Truace
Subscribe

ADMP-GCN: Interpretable Adaptive Message Passing Graph Learning for Patient-specific EEG Seizure Prediction with Human-in-the-Loop Validation

Automated seizure prediction based on multichannel Electroencephalography (EEG) using Artificial Intelligence (AI) is a challenging task that mandates accurate and interpretable results for improved clinical decisionmaking.However, brain connectivity modelling on the evolving patient-specific and seizure-dynamic EEG signals makes the problem highly complex.Recently, Graph Neural Network (GNN)-based approaches have shown promising results for improved seizure prediction by modelling brain functional connectivity…

By the Truvace desk1 min read1 sourcePeer-reviewedTRV-2026-1694· cite & verify
ADMP-GCN: Interpretable Adaptive Message Passing Graph Learning for Patient-specific EEG Seizure Prediction with Human-in-the-Loop Validation
Use case analysis: the ambulatory EEG in navy medicine for traumatic brain injuries by Burner, Tahnyah D.. Public domain

In brief

Researchers present ADMP-GCN, a graph convolutional network that models EEG brain connectivity with adaptive message propagation depth and reduced channels. After training, SHAP, Integrated Gradients and GNNExplainer are used to interpret which features, layers and graph structures drive each seizure prediction.

The work matters because clinical seizure prediction needs results that are both accurate and interpretable for decision-making. The evidence is a validation on one public dataset, so how well the approach works for patients in real clinical settings is not established by the text.

Main points

  1. Earlier graph neural network approaches used redundant channel representations and a fixed propagation mechanism, which the authors say limit practicality.
  2. ADMP-GCN uses topology-conditioned message passing to give heterogeneous receptive fields within EEG connectivity graphs.
  3. A human-aligned composite supervision balances prediction accuracy against explanation quality.
  4. Validation on CHB-MIT gave 95.10% accuracy and 96.33% sensitivity.

The gain

A topology-aware graph model with channel reduction and layered explanations reached 95.10% accuracy and 96.33% sensitivity on the CHB-MIT dataset, outperforming existing graph-based seizure predictors.

The rundown

The paper addresses automated seizure prediction from multichannel EEG, noting that patient-specific and seizure-dynamic signals make brain connectivity modelling highly complex. Graph neural networks had already shown promise by modelling functional connectivity and inter-channel interactions.

The proposed framework combines an explanation-constrained, topology-aware graph model with channel reduction aimed at efficiency and robustness. Explanations are produced at the feature, layer and graph levels after training.

Results on the CHB-MIT dataset showed higher accuracy and sensitivity than existing graph-based seizure prediction models, and the authors conclude the approach is viable and clinically relevant.

Sources

  1. Peer-reviewedInternational Journal of Intelligent Engineering and Systems2026-10-16

The debate