Seizure Onset Zone Localization

• Seizure Onset Zone Localization is the process of identifying the minimal brain regions responsible for initiating epileptic seizures, crucial for effective surgical interventions.
• It employs both noninvasive (scalp EEG, fMRI) and invasive (iEEG/SEEG) methods to capture diverse electrophysiological and imaging signatures of seizure onset.
• Advanced techniques like deep learning, Bayesian modeling, and network science improve accuracy by addressing signal noise, label ambiguity, and multimodal data fusion challenges.

Seizure onset zone (SOZ) localization refers to the identification of brain regions where epileptic seizures originate. Accurate SOZ localization is essential for surgical planning in drug-resistant epilepsy, directly impacting patient outcomes. The process spans noninvasive and invasive modalities—scalp EEG, iEEG/SEEG, electrical stimulation, fMRI, and multimodal data fusion—each presenting unique algorithmic, statistical, and interpretability challenges. Recent research integrates deep learning, Bayesian modeling, network science, and expert-embedded approaches, with increasingly rigorous evaluation on large, well-annotated datasets.

1. Fundamental Principles and Definitions

The seizure onset zone is the minimal cortical territory whose removal leads to cessation of clinical seizures. SOZ localization operationally seeks EEG channel sets, imaging-derived regions, or network nodes with the earliest and most characteristic pre-ictal and ictal transitions. Gold-standard ground truth is typically provided by expert review of invasive EEG during spontaneous seizures, sometimes validated by postsurgical seizure-freedom (e.g., Engel I outcomes). Localization methods must navigate signal noise, spatial sampling biases, heterogeneous onset patterns (LVFA, suppression, slow waves), and annotation ambiguities.

Key desiderata in SOZ localization algorithms include:

• Robustness to label noise and inter-rater disagreement
• Quantitative performance under varying SNR and label segmentation
• Integrated explainability with channel- or region-level saliency
• Multimodal fusion capacity (EEG, MRI, fMRI, stimulation, connectomics)
• Statistically validated correspondence to clinical outcomes

2. Noninvasive and Invasive Electrophysiological Approaches

Scalp EEG-Based Methods

Scalp EEG-based SOZ localization is challenged by low spatial resolution and attenuation of deep sources. Statistical frameworks such as BUNDL apply Bayesian uncertainty-aware deep learning with a KL-divergence loss, directly modeling mismatches between noisy clinician labels and latent clean labels. MC-dropout–derived per-sample epistemic uncertainty weights the influence of ambiguous labels, mitigating the negative impact of annotation noise on both detection and subsequent localization. DeepSOZ, a backbone model incorporating Transformer encoders, LSTM, and attention pooling, outputs channel-wise attention scores at onset, used post-thresholding to identify SOZ channels. Compared to baseline cross-entropy training, BUNDL delivers a 3.2 percentage point improvement in seizure-level SOZ accuracy and a 2.9 point improvement at the patient level in the TUH dataset, notably by reducing detection-linked false positives propagating to localization errors. Robustness holds over multiple noise regimes, SNR settings, and label segmentation errors (Shama et al., 17 Oct 2024).

Scalp EEG using blind source separation (ICA) and dipole modeling allows the isolation of ictal sources via independent component analysis (ICA) and localization of their equivalent current dipole (ECD) (Barborica et al., 2021). ECD fits provide sublobar "first guess" localization with accuracy as high as 10 mm for superficial foci, but with errors increasing (to 47 mm mean) for deeper mesial structures. Validation with simultaneous SEEG using wavelet coherence demonstrates the dependency of noninvasive visibility on SOZ depth and the necessity of individualized modeling.

Intracranial EEG: Algorithmic Diversity and Methodological Pitfalls

Modern SOZ extraction from iEEG/SEEG employs a diverse algorithmic landscape. Recent comparative work establishes that seemingly marginal algorithmic decisions—choice of baseline (fixed channel-specific vs moving or global), frequency range (inclusion/exclusion of low-frequency components), treatment of electrodecrements (decreases in activity)—substantially alter the identified onset channels, yielding only minimal overlap (1.8–2.9%) between representative algorithms: Imprint (deviation scores in line-length and band-power), Epileptogenicity Index (band-ratio + change-point detection), and Low Entropy Map (time-frequency activation entropy) (Gascoigne et al., 17 Oct 2024). Quantified effect sizes (r>0.3) and low within-subject kappa (κ<0.5) indicate low cross-algorithm stability, mandating explicit methodological transparency and, when possible, fusion of complementary markers for robust mapping.

