Attention-Enhanced Dual-Branch ConvNeXt-BiLSTM Network for Subject-Independent EEG Seizure Detection

Maimuna Chowdhury; Sk. Imran Hossain

eess.SPAug 25, 2026

Attention-Enhanced Dual-Branch ConvNeXt-BiLSTM Network for Subject-Independent EEG Seizure Detection

Authors: Maimuna Chowdhury, Sk. Imran Hossain

Organizations: Dept. of Computer Science and Engineering Khulna University of Engineering & Technology Khulna, Bangladesh

Abstract

Automated seizure detection from scalp electroencephalography (EEG) is difficult because seizure morphology varies among patients and seizure samples are substantially outnumbered by non-seizure samples. This paper presents an attention-enhanced dual-branch network that jointly learns time--frequency and temporal representations from the same EEG segment. A continuous wavelet transform converts each segment into a scalogram processed by an ImageNet-pretrained ConvNeXt-Tiny backbone and squeeze-and-excitation attention. In parallel, a bidirectional long short-term memory network followed by multi-head self-attention models the raw signal. The two feature vectors are concatenated and classified by a weighted multilayer perceptron. Experiments use 14 subjects from the CHB-MIT scalp EEG database with subject-wise partitioning performed before overlapping segmentation. The model obtains $97.88\%$ accuracy and $97.51\%$ F1-score over ten across-subject splits, and $97.51\%$ accuracy, $96.59\%$ F1-score, and $98.02\%$ area under the ROC curve under 14-fold leave-one-subject-out validation. Removing temporal attention causes the largest ablation loss. The model requires 28.26 million parameters and 4.56 GFLOPs, with a measured network-only inference latency of 4.64 ms per segment.

Figures & tables

Fig. 1: Proposed pipeline. Each filtered EEG window is represented simultaneously as a raw sequence and a CWT scalogram. The two attention-refined branches are concatenated before binary classification.

Protocol	Acc.	Prec.	Sens.	Spec.	F1	AUC
Across-subject	$97.88\pm 0.52$	$97.32\pm 0.57$	$97.70\pm 0.43$	$98.12\pm 0.42$	$97.51\pm 0.49$	$98.52\pm 0.30$
LOSO	$97.51\pm 2.01$	$96.21\pm 2.11$	$96.97\pm 2.05$	$97.76\pm 1.93$	$96.59\pm 2.08$	$98.02\pm 1.84$

TABLE I: Subject-Independent Detection Performance (Mean

\pm

Standard Deviation, %)

Test subject	Accuracy	Precision	Sensitivity	Specificity	F1-score	AUC
CHB01	97.89	96.67	97.45	98.45	97.06	98.67
CHB03	99.23	98.12	98.78	99.56	98.45	99.67
CHB06	93.45	91.89	92.78	94.23	92.33	94.89
CHB08	99.45	98.34	99.01	99.78	98.67	99.89
CHB10	98.12	96.89	97.67	98.78	97.28	98.89
CHB11	95.67	94.12	95.01	96.34	94.56	96.78

TABLE II: Complete Subject-Wise LOSO Results (%)

Fig. 2: Training and validation accuracy (left) and loss (right) for a representative LOSO fold.

Model	Input	Accuracy (%)	F1 (%)
CNN	Scalogram	84.21	84.09
ResNet50	Scalogram	92.58	93.71
BiLSTM	Raw EEG	88.07	90.86
ConvNeXt-only	Scalogram	93.66	93.21
Proposed	Both	97.88	97.51

TABLE III: Baseline Models Under the Same Across-Subject Protocol

Variant	Accuracy (%)	F1 (%)
Full model (SE + MHSA)	97.88	97.51
MHSA only (without SE)	96.40	95.33
SE only (without MHSA)	94.99	94.76
Without SE and MHSA	94.08	93.97

