Home / Research
Research

The Facial Expression Recognition and Affective Computing Research Group at the School of Psychology ProposesEpileptiform Discharge Network Subtyping and Subtype-Enhanced Prediction

2026-11-18

At the School of Psychology, Shanghai Jiao Tong University, the Facial Expression Recognition and Affective Computing Research Group’s Co-PI Dr. Feng Liu collaborated with a multidisciplinary team to develop an MEG-based method for subtyping epileptiform discharge networks. The team identified four network subtypes and used subtype information to improve patient-level prediction, offering a new approach to medical diagnosis using noninvasive brain–computer interface technologies such as OPM-MEG. The findings, entitled “Unsupervised Subtyping of Focal Epilepsy IEDs via Graph-Based MEG Networks,” were first published online on September 22, 2026 (Early Access) in the well-regarded biomedical engineering journal IEEE Transactions on Biomedical Engineering (TBME).

Research at a Glance

The study found that multiple epileptiform discharge network states can coexist within the same patient; incorporating their subtype structure and each patient’s subtype composition into a prediction model can improve patient-level classification of the spatial distribution of discharges.

Research Overview

Interictal epileptiform discharges (IEDs) provide an important window into pathological brain networks in epilepsy. Previous classification approaches have focused mainly on waveform morphology, while the brain regions involved in these discharges and their connectivity patterns remain to be characterized in greater detail. Noninvasive magnetoencephalography (MEG) offers a way to investigate these network states.

Research Motivation

The team asked two questions: Can epileptiform discharges be subtyped according to brain network organization? Can subtype-specific connectivity patterns and within-patient subtype composition improve individual-level prediction? To address these questions, the study proceeded from event-level network construction to unsupervised subtyping and patient-level validation.

Research Contributions

1. Graph Construction and Unsupervised Subtyping of Epileptiform Discharges

The study included 55 patients with focal epilepsy. Each underwent one hour of resting-state MEG, with structural MRI used for source reconstruction. A total of 8,509 IEDs were manually identified and analyzed. Phase-locking values (PLVs) between 148 brain regions were calculated to construct 148 × 148 functional connectivity matrices, one per event. After comparing various graph embedding and clustering methods, the team selected multi-view graph representation learning (MVGRL) and K-means clustering, identified four IED network subtypes, and extracted subtype-specific nodes and connectivity subgraphs.




Figure 1.Workflow for IED network construction and unsupervised subtyping based on MEG source-space functional connectivity.

2.Subtype-Specific Network Information and Within-Patient Differences

The four subtypes differed in network efficiency and regional connectivity patterns, with global efficiency ranked as Subtype 4 > Subtype 1 > Subtype 2 > Subtype 3. Representative nodes in Subtype 1 involved superior and inferior temporal regions; those in Subtype 2 involved occipitotemporal and parahippocampal regions; and those in Subtype 3 involved regions including the temporal pole and parieto-occipital areas. Subtype 4 prominently involved the bilateral subcallosal areas. All four subtypes occurred in 41 of the 55 patients, showing that multiple epileptiform discharge network states can coexist within the same individual. The spatial features of the subtypes and differences in their composition provide a more detailed characterization of within-patient network heterogeneity.

Figure 2 Mean connectivity matrices and network metrics for the four IED network subtypes. Subtype 4 had the highest global efficiency and shortest characteristic path length, while Subtype 3 showed the opposite pattern.

3.Subtype-Guided Patient-Level Prediction Network

The team treated each patient’s IED networks as a bag of instances and used attention-based multiple-instance learning to aggregate a patient representation. The model incorporates four learnable subtype prototypes and uses soft routing to weight and combine four expert branches according to the patient’s subtype composition. It also combines clustering alignment and contrastive learning to preserve the subtype structure discovered through unsupervised learning during downstream training. The model classifies the spatial distribution of MEG discharges at the patient level as temporal or other.

Figure 3 Patient-level classification model guided by subtype soft routing. Attention pooling aggregates IED representations, subtype prototypes determine the weights for combining expert branches, and auxiliary losses preserve subtype structure.

