Koopman Operator Analysis of EEG Microstate Dynamics Reveals a Cross-Dataset, Hearing-Level Adjusted Signature of Tinnitus
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Koopman Operator Analysis of EEG Microstate Dynamics Reveals a Cross-Dataset, Hearing-Level Adjusted Signature of Tinnitus".
Marcus: Here's the gist of the paper. Resting-state EEG biomarkers of tinnitus robust to cross-subject and cross-platform variation provide a way to investigate neural dynamics associated with tinnitus The PCA-based Koopman features yield the strongest cross-dataset metrics across both transfer directions, outperforming microstate-derived features Feature Extraction Frameworks The study systematically evaluates two feature extraction frameworks applied to resting-state EEG from three independently collected datasets for binary classification of tinnitus The methods employed include microstate analysis and Koopman operator analysis 1.
Ines: Microstate Analysis involves extracting topographic states using the k-means clustering algorithm, which results in transition probability and state duration features 2.
Marcus: Koopman Operator Analysis utilizes Hankel-embedded Dynamic Mode Decomposition (DMD) to extract complex eigenvalues, where the magnitude encodes oscillation stability and the phase encodes oscillation frequency Cross-Dataset Generalization Performance The classification performance was assessed in a strict cross-dataset paradigm where a linear support vector machine is trained on one dataset and evaluated on another without retraining The results demonstrate that PCA-based Koopman features yield the strongest cross-dataset metrics across both transfer directions, achieving an AUROC of 0.7719 in D1 → D2 and an Accuracy of 0.7623 in D2 → D1 The Wasserstein consistency analysis further reveals that Koopman eigenvalue magnitude generalizes across datasets with a mean consistency ratio of ρ¯ = 0.685, whereas the eigenvalue phase does not generalize with a mean ratio of ρ¯ = 1.583 Biomarker Interpretation The analysis provides interpretable evidence that altered oscillatory decay rates, rather than frequency shifts, constitute the more robust tinnitus biomarker This is supported by the observation that the eigenvalue magnitude distribution shows a concentration near the unit circle, indicating persistent, stable energy changes within time windows The complex-valued nature of Koopman modes allows for capturing both amplitude and phase information of underlying neural oscillations Modes with λ ≈ 1 are assigned high reliability scores, suggesting that these modes represent genuine tinnitus-related neural signatures decoupled from acquisition-specific artifacts The study concludes that the oscillation stability (eigenvalue magnitude) is the more reliable and generalizable tinnitus biomarker This robustness is confirmed because the consistency ratio ρ = Wcross-dataset / Wwithin-dataset < 1 indicates that class-related distributional shifts in oscillation stability generalize across datasets The spatial analysis of the most discriminative PCA component reveals a predicted spatial pattern of the tinnitus-associated dynamical signature, linking it to known neuroanatomical substrates like auditory cortex or fronto-parietal networks Limitations Several limitations were noted in the study, including relatively small clinical datasets which reduces the reliability of estimated classification metrics and makes the D2 → D1 direction more susceptible to high-variance performance estimates The classifier is restricted to a linear SVM, which may not capture higher-order interactions between feature dimensions Furthermore, the current analysis is restricted to binary tinnitus versus control classification and does not address tinnitus severity grading or subtype identification, which would be clinically relevant applications The DMD embedding features were computed for various delay and lag parameters but their full potential warrants further investigation References
1: J.
Ines: M.
Paper summary: Ines: So we're looking at this paper called "Koopman Operator Analysis of EEG Microstate Dynamics Reveals a Cross-Dataset, Hearing-Level Adjusted Signature of Tinnitus." The main thing they're trying to do is find a reliable neural signature for tinnitus that actually works across different subjects and even different machines.
Marcus: Right. They’re looking at EEG data from two different datasets and seeing if they can build a classification system that doesn't just work on the data it was trained on. The core thesis here is that microstate theory has been around for characterizing those quasi-stable spatial configurations in EEG, and this paper wants to improve on existing dynamical systems analysis because generalization is really hard.
Yuki: From a population genetics standpoint, I'm curious about how much of this finding we can trust when we look at the wider history of tinnitus across different groups. If it's not robust across datasets, how much weight do these individual subject findings carry?
Ines: That’s what they address by testing the microstate features and then moving to a Koopman operator framework. They found that while standard microstate features don't generalize well, they get AUROC values close to zero point five in both directions when you try to transfer them between datasets <ref:2604.22116#pg3>.