TABLE IV: Attention Ablation Under the Across-Subject Protocol

Study	Technique	Protocol	Acc. (%)	F1 (%)
Shen et al. [ 14 ]	CWT + GoogLeNet	LOSO	93.20	–
Shen et al. [ 14 ]	STFT + GoogLeNet	LOSO	97.74	–
Song et al. [ 12 ]	STFT + ResCon-LSTM	Across	96.59	96.37
Abdulwahhab et al. [ 13 ]	STFT + CNN-LSTM	Hold-out	97.12	97.27
Zhao et al. [ 15 ]	S-trans. + ConvNeXt	8:1:1	98.83	–
Proposed	CWT + raw EEG	LOSO	97.51	96.59

TABLE V: Comparison with Representative CHB-MIT Studies

Explore similar work

DiffEEG: A Self-Supervised Denoising Diffusion Model for Learning EEG Generic Representations
Jul 13, 2026Abdulkader Helwan, Lina Abou-Abbas, Hussein El Amouri +2Electroencephalography EEG Seizure Detection
Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon · Institute of Applied Artificial Intelligence, T´ELUQ University, Montreal, Canada
Spatiotemporal Hyperedges for EEG Seizure Detection and Prediction
Sep 29, 2026Hyunju Kim, Sheo Yon Jhin, Noseong Park +1EEG Seizure Detection Graph Neural Networks
School of Computational Science and Engineering Georgia Institute of Technology Atlanta, GA, USA · School of Computing Korea Advanced Institute of Science and Technology Daejeon, Republic of Korea
ViTexSZ: Heterogeneous Vision-Text Knowledge Distillation for EEG Seizure Detection
Aug 19, 2026Chenxi Liu, Mingzhao Li, Yicong Liu +4EEG Seizure Detection Electroencephalography
Centre for Artificial Intelligence and Robotics, Hong Kong Institute of Science & Innovation, Chinese Academy of Sciences, Hong Kong SAR · The First Affiliated Hospital, Sun Yat-sen University, China · The Hong Kong Polytechnic University, Hong Kong SAR +2

Jul 13, 2026J/K move · Enter open · S save

DiffEEG: A Self-Supervised Denoising Diffusion Model for Learning EEG Generic Representations

Abdulkader Helwan, Lina Abou-Abbas, Hussein El Amouri, Belkacem Chikhaoui+1

Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon · Institute of Applied Artificial Intelligence, T´ELUQ University, Montreal, Canada

Electroencephalography EEG Seizure Detection Self-Supervised Time Series Representation Learning EEG Foundation Models Diffusion Models

Open paper AskPDF arXiv

Deep learning for EEG-based seizure detection faces critical challenges: severe annotation scarcity and extreme class imbalance, where ictal events comprise less than 10% of clinical recordings. We present DiffEEG, a 9.6M-parameter self-supervised foundation model that addresses both limitations through denoising diffusion pre-training and reinforcement learning (RL)-based fine-tuning. Pre-trained on 1.3M unlabeled segments from the Temple University Hospital Seizure Corpus (TUHSZ), DiffEEG learns generic neural representations via a 1D U-Net with multi-head self-attention. For downstream adaptation, a reinforced decision layer employs policy gradient optimization to directly maximize F1-score, prioritizing sensitivity to rare seizure events over overall accuracy. Under strict patient-wise evaluation (279 patients, Leave-One-Fold-Out), DiffEEG achieves 61% accuracy and 59% F1 for 4-class seizure subtyping, and 81% accuracy with 85% weighted F1 for binary detection, maintaining clinically viable seizure recall (59%) despite extreme imbalance (6.7% prevalence). Segment-level evaluation establishes an upper bound of 97.6% accuracy, confirming strong architectural capacity. DiffEEG demonstrates that diffusion-based pre-training combined with metric-aware reinforcement learning enables clinically deployable seizure monitoring with minimal labeled data requirements.

Abstract

Explore similar work

DiffEEG: A Self-Supervised Denoising Diffusion Model for Learning EEG Generic Representations

Spatiotemporal Hyperedges for EEG Seizure Detection and Prediction

ViTexSZ: Heterogeneous Vision-Text Knowledge Distillation for EEG Seizure Detection