Research map / Label-efficient and robust learning / OOD and anomaly detection

DMNet: Self-comparison Driven Model for Subject-independent Seizure Detection

NeurIPS 2024 · Label-efficient and robust learning › OOD and anomaly detection · approach: Deep one-class and tabular anomaly detection · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DMNet, a subject-independent iEEG seizure detector that compares each target segment with contextual and channel-level references via a fully differenced difference matrix encoded by a small CNN.

Problem

Subject-independent automated seizure detection from intracranial EEG is hampered by large domain shifts across subjects, time intervals, and channels, so models trained on some subjects generalize poorly to unseen ones.

Approach

Self-comparison: each target segment's DFT representation is differenced pairwise against adjacent normal segments (contextual reference) and K-Means cluster centroids of the channel (channel-level reference), producing a min-max normalized difference matrix that a CNN classifies. This exploits the universal pattern that seizures have higher amplitude than adjacent normal activity.

Key results

• On the clinical dataset DMNet reaches 55.93 F2 versus 51.12 for the next best (PPi), averaging about 9.41% F2 gain over SOTA.
• On MAYO DMNet attains 81.54 F2 vs 71.02 for PPi; on FNUSA 75.15 F2 vs 71.59 for PPi.
• DMNet has 0.316MB parameters and 45.5 s average inference per subject, the fastest among compared models.

Measured effects reported in the paper

Surprising findings

• Standard domain generalization baselines (e.g., SelfReg, GroupDRO, CORAL) underperform EEG-based and iEEG-specific methods on this task.
• The paper reports that the iEEG domain-adversarial methods (SEEG-Net, SICR) can suffer negative transfer.

Limitations

• Clinical dataset is private, so the main results cannot be reproduced independently.
• Lacks a theoretical foundation, as acknowledged by the authors.
• Channel-level reference is removed for the public datasets because they lack full channel data.

Experimental setup

Datasets: clinical iEEG dataset (8 subjects) MAYO FNUSA TUSZ

Benchmarks: clinical iEEG dataset MAYO FNUSA TUSZ

Metrics: precision recall F1-score F2-score

Baselines: SICR SEEG-Net PPi Abou-Abbas et al. Zhao et al. Dissanayake et al. SelfReg GroupDRO CORAL CDANN

Builds on: domain generalization K-Means discrete Fourier transform

Closest papers by meaning

• REST: Efficient and Accelerated EEG Seizure Analysis through Residual State Updates (ICML 2024 · similarity 0.87)
• Long-Term EEG Partitioning for Seizure Onset Detection (AAAI 2025 · similarity 0.87)
• EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks (NeurIPS 2025 · similarity 0.86)
• Seizure-Semiology-Suite($S^3$): A Clinically Multimodal Dataset, Benchmark, and Models for Seizure Semiology Understanding (ICML 2026 · similarity 0.85)
• NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.85)
• FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction (NeurIPS 2025 · similarity 0.84)