The study used validation at both the matrix and patient levels. A support vector machine using graph-theoretic features identified the four subtypes with an accuracy of 93.9%. Across 20 repetitions of leave-one-patient-out cross-validation, the full model achieved a patient-level balanced accuracy of 58.8%, exceeding a random forest using graph-theoretic features (50.3%) by 8.5 percentage points and a deep multiple-instance model without subtype routing (50.7%) by 8.1 percentage points.

Research Innovations

This study extends the focus of analysis from discharge waveforms to source-space brain networks and interprets subtype differences through global topology, regional nodes, and within-patient subtype composition. Subtype-specific subgraphs link data-driven clustering results to specific brain regions and their connections, providing a network-level basis for understanding different epileptiform discharge states.

For predictive modeling, the study uses subtypes discovered through unsupervised learning as learnable structural priors, allowing each patient’s composition of discharge states to inform classification decisions. The framework connects network-state discovery with patient-level prediction, providing a methodological foundation for decoding pathological brain states with network-level interpretability and for research on personalized noninvasive brain–computer interfaces.

Limitations and Future Directions

This study used a retrospective cohort and relied on manual IED identification. The patient-level classification labels reflect the distribution of interictal activity described in clinical MEG reports. The clinical value of this classification for localization still requires validation using independent seizure onset zone, epileptogenic zone, and outcome data. Prospective studies and online brain-state decoding are directions for future work.

Research Team

Fudan University Ph.D. student Qijian Zheng is the paper’s first author and is jointly supervised by Professor Yang Chen of Fudan University and Dr. Feng Liu of our school. Dr. Yafei Liu, Dr. Chunyan Cao, and Dr. Feng Liu of our school are co-corresponding authors, with Dr. Feng Liu serving as the lead corresponding author. The study was conducted jointly by researchers from the School of Psychology at Shanghai Jiao Tong University, Fudan University, Army Medical University, China Medical University, Shanghai University, and Ruijin Hospital affiliated with Shanghai Jiao Tong University School of Medicine. The study was supported by the 2025 Shanghai Magnolia Talent Plan Pujiang Project and the Startup Program for Newly Recruited Faculty at Shanghai Jiao Tong University.

About the Journal

IEEE Transactions on Biomedical Engineering (TBME) is the flagship journal of the IEEE Engineering in Medicine and Biology Society (EMBS). It publishes basic and applied research in biomedical engineering, covering engineering methods and technologies for biomedical applications, as well as experimental and clinical research with engineering contributions.

Publication Details

Title: Unsupervised Subtyping of Focal Epilepsy IEDs via Graph-Based MEG Networks

Online publication date: September 22, 2026 (Early Access)

Authors: Qijian Zheng¹, Yafei Liu²˒³*, Zhongyi Xie⁴, Ke Dong⁵, Zhi Liu², Yujie Chen², Jian Sun³, Chunyan Cao⁶*, Feng Liu⁷*

Affiliations: ¹College of Computer Science and Artificial Intelligence, Fudan University; ²Department of Neurosurgery, Southwest Hospital, Army Medical University; ³School of Biomedical Engineering, Army Medical University; ⁴School of Intelligent Medicine, China Medical University; ⁵School of Microelectronics, Shanghai University; ⁶Department of Neurosurgery, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine; ⁷School of Psychology, Shanghai Jiao Tong University.

* Corresponding authors  Contact email: liu.feng@sjtu.edu.cn

DOI: https://doi.org/10.1109/TBME.2026.3736741

IEEE article link: https://ieeexplore.ieee.org/document/11703921

Code: https://github.com/Cross-Innovation-Lab/MEGnet4FocalEpilepsy

Citation:

Q. Zheng, Y. Liu, Z. Xie, K. Dong, Z. Liu, Y. Chen, J. Sun, C. Cao, and F. Liu, “Unsupervised Subtyping of Focal Epilepsy IEDs via Graph-Based MEG Networks,” IEEE Transactions on Biomedical Engineering, early access, Sep. 22, 2026, doi: 10.1109/TBME.2026.3736741.