New imaging framework transforms epilepsy surgery planning
A unified computational framework developed at Carnegie Mellon University enables clinicians to analyse all major epileptic biomarkers within a single system, potentially transforming presurgical evaluation for drug-resistant epilepsy patients. The spatial-temporal-spectral imaging approach achieves localisation accuracy within millimetres of invasive monitoring whilst eliminating the need for separate analysis pipelines.
Novel approach addresses clinical imaging challenge
Researchers led by Professor Bin He at Carnegie Mellon University have developed a machine learning-based approach called spatial-temporal-spectral imaging (STSI) that represents the first technology capable of analysing every major type of epileptic brain signal within a unified computational framework. The work, published in the Proceedings of the National Academy of Sciences on 8 December 2025, addresses a fundamental challenge in presurgical planning for the estimated 30% of epilepsy patients whose seizures cannot be controlled with medication alone.
“For the first time, one algorithm can handle all epileptic biomarkers,” He emphasised. “This unified computational approach has never been done before.”
The current gold standard for presurgical evaluation relies on intracranial electroencephalography (iEEG), where electrodes are surgically placed over presumed epileptogenic regions. Whilst accurate, this invasive procedure requires lengthy hospital stays and carries inherent surgical risks. Noninvasive scalp EEG offers a safer alternative, but clinicians have lacked clarity on which biomarkers – interictal spikes, high-frequency oscillations (HFOs), or seizures – most reliably identify seizure-generating tissue.
Pathological HFOs emerge as optimal interictal biomarker
The research team conducted a multi-year study analysing 2,081 individual EEG events from 42 drug-resistant epilepsy patients in collaboration with Mayo Clinic clinicians. This represents the first rigorous quantitative comparison of all major epileptic biomarkers for source localisation. The authors noted in their paper: “The STSI enabled quantitative comparisons across key EEG epilepsy-related biomarkers, with averaged localisation errors of 6.67 mm for seizures, 8.73 mm for HFOs overlapping with spikes (pHFO), 10.28 mm for HFO-riding spikes (pSpike), 19.59 mm for general spikes (aSpike), and 36.53 mm for general HFOs (aHFO), respectively, for seizure-free patients.”
The findings revealed that pathological HFOs – which occur only when HFOs overlap with interictal spikes – are the most spatially accurate interictal biomarker for mapping the epileptogenic zone. These pathological HFOs localised the epileptogenic zone within approximately nine millimetres of invasive seizure mapping, approaching the seven millimetre accuracy achieved using actual seizures.
“You can record pathological HFOs in under an hour, instead of waiting days for a seizure,” He explained. “The accuracy is only two to three millimetres different.”
In contrast, general HFOs, once considered promising biomarkers, performed poorly. This finding helps clarify years of inconsistent results across clinical studies and highlights the importance of distinguishing pathological from physiological high-frequency oscillations.
Technical innovation enables unified biomarker analysis
The STSI framework operates by jointly analysing where, when, and at what frequencies brain activity occurs. The system decomposes complex, high-dimensional EEG data into low-dimensional representations, delineates spectral characteristics of different biomarkers, and maps the location, dynamics, and extent of sources generating these biomarkers through data-driven optimisation.
The framework relies on tensor decomposition analysis, which enables low-dimensional representation of complex neurophysiological data. As the authors describe in their paper: “STSI decomposes complex, high-dimensional EEG data into low-dimensional representations, delineates the spectral characteristics of different biomarkers in EEG measurements, and then maps the location, dynamics, and extent of sources generating these biomarkers with data-driven L1 norm-based optimisation.”
When benchmarked against established methods including standardised low-resolution brain electromagnetic tomography (sLORETA) and linearly constrained minimum variance (LCMV) beamforming, STSI demonstrated superior accuracy in estimating both source location and extent across all frequency ranges. The system showed significantly lower localisation error and spatial dispersion, with higher precision and specificity compared to conventional approaches.
Performance stratified by surgical outcome
The research revealed distinct performance patterns when stratifying patients by surgical outcome. Among the 42 patients studied, 30 achieved seizure freedom (ILAE I-II) whilst 12 remained non-seizure-free (ILAE III-V). Seizure source localisation showed consistent performance across both groups, whilst pathological HFOs and pathological spikes exhibited statistical differences in localisation error and spatial dispersion between groups.
For seizure-free patients, pathological spikes showed significantly higher spatial dispersion compared with seizures and pathological HFOs. The authors note: “These findings indicate that HFOs overlapping with spikes is the most spatially accurate interictal biomarker for mapping the EZ.”
When comparing results to iEEG-defined seizure onset zones in 20 patients with available invasive monitoring data, all three biomarkers (seizures, pathological HFOs, and pathological spikes) exhibited similar performance, with the seizure-free group showing significantly lower localisation error values across all biomarkers.
Broader applications beyond epilepsy
The implications of STSI extend beyond epilepsy surgery planning. The framework’s capability to analyse any EEG or magnetoencephalography signal, whether transient or oscillatory, opens applications across multiple domains of neuroscience and clinical medicine.
As the authors state in their paper: “STSI holds significant promise for cognitive neuroscience and clinical research. It can be extended to map event-related potentials (ERPs) like the P300 and N400, which are central to cognitive functions such as attention, memory, and language processing, or evoked potentials such as visual, auditory, and somatosensory evoked potentials, that are widely used in brain research and clinical neurophysiology testing.”
The framework could enable imaging sources of neural oscillations and dynamic brain-state modelling, potentially advancing understanding of network dynamics in conditions including Parkinson’s disease, Alzheimer’s disease, and psychiatric disorders.
Future development and clinical translation
He aims to secure funding to validate the technique in larger patient cohorts and prepare it for clinical adoption. The team has made the computer code publicly available via GitHub < https://github.com/bfinl/STSI > to facilitate broader research application and development.
“The whole point is to help others,” He said. “If we can provide a noninvasive, precise alternative that spares patients from days of invasive monitoring, that would have a major impact. We’re committed to improving the patient experience through our expertise.”
The research represents a collaboration between Carnegie Mellon University’s Department of Biomedical Engineering and the Mayo Clinic’s Department of Neurology, with data collection and clinical validation conducted at Mayo Clinic, Rochester, Minnesota.
Reference
Jiang, X., Cai, Z., Gonsisko, C., Worrell, G. A., & He, B. (2025). Mapping epileptogenic brain using a unified spatial-temporal-spectral source imaging framework. Proceedings of the National Academy of Sciences, 122(50), e2510015122. https://doi.org/10.1073/pnas.2510015122
