🧠 neurology

Localising the epileptogenic zone from single-pulse electrical stimulation responses using cross-trial attention

This study demonstrates that an interleaved Hierarchical Attention Transformer (HAT) model, which explicitly captures cross-trial dependencies in single-pulse electrical stimulation responses without averaging, achieves superior concordance with the clinical seizure onset zone compared to traditional averaging methods while remaining robust to limited trial availability, though its association with postsurgical seizure freedom remains inconclusive.

Original authors: Norris, J., van Blooijs, D., Chari, A., Cooray, G., Tisdall, M., Friston, K., Smith, S. D. W., Rosch, R.

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Norris, J., van Blooijs, D., Chari, A., Cooray, G., Tisdall, M., Friston, K., Smith, S. D. W., Rosch, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain is a bustling, chaotic city. Usually, traffic flows smoothly, but in people with epilepsy, certain neighborhoods get stuck in gridlock, causing sudden, overwhelming traffic jams we call seizures. For many, medicine acts like a traffic controller, keeping things moving. But for about one-third of patients, the lights don't work, and the jams keep happening. To fix this, doctors sometimes have to perform brain surgery, aiming to remove the specific "neighborhood" where the jams start. The tricky part is finding that exact spot. It's like trying to locate a single faulty wire in a massive, tangled ball of yarn without cutting the whole thing apart.

To find this "faulty wire," doctors use a tool called Single-Pulse Electrical Stimulation (SPES). Think of it as tapping a specific spot on the brain with a tiny, gentle electric finger and listening to how the rest of the city reacts. If you tap a normal spot, the reaction is predictable. But if you tap the trouble spot, the brain might react strangely, sending out weird signals or reacting differently every single time you tap it. Traditionally, doctors have tapped the same spot many times and then averaged all those reactions together to get a clear picture, much like taking a long-exposure photo to blur out the noise. But what if the most important clues are hidden in the tiny differences between each tap? That's the question this new research asks.

The authors of this study, led by Jamie Norris, decided to stop averaging the taps and instead looked at every single reaction individually using a fancy new type of computer brain called a "Hierarchical Attention Transformer" (or HAT for short). Imagine you are trying to identify a suspect in a crowd. The old way was to take a blurry photo of the whole group and hope the suspect stands out. The new HAT method is like having a super-smart detective who watches every single person in the crowd, notices how they react to different people walking by, and remembers how they acted in the past. It doesn't just look at the average; it pays attention to the unique, quirky details of every single moment.

The researchers tested this detective on data from 35 patients. They compared their new HAT method against the old "averaging" method and a middle-ground method. The results were promising: the new HAT detective was better at spotting the trouble spots than the old averaging method. Specifically, it got a score of 0.762 out of 1.0, while the old method only managed 0.721. It's a small but statistically significant improvement, suggesting that paying attention to the tiny, trial-by-trial differences really does help find the epileptic zone.

Here is the really cool part: the new method was so good that it didn't even need to see all the taps to do its job. Even when the researchers only gave it the very first tap (just one trial) to make a decision, the HAT detective performed almost as well as when it saw all the data. This suggests that in the future, doctors might not need to tap the brain dozens of times to get a good reading; maybe just a few taps would be enough.

However, the story doesn't end with a perfect victory. The researchers also asked a crucial question: "If we remove the spots the new detective says are bad, do the patients stop having seizures?" They looked at the surgery results for these patients and found... well, they couldn't say for sure. The data didn't show a strong enough link to prove that removing the spots the model identified would guarantee a seizure-free life. The authors are careful to point out that this might be because the model was trained to find the "clinical" trouble spots, which aren't always the exact same thing as the "real" trouble spots that cause seizures.

So, what's the takeaway? This study suggests that we can find the brain's trouble spots more accurately by using a smart computer model that listens to every single reaction instead of just averaging them out. It's a step forward, a proof that looking at the details matters. But it's not a magic wand yet. The model is a helpful tool that might one day guide surgeons, but it hasn't been proven to cure epilepsy on its own just yet. The authors suggest that to make it truly useful, we need to teach it with even better data and test it on more patients in different hospitals. For now, it's a very clever, very promising new way of listening to the brain's whispers.

Technical Summary: Localising the Epileptogenic Zone from Single-Pulse Electrical Stimulation Responses Using Cross-Trial Attention

Problem Statement
The delineation of the epileptogenic zone (EZ) is critical for surgical success in drug-resistant epilepsy, yet current reliance on the clinically defined seizure onset zone (SOZ) yields suboptimal outcomes, with only about half of patients achieving seizure freedom five years post-surgery. A significant limitation in current methodologies is the analysis of Single-Pulse Electrical Stimulation (SPES) responses. Standard clinical practice involves averaging repeated stimulation trials to improve signal-to-noise ratio. However, this averaging process may obscure clinically relevant trial-to-trial variations, such as delayed responses or variable latencies, which have been linked to epileptogenicity. Furthermore, existing machine learning approaches often treat individual trials independently or rely on trial-averaged inputs, potentially discarding the rich cross-trial dependencies that characterize epileptic tissue.

