Together, these results support the predictive value of modeling how brain states evolve over time.

Human sleep studies motivate the link to stress regulation. Acute social stress before bedtime alters EEG activity during subsequent N2 and slow-wave sleep [1], and greater NREM slow-wave activity is associated with lower anxiety the following morning [2]. These findings motivate testing whether individual differences in sleep dynamics also predict recovery after stress. Figure 6 establishes a sustained group difference in deviation from typical dynamics. For future work, we plan to test whether the learned EEG measures predict independently assessed recovery following a standardized stressor.

[1] Emil Hein et al. Does sleep promote adaptation to acute stress: An experimental study. Neurobiology of Stress, 29, 100613, 2024.

[2] Eti Ben Simon et al. Overanxious and underslept. Nature Human Behaviour, 4, 100–110, 2020.

AuthorsOpenReview ↗

R2. Role of the Koopman operator, spectral constraint, and residual pathway (Weakness 2).

Thanks for your helpful comments. Koopman operator theory provides a way to model nonlinear dynamics by describing how observables of a system evolve under a linear operator (Lines 129–133). We use this idea to learn how EEG activity typically evolves within a given brain-state condition. The observable we predict is the encoder embedding of the next EEG clip. For each condition, we learn a linear operator shared across training subjects, apply it to a summary of recent EEG embeddings, and decode the result after normalization. Equation (9) constrains this linear step so that it cannot amplify differences between two context representations. The resulting prediction serves as a reference for typical evolution under that condition. The residual in Eq. (10) measures how far an individual transition departs from this reference (Lines 198–219).

Following your suggestion, we extended the ablation study in Fig. 3 (Lines 344–354). Under the same experimental setting, we evaluated three configurations that remove the spectral constraint, the residual pathway, and the Koopman operator, respectively. In the third configuration, removing the operator also removes the spectral constraint applied to it and the residual pathway that uses its prediction.

Koopman operatorSpectral constraintResidual pathwayNight BAccNight AUCDay BAccDay AUC
✓✗✓0.6770.6460.5900.661
✓✓✗0.6630.6240.6120.687
✗✗✗0.6480.6510.5850.650
✓✓✓0.7240.7400.6350.700

As shown in the table, the full model achieves the highest BAcc and AUC on both datasets. Without the spectral constraint, AUC decreases from 0.740 to 0.646 on Night and from 0.700 to 0.661 on Day. Without the residual pathway, AUC decreases to 0.624 and 0.687, respectively. These results suggest that the Koopman operator, spectral constraint, and residual pathway together contribute to the performance gains of the full model.

AuthorsOpenReview ↗

R3. Operational definition and evidence for recovery (Question 1).

Thanks for your helpful comments. We define deviation in Eq. (10) as the difference between the observed state embedding and its conditional prediction. Given the current brain-state condition and recent EEG trajectory, the model uses a Koopman operator shared across training subjects to predict the next clip embedding. The spectral constraint ensures that applying the same linear operator to two input representations cannot increase the distance between them. This property motivates its use as a reference for stable evolution (Lines 198–219). We obtain one transition residual every 30 seconds on RESIST Night and every 10 seconds on RESIST Day. A decrease in residual magnitude means that subsequent transitions more closely follow the predictions made for their respective contexts. Because the reference updates with the recent trajectory, a smaller residual alone cannot establish a return to the pre-disturbance state. We did not measure a post-disturbance return time or decay rate.

Figure 4 projects the RESIST Night dynamic embeddings into three dimensions. The two groups’ mean trajectories remain separated throughout the night, and the high-resilience group follows a smoother mean trajectory (Lines 380–388). Figure 6 examines the transition residuals directly. Relative to their respective waking baselines, the groups have similar mean residuals before sleep onset. During stable NREM sleep, the low-resilience group shows persistently larger mean residuals (Lines 408–414). Its transitions therefore deviate further, on average, from the typical evolution learned across subjects during this period.

The role of sleep in stress and emotion regulation motivates this study. For example, acute social stress before bedtime affects EEG activity during subsequent NREM sleep [1], and greater NREM slow-wave activity is associated with lower anxiety the following morning [2]. In our study, CD-RISC-10 provides a subject-level measure of self-reported resilience. We agree with the reviewer and will qualify the recovery claim accordingly. The predictability of brain-state transitions during sleep, relative to a shared dynamical reference, is associated with questionnaire-assessed resilience. These transitions help predict an individual’s resilience group. Establishing a link to adaptive recovery after adversity requires independent behavioral measures.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.