Marcus: Exactly. So they took the PCA-reduced data and used a Hankel Dynamic Mode Decomposition algorithm to embed it into a lifted space, then fit a Koopman operator on that space to estimate linear dynamics <ref:2604.22116#pg3>. That’s where they extract the Koopman modes and eigenvalues.
Yuki: So the paper moves from spatial patterns of scalp potential to these underlying mathematical descriptions of how the system evolves over time, which is a bigger leap for linking it to biological mechanisms. What does this dynamical description tell us about what's actually going on in the brain?
Ines: Well, they look at those eigenvalues—the Koopman modes—and the magnitude of those eigenvalues encodes oscillation stability, and the phase encodes the oscillation frequency <ref:2604.22116#pg2>. They found that focusing on these eigenvalue magnitudes gives them a much more stable biomarker than focusing on the phase <ref:2604.22116#pg3>.
Marcus: That's significant because it means altered oscillation decay rates, rather than just frequency shifts, are what they see as the most robust tinnitus biomarker <ref:2604.22116#pg3>. They saw that the eigenvalue magnitude distribution clusters near the unit circle, suggesting persistent changes in energy within time windows.
Yuki: So if we think about population genetics, this suggests that these stable energy changes might be more fundamental than transient fluctuations, which could have implications for understanding tinnitus as a chronic condition rather than just an acute symptom.
Ines: It really does. This robustness is quantified by the Wasserstein consistency analysis, where they found a mean consistency ratio of rho equal to zero point six eighty five for the eigenvalue magnitude <ref:2604.22116#pg3>. That means class-related distributional shifts in oscillation stability actually generalize across datasets.
Paper summary: Marcus: That’s a key piece of evidence for clinical translation because it shows that the tinnitus signature isn't just a fluke from one specific patient group or scanner setup. It generalizes across different acquisition conditions, which is exactly what we need when moving toward real-world diagnostics.
Yuki: And they also looked at the spatial pattern derived from the most discriminative principal component, and they found it predicts a spatial pattern linked to known neuroanatomical substrates like auditory cortex or fronto-parietal networks <ref:2604.22116#pg3>. That connects the math back to where in the brain this dynamic signature is happening.
Ines: So, to summarize this paper, it’s about using Koopman operator analysis on EEG microstate dynamics to find a tinnitus biomarker that isn't just good on its training data but actually holds up when you try to apply it elsewhere.
Marcus: The authors took two resting state EEG datasets and compared microstate features against Koopman eigenvalues, finding that the stability of the oscillation magnitude is the most reliable feature, achieving an AUROC of zero point seven seven one nine in one direction and an Accuracy of zero point seven six two three in the other direction <ref:2604.22116#pg3>.
Yuki: And what they’re saying is that this stability metric is consistent across different acquisition settings, which is important for any large-scale study trying to build a universal model.
Ines: It really points toward a more fundamental description of the neural dynamics involved in tinnitus than just looking at simple frequency changes, which we often see. This paper suggests that the underlying way the brain's oscillations are decaying or persisting is what’s truly characteristic of the condition.
Marcus: The implication for us as data scientists is that we should be looking beyond static features and trying to capture these nonlinear dynamics using tools like DMD and Koopman operators when dealing with complex signals like EEG.
Yuki: For me, it reinforces the idea that population-level studies need to look at these dynamic properties, not just static averages, because if the underlying dynamical structure is stable across populations, then we have a stronger basis for understanding the condition’s biological roots.
Ines: It’s about moving from simply finding patterns to understanding the rules governing those patterns in a way that remains consistent across different subjects and platforms. That's what this work offers.
Marcus: So, to wrap up this discussion on "Koopman Operator Analysis of EEG Microstate Dynamics Reveals a Cross-Dataset, Hearing-Level Adjusted Signature of Tinnitus," the paper shows that the stability of oscillation magnitude is a robust biomarker because it generalizes across datasets with a mean consistency ratio of rho equal to zero point six eighty five <ref:2604.22116#pg3>.
Yuki: And this robustness means we can have more confidence in using these biomarkers for clinical applications, even when the data comes from a different source than what was used for training.
Ines: It’s about finding that consistent dynamical signature that decouples the neural activity from the specific acquisition hardware or subject population noise. We need to see how this translates into actual diagnostic tools down the line, because right now it’s just a strong mathematical prediction based on these cross-dataset metrics.