Methodology
The authors propose a deep learning framework designed to model cross-trial and cross-channel dependencies in SPES data without prior averaging. The study utilizes a retrospective cohort of 35 patients from a single center (UMC Utrecht) who underwent intracranial EEG monitoring with SPES.

Data Processing: The data follows a "convergent paradigm," where responses at a specific recording site are grouped based on stimulation from various other locations. This captures the inherent excitability of the recording site. Preprocessing included band-pass filtering (1–150 Hz), artifact removal, and exclusion of the first 9 ms post-stimulation to mitigate stimulation artifacts.
Model Architecture: The core contribution is an interleaved Hierarchical Attention Transformer (HAT).
- Input: Unlike baselines that average trials, the HAT processes individual trial responses. Each trial is embedded via a Multilayer Perceptron (MLP) and concatenated with the Euclidean distance between the stimulation and recording sites.
- Architecture: The model employs an encoder-only Transformer with interleaved layers. It alternates between Channel-wise Encoders (CWE), which apply cross-trial attention to capture dependencies between repeated responses from the same stimulation-recording pair, and Cross-Channel Encoders, which model relationships between different recording sites.
- Baselines: The HAT was compared against two baselines:
  1. Baseline: Trial-averaged SPES responses embedded and fed into a cross-channel Transformer.
  2. Mean Trial Embedding (MTE): Individual trials embedded, then mean-pooled across trials before entering the cross-channel Transformer.
Evaluation Strategy: The models were evaluated using patient-held-out, repeated five-fold cross-validation (25 total runs). Performance was measured primarily by AUROC concordance with the clinical SOZ. Robustness was tested by restricting inference to 1 or 5 trials. An exploratory analysis assessed whether the proportion of model-positive channels resected correlated with postsurgical seizure freedom (12 seizure-free vs. 23 non-seizure-free patients).

Key Contributions

Novel Architecture: Introduction of the HAT, the first architecture to explicitly model cross-trial dependencies in SPES responses using interleaved attention mechanisms, avoiding the information loss associated with trial averaging.
Performance Improvement: Demonstration that modeling cross-trial dependencies yields statistically significant improvements in SOZ localization compared to trial-averaged approaches.
Robustness Analysis: Evidence that the model's performance remains stable even when inference is restricted to a single trial, suggesting potential for reduced acquisition time, though this was not prospectively tested.
Outcome Analysis: A rigorous, albeit negative, exploration of the link between model predictions and surgical outcomes, highlighting the limitations of using clinical SOZ labels as a proxy for the true epileptogenic zone.

Results

SOZ Concordance: The HAT achieved a mean AUROC of 0.762, significantly outperforming the trial-averaged Baseline (AUROC 0.721; mean difference 0.041, Holm-adjusted p = 0.0197) and the MTE model (AUROC 0.734). The HAT also showed higher Youden's index (0.364) compared to the Baseline (0.309).
Trial Subset Analysis: When inference was restricted to the first trial ( $n=1$ ), the HAT and MTE models showed minimal degradation in performance. In contrast, the Baseline model suffered a notable drop in Youden's index (from 0.309 to 0.261) due to reduced sensitivity.
Surgical Outcome Association: The exploratory analysis found no statistical evidence that a higher proportion of resected model-positive channels (HATEZ) was associated with seizure freedom (AUROC 0.634, p = 0.102). Similarly, intersections or unions of HATEZ and clinical SOZ did not yield significant associations. Among seizure-free patients, the mean proportion of HATEZ channels resected was 41.0%, indicating that seizure freedom occurred even when most model-predicted epileptogenic channels were preserved.

Significance and Claims
The paper claims that explicitly modeling cross-trial dependencies in SPES responses improves the concordance of machine learning models with the clinical SOZ compared to standard trial-averaged methods. The authors emphasize that this improvement is achieved while maintaining robustness to reduced trial availability.

However, the authors are modest regarding clinical utility. They explicitly state that the study does not establish a clinical role for the model. The lack of association with surgical outcomes is attributed to the training labels (clinical SOZ) being an imperfect proxy for the true epileptogenic zone. Consequently, the authors conclude that the model should not be used to direct resection strategies in its current form. They suggest that future work must focus on multicentre validation, outcome-aligned supervision (training on resection outcomes rather than SOZ labels), and the integration of multimodal data to bridge the gap between improved SOZ localization and actual surgical success.

Technical Summary: Localising the Epileptogenic Zone from Single-Pulse Electrical Stimulation Responses Using Cross-Trial Attention

More like this