For interictal iEEG, AI-based frameworks that combine multiple biomarkers (HFOs, IEDs, PAC) and their temporal features in an RBF-SVM, rather than relying on single-rate or hard-coded thresholds, yield robust SOZ discrimination (AUC = 0.73 in subjectwise cross-validation, sensitivity 57.5%, specificity 79.5%). Recording durations of 90–120 min are sufficient to converge to these accuracies (Varatharajah et al., 2018).

3. Graph-Based, Network-theoretic, and Bayesian Approaches

SOZ mapping increasingly leverages brain network theory. High-dimensional state-space or MAR models, equipped with stochastic block-model community priors, jointly infer both directed connectivity and latent region clusters, enabling identification of both SOZ and its community-level context (Li et al., 2020, Wang et al., 2022). SOZ localization is then performed by ranking nodes with maximal post-ictal increases in average outgoing connectivity (ADC), controlling false positives by nonparametric null calibration. In real iEEG, these approaches achieve 100% TPR and ≤3% FPR, outperforming standard MAR, PDC, and glasso benchmarks.

Recent graph neural network (GNN) architectures treat SEEG as a dynamic functional graph (nodes: contacts; edges: e.g., Pearson or DTF connectivity). Dual contrastive learning frameworks, such as Seizure-NGCLNet, execute global and local mutual information maximization to encode spatial pathological patterns. Fine-tuned with top-k localized GAT, attention weights directly highlight SOZ channels, achieving >85% concordance with expert labels and performance >95% on seizure detection and SOZ localization (Wang et al., 19 Nov 2025). Classical GNNs with multi-scale attention mechanisms also rank SOZ-associated contacts, with feature attribution analyses and relationship to postsurgical outcome (Agaronyan et al., 21 Feb 2025, Amir et al., 29 May 2025).

4. Structural, Functional, and Multimodal Imaging Fusion

Structural MRI (T1, dMRI) and fMRI are increasingly used to complement electrophysiological data. Structural connectivity abnormalities are derived as fractional anisotropy (FA) z-scores per region and combined with functional iEEG z-scores via decision trees. Retrospective multicenter analysis demonstrates that patients in whom maximal abnormalities (structural and/or functional) were resected had 15 times higher odds of seizure-freedom postoperatively (p=0.008); combining both modalities in a 2-split CART classified patient outcomes with 84% accuracy (Horsley et al., 2023).

Resting-state fMRI applied to SOZ localization, particularly in pediatric PRE cohorts, benefits substantially from hybrid frameworks uniting deep learning (2D CNNs on ICA spatial maps) and expert knowledge modules (feature-based SVMs using activation topology and morphology) (Kamboj et al., 2023, Kamboj et al., 2023). Explicit incorporation of features such as white matter extension into ventricles and multi-cluster spatial asymmetry is essential for high precision (e.g., 91.7% F1, 84.8% accuracy; ablations show –27% F1 if "white matter overlap" is omitted). These approaches are robust to class imbalance and greatly reduce expert IC-sorting workload, facilitating application as a low-cost screening tool.

Spatio-temporal ICA on rs-fMRI, with sorting procedures based on frequency, lateralization, local network measures, and non-Gaussianity, achieves 90% lobar concordance with postsurgical resection in TLE (Sadjadi et al., 3 Jan 2025). Structured joint matrix-tensor factorization (sCMTF) on EEG-fMRI simultaneously extracts spatial, spectral, and temporal signatures, and region-specific HRF biomarkers, localizing SOZ in all tested patients (each by at least one metric), and outperforming standard GLM EEG-fMRI (Eyndhoven et al., 2020).