Conclusion: Ines: So we're wrapping up our look at this paper on Koopman Operator Analysis of EEG Microstate Dynamics Reveals a Cross-Dataset, Hearing-Level Adjusted Signature of Tinnitus and what all that means for finding these neural signatures.
Marcus: Basically, the authors are saying they took two different datasets and tried to see if they could build a classification system that worked across those different sources without retraining.
Yuki: And the key finding is that by focusing on the stability of those underlying oscillation magnitudes, which we call eigenvalues, you get a much more consistent signature than looking at the frequency shifts alone.
Ines: Exactly. They found this stability metric actually works well in both directions—trying to predict tinnitus from one dataset to another and vice versa.
Marcus: The numbers they got are pretty strong here; they hit an AUROC of zero point seven seven one nine in one direction, and an accuracy of zero point seven six two three in the other.
Yuki: That consistency across different acquisition settings is important because it suggests this isn't just a fluke tied to a specific machine or a narrow patient group.
Ines: It points toward something more fundamental about how the brain's rhythms are persisting even when you change the data source.
Marcus: The authors also mentioned that they found this stability metric is consistent across datasets with a mean consistency ratio of rho equal to zero point six eighty five, which is pretty solid.
Yuki: And they linked that spatial pattern—the most discriminative part of the principal component—to known brain areas like the auditory cortex or fronto-parietal networks.
Ines: So what this really changes for us is moving away from simple static measurements and using these dynamic tools to describe the underlying rules governing those neural signatures.
Marcus: It means we should be looking for these kinds of dynamical features when we analyze complex signals like EEG in large cohorts.
Yuki: And that opens up avenues for understanding tinnitus not just as a symptom, but as a condition with consistent underlying neural dynamics across populations.
University of California San Diego
q-bio.NC, eess.SP
Submitted: 2026-04-23
Updated: 2026-11-16
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 82/100
The gist: Resting-state EEG biomarkers of tinnitus robust to cross-subject and cross-platform variation provide a way to investigate neural dynamics associated with tinnitus The PCA-based Koopman features…
Key concepts
- Microstate Analysis
- This method uses k-means clustering to group EEG data into distinct topographic states. It then analyzes the transitions between these states and how long they persist. This helps identify specific, stable patterns of brain activity associated with tinnitus.
- Koopman Operator Analysis
- This technique uses Hankel-embedded Dynamic Mode Decomposition (DMD) to extract complex eigenvalues from EEG data. The magnitude of these eigenvalues reflects the stability of neural oscillations, while the phase relates to their frequency, providing a rich mathematical description of brain dynamics.
- Eigenvalue Magnitude
- The magnitude of the Koopman eigenvalue is considered a robust tinnitus biomarker because modes with values near one indicate persistent, stable energy changes in time. This stability suggests that these oscillatory patterns are genuine neural signatures related to tinnitus rather than temporary noise or acquisition artifacts.
Terminology
Summary
Resting-state EEG biomarkers of tinnitus robust to cross-subject and cross-platform variation provide a way to investigate neural dynamics associated with tinnitus The PCA-based Koopman features yield the strongest cross-dataset metrics across both transfer directions, outperforming microstate-derived features
Feature Extraction Frameworks
The study systematically evaluates two feature extraction frameworks applied to resting-state EEG from three independently collected datasets for binary classification of tinnitus The methods employed include microstate analysis and Koopman operator analysis
-
Microstate Analysis involves extracting topographic states using the k-means clustering algorithm, which results in transition probability and state duration features
-
