WorksEngineering
Reading Across Space, Time and Frequencies: Tensor Decomposition of SEEG Signals to Support Epileptogenic Zone Definition
Abstract
Stereo-electroencephalography (SEEG) is an invasive technique used in the presurgical evaluation of drug-resistant epilepsy to support the localization of the Epileptogenic Zone (EZ). SEEG data are particularly complex to analyze, as they include hundreds of components, named channels, each associated with a different brain site, with interactions varying in both time and frequency. In this study, we propose a new framework for SEEG analysis, based on the Canonical Polyadic (CP) decomposition, to extract the latent modes underlying seizure generation. We model SEEG signals as a third-order tensor (channel x time x frequency), and factorize the tensor into rank-1 factors. We then exploit the channel-specific factor as an unsupervised feature set for identifying the EZ. The framework is validated on SEEG-recorded seizures from 35 patients with drug-resistant epilepsy. The results show that the joint combination of multiple frequency bands significantly enhances classification performance.
Introduction
The goal of epilepsy surgery is to resect the Epileptogenic Zone (EZ), defined as the cortical area indispensable for the generation of seizures [1]. When non-invasive evaluation is inconclusive, EZ location can be inferred via Stereo-Electroencephalography (SEEG), an invasive procedure to record electrical activity from cortical and subcortical structures through intracranial electrodes with multiple contacts. The result is a multi-dimensional signal, in which each channel reflects the activity of a brain site. In clinical practice, SEEG recordings are reviewed by highly specialized neurophysiologists, with particular attention to the transition from the inter-ictal (between seizures) to the ictal period (during seizures) [2]. This analysis, based on visual inspection, allows one to identify the Seizure Onset Zone (SOZ), the region where seizures originate, and a key proxy for the EZ.
In recent years, many techniques have been developed to support SEEG interpretation, analyzing the spectral energy of single channels or functional connectivity between them. The second option is particularly relevant given that seizures can originate from the interaction of distributed brain structures [3]. In particular, SEEG signals can be modeled as time-varying graphs, where nodes represent channels and edges represent connectivity relationships. A variety of techniques have been proposed to estimate SEEG connectivity, with results that change depending on the frequency band analyzed. The SOZ channels tend to desynchronize within the first seconds of seizure, considering slow oscillations [4], and increase high-frequency connectivity as ictal discharges propagate through the brain [5]. Hence, integrating connectivity information across different frequency bands is fundamental to provide a comprehensive description of SOZ.
To cope with the inherent complexity of SEEG data, we can apply dimensionality reduction techniques to discover latent patterns behind brain activity. In this context, [6] proposed a joint matrix factorization to identify connectivity subgraphs shared across different seizures. More recently, [7] applied Dynamic Mode Decomposition (DMD) to separate frequency-specific modes involved in seizure generation. However, these approaches do not integrate information from different oscillations: in [6] connectivity is computed in a single frequency band, whereas in [7] frequency is only used to sort the decomposition modes. To jointly capture the effect of different frequency components, SEEG signals can be represented as tensors whose modes are nodes (or links), time, and frequency [8, 9]. This strategy was explored in [10], which builds wavelet-power tensors of inter-ictal SEEG activity, and in [11], which considers connectivity tensors during seizures. Nevertheless, these studies were validated in very small cohorts, and none of them assessed whether tensor modeling can support the identification of the EZ.
Inspired by the previous works in the literature, we develop a framework where the features of SEEG signals are modeled as a channel time frequency tensor. We then apply the Canonical Polyadic / Parallel Factors (CP/PARAFAC) decomposition to factorize the tensor into a reduced number of components. The factors associated with the spatial dimension are then exploited as an unsupervised feature set to identify the SOZ channels. We validate the framework on ictal data from 35 patients with drug-resistant epilepsy, who underwent SEEG monitoring at La Timone Hospital (Marseille, France). To the best of our knowledge, this is the first study applying tensor decomposition to ictal SEEG data from a large patient cohort. Our results show that the integration of multiple frequencies into a single tensor model improves classification performance, suggesting that our framework is a promising tool for analysis of SEEG signals.
System Model
Signal Model
We model an SEEG recording as a multivariate discrete-time signal , , where is the value of channel at sample . Given the sampling frequency and the recording duration , the total number of samples is . We consider a set of frequency bands, with each band defined by the interval in Hz, with . Hence, we denote by the signal of channel filtered in band . The signal is segmented into overlapping windows of duration , with a time shift between consecutive windows. We choose and as multiples of the sampling period , so that each window contains an integer number of samples , and consecutive windows are spaced by samples. The total number of windows is therefore , and window contains the samples from to . Finally, we denote by the -th sample of channel within the -th window for band , i.e., , with .
From the windowed signal, we now extract a set of features that characterize it through both local and connectivity-based measures. First, for each channel , window and frequency band , we compute the energy
Then, we compute the instantaneous phase of as the argument of its analytic signal:
where denotes the Hilbert transform [12]. We partition the phase domain into uniform bins , according to Sturges’ rule [13], and compute the entropy of channel , window , and band , as
where is the empirical probability that the phase samples fall into bin . Finally, for window and a frequency band , we measure the connectivity between channels and through the Mutual Information (MI) between the phase distributions:
where and are defined as above, and is the joint entropy of the two distributions [14].
Tensor Decomposition
We consider three tensor models of the SEEG signal, depending on the feature used to characterize it:
an energy tensor , whose entries are the signal energies of each channel, window, and frequency band, i.e., ;
an entropy tensor , whose entries are the phase entropies, i.e., ;
a MI tensor , whose entries are the functional connectivity of each channel pair, window and frequency band, i.e., , where and .
To ease the reading, in the following we refer to and , but the same formulation applies to , with the first mode indexing channel pairs instead of channels. Following the CP/PARAFAC approach [15], we approximate the tensor as a sum of non-negative rank-1 tensors:
where denotes the vector outer product, , , and . This representation decomposes the SEEG signal into the superposition of latent brain components. The weight quantifies the importance of component , while the vectors , and are its spatial (channel), temporal (window) and spectral (frequency) signatures.
To solve (5), we employ the Alternating Least Square (ALS) algorithm. We first collect the factor vectors of the components into the factor matrices , and , whose -th columns are , and , respectively. The algorithm alternately updates each factor matrix, while the other two are fixed. In particular, the update of is
where is the Khatri-Rao product, is the Hadamard (element-wise) product, and is the mode-0 unfolding of the tensor, defined as . The matrices and are updated analogously, using the corresponding unfoldings of . Hence, each column of the factor matrices is normalized to unit norm, and the norms are absorbed into the weights:
The updates (6)–(7) are repeated until convergence, i.e., until the relative change of the reconstruction error falls below a predefined tolerance.
Unsupervised Feature Extraction
We assume that the factors capture the spatial organization of the SEEG signal, and can therefore be used to characterize epileptogenic channels. In the energy-based tensor, each channel is associated with the vector , describing the participation of in the latent components of the signal. Hence, we can train a classifier that takes as input and outputs the class (EZ vs non-EZ) of channel . In the case of the entropy-based tensor, we follow the same approach, using as the classification input. We note that the dimensions of and do not depend on the number of windows or frequency bands , but only on the rank .
With the MI-based tensor, the rows of are associated with channel pairs, so there is no one-to-one correspondence between channels and the decomposition output. To obtain channel-level features, we aggregate the links incident to each channel. Let be the set of links involving channel . For each component and each percentile threshold , we define the thresholded weighted degree of as
where is the indicator function and is the -th percentile of the links associated with the -th component. The unsupervised feature vector is finally obtained as
A schematic representation of the proposed methodological pipeline is shown in Figure 1.

Figure 1. Methodological pipeline.
Experimental Results
Dataset and Settings
We retrospectively analyzed SEEG data of 35 patients from the EPINOV database of La Timone Hospital, (Marseille, France). The patients were selected according to the following inclusion criterion: at least one seizure starting with low-voltage fast activity. The signals were sampled at Hz and then converted to a bipolar montage. For each patient, we selected an epoch of s spanning 35 s before and 15 s after the electrographic seizure onset, i.e., the start of fast oscillations. We segment each epoch into windows lasting s, with a shift of s.
Following the model in Sec. 2.1, we computed the energy , entropy , and connectivity for each channel , link , window , and frequency band , where , , , , , Hz. To compute ictal-related changes for each feature, we applied a -score normalization using the first 30 s (i.e., the first 120 windows) of each recording to estimate the mean (E) and standard deviation (STD) of each feature. For energy, this yields
with analogous expressions for entropy and MI.
Following the model in Sec. 2.2, we then constructed the tensors , , and from the remaining 20 s of each epoch, with the seizure starting after 5 s. The tensor dimensions are determined by the number of channels, which is patient-specific, the number of windows, and the number of frequency bands. In our experiments, we vary the portion of the ictal signal analyzed and the composition of the frequency band set , obtaining different results for and , where denotes the last window of the epoch. Finally, we set the rank of the decomposition to , and the percentiles used to compute the weighted degrees to .
We used the unsupervised feature vectors , , and as inputs to a logistic regression classifier to identify channels labeled as part of the SOZ. Ground truth was defined by the clinical annotations provided by neurophysiologists following visual inspection of the SEEG recordings. Classification performance was assessed using the Area Under the Receiver Operating Characteristic Curve (AUROC) and the Area Under the Precision-Recall Curve (AUPRC). Our classification task presents strongly imbalanced classes and, thus, the AUPRC is a preferable metric for system evaluation[16]. Since the goal of this study is to assess the intrinsic discriminative power of the unsupervised features, performance was evaluated in a within-patient setting, while cross-patient generalization is left for future work.

Figure 3. Time-averaged performance distribution in the patient cohort for different frequency sets .
Numerical Results
We first consider the edge scenario where the feature set includes a single frequency band, so that the tensor has shape for energy and entropy, and for MI. Figure 2 reports the AUPRC, averaged across the patient cohort, as a function of the feature used to build the tensor and of the extension of the signal analyzed. We note that performance exhibits different trends depending on the feature used to build the tensor. The energy-based model achieves its best results during the first seconds of the seizure (which starts after 5 s), particularly when using the and the bands. Instead, the entropy-based model performs best in the low portion of the frequency spectrum and at later stages of the seizure. Finally, the MI-based model yields the highest AUPRC overall, with maxima in the and bands and approximately 5 s after seizure onset.

Figure 2. Average PR-AUC in the patient cohort as a function of time, using a single frequency band in the tensor model.
Fig. 3 reports the distribution across patients of the time-averaged AUROC and AUPRC as a function of the frequency bands included in the tensor. Each distribution is shown as a box plot, where the central line is the median, the edges are the first and third quartiles, and the whiskers are the 5-th and 95-th percentiles. We denote by the configuration with the and bands and, for increasing , the configurations obtained by adding one band at a time. For the energy-based tensor, the best performance is obtained with , while including additional bands does not improve either AUROC or AUPRC. The entropy-based tensor shows the opposite trend: the best performance is obtained with the low-frequency bands, and it decreases monotonically as higher portions of the spectrum are included.
For the MI-based tensor, the performance trends are similar to those of entropy, even though with higher AUROC and AUPRC values in all frequency configurations. In this case, combining multiple bands yields more robust unsupervised features: all the multi-band configurations improve the AUROC and AUPRC in both the upper tails with respect to the single-band models. In particular, the configuration integrating all frequencies () reaches mean AUROC and AUPRC values of 0.887 and 0.655, respectively. Performing a one-sided Wilcoxon signed-rank test [17], with the alternative hypothesis that achieves a higher AUPRC than single-band tensors, we obtain significant differences in all comparisons. The best single-band configuration () is outperformed with a p-value , while all remaining configurations are outperformed with .
Conclusions
In this work, we proposed a new framework to model SEEG signals across space (channels), time, and frequency. We built three SEEG tensor models that characterize the signal through its energy, entropy, and functional connectivity. We applied the CP/PARAFAC approach to decompose the tensor into rank-1 factors, and exploited the spatial factor as an unsupervised feature set for EZ definition. The framework was evaluated on a dataset of 35 patients who underwent SEEG monitoring at La Timone Hospital (Marseille, France).
Our results show that the performance trends across frequency bands depend on the feature used to build the tensor, and that integrating multiple frequency bands in the same model improves the identification of the EZ in terms of AUROC and AUPRC. Future work will address the validation of the framework on an independent patient cohort, using cross-validation strategies to assess the generalizability of the results beyond the within-patient settings.
Compliance with Ethical Standards
This retrospective study exclusively used pseudonymised data acquired as part of EPINOV (NCT03643016) prospective study. All subjects gave their informed written consent to participate in the study according to the Declaration of Helsinki. In accordance with French regulations on retrospective research using health data, all patients received written information about the reuse of their data for research and did not object (non-opposition procedure). No potentially identifying information is reported.
Acknowledgments
Part of this work was supported by the European Union under the Marie Skłodowska-Curie Actions (MSCA) postdoctoral fellowships program of Horizon Europe, Grant Agreement No. 101270737 (HYPHEN). Part of this work received funding from the French National Research Agency (ANR) under the second “Investissements d’Avenir” program (ANR-17-RHUS-0004, EPINOV).
References
- [1]F. Rosenow and H. Lüders, “Presurgical evaluation of epilepsy,” Brain, vol. 124, no. 9, pp. 1683–1700, September 2001.pubmed.ncbi.nlm.nih.gov/11522572
- [2]F. Bartolomei, A. Nica, M. P. Valenti-Hirsch, C. Adam, and M. Donneulle, “Interpretation of SEEG recordings,” Neurophysiologie clinique, vol. 48, no. 1, pp. 53–57, 2018.DOI
- [3]F. Bartolomei et al., “Defining epileptogenic networks: contribution of SEEG and signal analysis,” Epilepsia, vol. 58, no. 7, pp. 1131–1147, 2017.DOI
- [4]F. Mason, L. Ferri, L. Di Vito, L. Alvisi, L. Zanuttini, M. Martinoni, R. Mai, F. Cardinale, P. Tinuper, R. Michelucci et al., “Desynchronization index: A new connectivity approach for exploring epileptogenic networks,” Biomedical Signal Processing and Control, vol. 127, p. 111166, 2026.DOI
- [5]F. Wendling, F. Bartolomei, J.-J. Bellanger, J. Bourien, and P. Chauvel, “Epileptic fast intracerebral EEG activity: evidence for spatial decorrelation at seizure onset,” Brain, vol. 126, no. 6, pp. 1449–1459, 2003.pubmed.ncbi.nlm.nih.gov/12764064
- [6]G. Frusque, P. Borgnat, P. Gonçalves, and J. Jung, “Semi-automatic extraction of functional dynamic networks describing patient’s epileptic seizures,” Frontiers in Neurology, vol. 11, p. 579725, 2020.DOI
- [7]A. Nieto Ramos, B. Krishnan, A. V. Alexopoulos, W. Bingaman, I. Najm, J. C. Bulacio, and D. Serletis, “Epileptic network identification: insights from dynamic mode decomposition of sEEG data,” Journal of neural engineering, vol. 21, no. 4, p. 046061, 2024.DOI
- [8]F. Cong, Q.-H. Lin, L.-D. Kuang, X.-F. Gong, P. Astikainen, and T. Ristaniemi, “Tensor decomposition of eeg signals: A brief review,” Journal of neuroscience methods, vol. 248, pp. 59–69, 2015.DOI
- [9]N. D. Sidiropoulos, L. De Lathauwer, X. Fu, K. Huang, E. E. Papalexakis, and C. Faloutsos, “Tensor decomposition for signal processing and machine learning,” IEEE Transactions on signal processing, vol. 65, no. 13, pp. 3551–3582, 2017.DOI
- [10]J. Li, J. P. Haldar, J. C. Mosher, D. R. Nair, J. A. Gonzalez-Martinez, and R. M. Leahy, “Scalable and robust tensor decomposition of spontaneous stereotactic eeg data,” IEEE Transactions on Biomedical Engineering, vol. 66, no. 6, pp. 1549–1558, 2018.DOI
- [11]P.-A. Chantal, A. Karfoul, A. Nica, and R. Le Bouquin Jeannes, “Dynamic brain effective connectivity analysis based on low-rank canonical polyadic decomposition: application to epilepsy,” Medical & Biological Engineering & Computing, vol. 59, no. 5, pp. 1081–1098, 2021.DOI
- [12]A. V. Oppenheim, R. W. Schafer, J. R. Buck et al., “Discrete-time signal processing,” 1989.
- [13]H. A. Sturges, “The choice of a class interval,” Journal of the american statistical association, vol. 21, no. 153, pp. 65–66, 1926.DOI
- [14]C. E. Shannon, “A mathematical theory of communication,” The Bell system technical journal, vol. 27, no. 3, pp. 379–423, 1948.
- [15]R. Bro, “Parafac, tutorial and applications,” Chemometrics and intelligent laboratory systems, vol. 38, no. 2, pp. 149–171, 1997.DOI
- [16]T. Fawcett, “An introduction to ROC analysis,” Pattern recognition letters, vol. 27, no. 8, pp. 861–874, 2006.DOI
- [17]F. Wilcoxon, “Individual comparisons by ranking methods,” in Breakthroughs in statistics: Methodology and distribution. Springer, 1992, pp. 196–202.