5. SOZ Localization via Electrical Stimulation and Machine Learning

Active probing modalities, such as SPES and CCEPs, provide yet another axis of SOZ information. In SPES, deep learning classifiers (multi-scale 1D ResNet CNNs and cross-channel Transformers) aggregate evoked potentials under divergent (outward) and convergent (inward) paradigms. The convergent approach, analyzing how a candidate response site reacts to stimulation elsewhere, significantly outperforms classic divergent approaches (AUROC 0.666 vs 0.574), and cross-channel Transformers—able to integrate variable electrode positions—further increase AUROC to 0.730–0.745. The self-attention operator confers both adaptivity to arbitrary layouts and sensitivity to spatially distributed SOZ patterns (Norris et al., 29 Mar 2024).

CCEP-based SOZ localization using machine learning (ensemble trees, CNNs) on raw (0–495 ms) post-stimulus waveform features and anatomical metadata achieves ROC-AUCs up to ~83%, with performance enhanced by ensemble aggregation and inclusion of anatomical context (Malone et al., 2022).

Unsupervised methods based on time-frequency chirp event outlier detection (LOF with spectro-temporal features) identify SOZ-associated electrodes with high index-based spatial concordance (precision ≈ 0.90 in seizure-free patients), showing utility as a complementary localization pipeline, especially in cases with clearly defined surgical outcomes (Bahador et al., 18 Aug 2025).

6. Limitations, Methodological Pitfalls, and Future Perspectives

SOZ localization remains sensitive to methodological choices. Label noise, sampling bias, and dependence on single-event or single-marker detection frameworks all risk either overfitting to artifacts or underdetecting true epileptogenic zones. Many AI or network-based approaches depend on reference annotations or indirect ground truth, propagating clinical subjectivity.

Current limitations include:

• Dependence of deep learning uncertainty estimation on pretraining (BUNDL)
• Lack of temporally correlated noise modeling in label-ambiguous frameworks
• Poor cross-validation of SOZ localization pipelines on non-TUH/non-adult datasets
• Incomplete handling of electrodecrements and suppression-type onsets in many conventional algorithms
• Modality-specific constraints (limited fMRI/fusion adoption, variable dMRI protocols)

Moving forward, recommended directions include:

• Hierarchical noise modeling and self-supervised pretraining to decouple SOZ detection from expert biases (Shama et al., 17 Oct 2024)
• Multi-rater consensus modeling for ambiguous or sparsely labeled data
• Integrating inter-modality priors via Bayesian frameworks or learned representations
• Large-scale, multi-center, prospective validation—including multimodal and pediatric populations

7. Summary Table: Representative Methodologies and Performance Metrics

Method/Modality	Quantitative SOZ Localization Results	Noted Limitations / Key Features
BUNDL (scalp EEG, DeepSOZ)	Seizure-level acc.↑3.2pp, patient-level acc.↑2.9pp	Label-uncertainty modeling, no extra params, SNR/label-noise robust
ICA+ECD (scalp EEG)	Mean error 10–47 mm (superficial→deep foci)	Poor deep SOZ detection; 40% seizure yield; requires expert IC pick
Dual GNN (SEEG)	Node-level acc. 94.7%, graph-level acc. 89.3%	Pediatric cohort, fixed windowing, Pearson-only edges
Seizure-NGCLNet (SEEG)	Concordance >85% with expert SOZ; acc. 95.9%	Node-graph contrast, attention-based localization
rs-fMRI + CNN + expert (DeepXSOZ)	F1=91.7%, accuracy=84.8% (hybrid)	White matter/ventricle features most critical; 6.7-fold workload↓
CCEP + ML ensemble	ROC-AUC ~83.2%, precision 68%, recall 76%	7-patient paper; raw timeseries inputs; metadata boosts accuracy
Chirp + LOF outlier (iEEG)	Index precision: 0.90 (seizure-free), 0.46 (failure)	Best in complete resection, small retrospective sample
sCMTF (EEG-fMRI fusion)	True SOZ covered in 12/12 pts (by any metric)	Advanced tensor/matrix factorization, HRF biomarkers

Methodological choices must be considered in the context of the clinical workflow, patient cohort, and desired trade-off between specificity and surgical sensitivity. Integration of robust statistical models, network science, deep learning, and expert-guided heuristics currently produces the most reliable outcomes.