Koopman Operator Analysis utilizes Hankel-embedded Dynamic Mode Decomposition (DMD) to extract complex eigenvalues, where the magnitude encodes oscillation stability and the phase encodes oscillation frequency
Cross-Dataset Generalization Performance
The classification performance was assessed in a strict cross-dataset paradigm where a linear support vector machine is trained on one dataset and evaluated on another without retraining The results demonstrate that PCA-based Koopman features yield the strongest cross-dataset metrics across both transfer directions, achieving an AUROC of 0.7719 in D1 → D2 and an Accuracy of 0. p. 27623 in D2 → D1 The Wasserstein consistency analysis further reveals that Koopman eigenvalue magnitude generalizes across datasets with a mean consistency ratio of ρ¯ = 0. p. 6685, whereas the eigenvalue phase does not generalize with a mean ratio of ρ¯ = 1. p. 6583
Biomarker Interpretation
The analysis provides interpretable evidence that altered oscillatory decay rates, rather than frequency shifts, constitute the more robust tinnitus biomarker This is supported by the observation that the eigenvalue magnitude distribution shows a concentration near the unit circle, indicating persistent, stable energy changes within time windows The complex-valued nature of Koopman modes allows for capturing both amplitude and phase information of underlying neural oscillations Modes with λ ≈ 1 are assigned high reliability scores, suggesting that these modes represent genuine tinnitus-related neural signatures decoupled from acquisition-specific artifacts
Conclusion on Robustness
The study concludes that the oscillation stability (eigenvalue magnitude) is the more reliable and generalizable tinnitus biomarker This robustness is confirmed because the consistency ratio ρ = Wcross-dataset / Wwithin-dataset < 1 indicates that class-related distributional shifts in oscillation stability generalize across datasets The spatial analysis of the most discriminative PCA component reveals a predicted spatial pattern of the tinnitus-associated dynamical signature, linking it to known neuroanatomical substrates like auditory cortex or fronto-parietal networks
Limitations
Several limitations were noted in the study, including relatively small clinical datasets which reduces the reliability of estimated classification metrics and makes the D2 → D1 direction more susceptible to high-variance performance estimates The classifier is restricted to a linear SVM, which may not capture higher-order interactions between feature dimensions Furthermore, the current analysis is restricted to binary tinnitus versus control classification and does not address tinnitus severity grading or subtype identification, which would be clinically relevant applications The DMD embedding features were computed for various delay and lag parameters but their full potential warrants further investigation
References
[1] J. M. Bhatt, H. W. Lin et al., “Prevalence, Severity, Exposures, and Treatment Patterns of Tinnitus in the United States,” JAMA Otolaryngology–Head & Neck Surgery, vol. p. 6 142, no. 10, pp. 959–965, Oct. p. 6 2016
[3] E. Van Der Loo, S. p. 6 Gais et al. p. 6, “Tinnitus Intensity Dependent Gamma Oscillations of the Contralateral Auditory Cortex,” PLoS ONE, vol. p. 6 4, no. 10, p. e7396, Oct. p. 6 2009
[4] N. Weisz, S. Moratti et al. p. 6, “Tinnitus Perception and Distress Is Related to Abnormal Spontaneous Brain Activity as Measured by Magnetoencephalography,” PLoS Medicine, vol. p. 6 2, no. 6, p. e153, Jun. 2005
[5] R. R. Llinas, U. Ribary ´ et al. p. 6, “Thalamocortical dysrhythmia: A neurological and neuropsychiatric syndrome characterized by magnetoencephalography,” Proceedings of the National Academy of Sciences, vol. p. 6 96, no. 26, pp. 15 222–15 227, Dec. p. 6 1999
[6] D. De Ridder, A. p. 6 B. Elgoyhen et al. p. 6, “Phantom percepts: Tinnitus and pain as persisting aversive memory networks,” Proceedings of the National Academy of Sciences, vol. p. 6 108, no. 20, pp. 8075–8080, May 2011
[7] D. Lehmann, H. p. 6 p. 6 Ozaki et al. p. 6 p. 6, “EEG alpha map series: brain micro-states by space-oriented adaptive segmentation,” Electroencephalography and Clinical Neurophysiology, vol. p. 6 p. 6 67, no. 3, pp. 271–288, Sep. p. 6 p. 6 1987
[8] A. J. Norena, “An integrative model of tinnitus based on a central gain controlling neural sensitivity,” Neuroscience & Biobehavioral Reviews, vol. p. 6 35, no. 5, pp. 1089–1109, Apr. p. 6 2011
[9] K. Nidal and A. p. 6 S. Malik, EEG/ERP Analysis: Methods and Applications
[10] E. Cardon, I. Joossen et al. p. 6, “Systematic review and meta-analysis of late auditory evoked potentials as a candidate biomarker in the assessment of tinnitus,” PLOS ONE, vol. p. 6 15, no. 12, p. e0243785, Dec. p. 6 2020
[11] D. Lehmann, H. Ozaki et al., “EEG alpha map series: brain micro-states by space-oriented adaptive segmentation,” Electroencephalography and Clinical Neurophysiology, vol. 67, no. 3, pp. 271–288, Sep. 1987
[12] C. M. Michel and T. p. 6 Koenig, “EEG microstates as a tool for studying the temporal dynamics of whole-brain neuronal networks: A review,” NeuroImage, vol. p. 6 180, pp. 577–593, Oct. p. 6 2018
[13] Y. Cai, S. Chen et al. p. 6, “Altered Resting-State EEG Microstate in Idiopathic Sudden Sensorineural Hearing Loss Patients With Tinnitus,” Frontiers in Neuroscience, vol. p. 6 13, May 2019
[14] Z. Raeisi, A. Sodagartojgi et al. p. 6, “Enhanced classification of tinnitus patients using EEG microstates and deep learning techniques,” Scientific Reports, vol. p. 6 15, no. 1, p. 15959, May 2025
[15] B. O. Koopman, “Hamiltonian Systems and Transformation in Hilbert Space,” Proceedings of the National Academy of Sciences, vol. p. 6 17, no. 5, pp. 315–318, May 1931
[16] I. Mezic, “Spectral Properties of Dynamical Systems, Model Reduction ´ and Decompositions,” Nonlinear Dynamics, vol. p. 6 41, no. 1, pp. 309–325, Aug. p. 6 2005
[17] M. Budisiˇ c, R. p. 6 Mohr ´ et al. p. 6, “Applied Koopmanisma),” Chaos: An Interdisciplinary Journal of Nonlinear Science, vol. p. 6 22, no. 4, p. 047510, Dec. p. 6 2012
[18] M. Wanner and I. p. 6 Mezic, “Robust Approximation of the Stochastic Koopman Operator,” SIAM Journal on Applied Dynamical Systems, vol. p. 6 21, no. 3, pp. 1930–1951, Sep. p. 6 2022
[19] S. Qian and C. p. 6-A. Chou, “A Koopman-operator-theoretical approach for anomaly recognition and detection of multi-variate EEG system,” Biomedical Signal Processing and Control, vol. p. 6 69, p. 102911, Aug. p. 6 2021
[20] N. Marrouch, J. p. 6 Slawinska et al. p. 6, “Data-driven Koopman operator approach for computational neuroscience,” Annals of Mathematics and Artificial Intelligence, vol. p. 6 88, no. 11, pp. 1155–1173, Dec. p. 6 2020
[21] C.-D. Wang, X.-R. Zhu et al., “Cross-Subject Tinnitus Diagnosis Based on Multi-Band EEG Contrastive Representation Learning,” IEEE Journal of Biomedical and Health Informatics, vol. p. 6 27, no. 7, pp. 3187–3197, Jul. p. 6 2023
[22] A. R. Cuevas-Romero, L. p. 6 M. Alonso-Valerdi et al. p. 6, “An Electroencephalography-based Database for studying the Effects of Acoustic Therapies for Tinnitus Treatment,” Scientific Data, vol. p. 6 9, no. 1, p. 500, Aug. 2022
[23] M. X. Cohen, “A tutorial on generalized eigendecomposition for denoising, contrast enhancement, and dimension reduction in multichannel electrophysiology,” NeuroImage, vol. p. 6 247, p. 118809, Feb. p. 6 2022
[24] F. Takens, “Detecting strange attractors in turbulence,” in Dynamical Systems and Turbulence, Warwick 1980, D. p. 6 Rand and L. p. 6-S. Young, Eds. p. 6, Berlin, Heidelberg: Springer, 1981
[25] H. Arbabi and I. p. 6 Mezic, “Ergodic Theory, Dynamic Mode Decomposi- ´tion, and Computation of Spectral Properties of the Koopman Operator,” SIAM Journal on Applied Dynamical Systems, vol. p. 6 16, no. 4, pp. 2096–2126, Jan. p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Robust Cross-Dataset Tinnitus Classification: The improved system can classify tinnitus robustly across different EEG recording conditions by leveraging PCA-based Koopman eigenvalue magnitude features, which
yield the strongest cross-dataset metrics in both transfer directions
(AUROC = 0.7719 in D1 -> D2 and Accuracy = 0.7623 in D2 -> D1). -
Biomarker Specificity through Oscillation Stability: The system can identify a more reliable biomarker by focusing on oscillation stability rather than frequency shifts, as
altered oscillatory decay rates, rather than frequency shifts, constitute the more robust tinnitus biomarker.
-
Enhanced Feature Reliability via Consistency Analysis: The improved system utilizes a Wasserstein consistency analysis to confirm that
Koopman eigenvalue magnitude, encoding oscillation stability, generalizes across datasets (rho bar = 0.685), whereas eigenvalue phase, encoding oscillation frequency, does not (rho bar = 1.583).
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks