ConvMambaNet: A Hybrid CNN-Mamba State Space Architecture for Accurate and Real-Time EEG Seizure Detection

Abstract: Epilepsy is a chronic neurological disorder marked by recurrent seizures that can severely impact quality of life. Electroencephalography (EEG) remains the primary tool for monitoring neural activity and detecting seizures, yet automated analysis remains challenging due to the temporal complexity of EEG signals. This study introduces ConvMambaNet, a hybrid deep learning model that integrates Convolutional Neural Networks (CNNs) with the Mamba Structured State Space Model (SSM) to enhance temporal feature extraction. By embedding the Mamba-SSM block within a CNN framework, the model effectively captures both spatial and long-range temporal dynamics. Evaluated on the CHB-MIT Scalp EEG dataset, ConvMambaNet achieved a 99% accuracy and demonstrated robust performance under severe class imbalance. These results underscore the model's potential for precise and efficient seizure detection, offering a viable path toward real-time, automated epilepsy monitoring in clinical environments.

• The paper presents ConvMambaNet, a hybrid CNN-Mamba SSM architecture that achieves 99% accuracy in real-time EEG seizure detection using efficient spatial and temporal feature extraction.
• It leverages convolutional encoding for spatial analysis and Mamba-based state-space modules with multi-head attention to capture long-range temporal dependencies.
• The model reduces epoch training time significantly and outperforms state-of-the-art benchmarks on clinical EEG data, supporting its potential for real-time bedside applications.

ConvMambaNet: Integrating CNN and Mamba SSM for Efficient Real-Time EEG Seizure Detection

Problem Context and Motivation

Automated EEG seizure detection faces significant challenges due to the temporal complexity and inter-individual variability of EEG signals. Previous paradigms, including classic ML with handcrafted features and end-to-end deep neural architectures like CNNs, RNNs, and Transformers, have demonstrated limited scalability, computational efficiency, and robustness for real-time clinical applications. Recent advances in hardware-efficient sequence modeling have invigorated research into SSMs, with Mamba SSM delivering reduced latency for long-range temporal dependencies. The paper introduces ConvMambaNet, a hybrid framework that synergistically integrates CNN-based spatial feature extraction with Mamba-SSM modules for temporal modeling. The architecture directly addresses the dual requirements of accurate and low-latency EEG-based seizure detection, crucial for clinical deployment.

Model Architecture

ConvMambaNet is architected as a multi-stage pipeline, beginning with convolutional encoding layers that abstract multichannel EEG signals into high-level spatial representations. These CNN-extracted features are then processed by integrated Mamba-SSM blocks, which employ selective state-space transformations with gating to efficiently capture long-range temporal dependencies in EEG time series. The use of multi-head attention mechanisms within Mamba-SSM enhances the sensitivity of temporal pattern recognition, and layer-appropriate initialization protocols (He/Kaiming, Glorot, stability-oriented parameterization for SSM matrices) mitigate exploding/vanishing gradient issues. The overall architecture is end-to-end trainable and highly hardware efficient.

Figure 1: Schematic of the ConvMambaNet pipeline, combining CNN spatial analysis with Mamba-SSM-based temporal modeling for robust EEG classification.

EEG Data Handling and Preprocessing

The model is validated on the CHB-MIT Scalp EEG dataset, comprising recordings from 22 pediatric epilepsy patients. Preprocessing encompasses patient-wise stratified splits, robust channel selection focusing on clinically-relevant bipolar derivations, and balanced sampling to mitigate severe class imbalance (non-seizure events vastly outnumber seizures). Overlapping sliding windows segment the signals, and binary seizure labels are precisely aligned using expert annotations and the WFDB toolkit. This generates a final dataset of 9,505 signals, ensuring no cross-contamination between training and testing splits, supporting strong external validity.

Figure 2: Multichannel EEG excerpts illustrating the signal heterogeneity present across the CHB-MIT dataset.

Training Protocol and Numerical Performance

Model training utilizes the Adam optimizer, batch normalization, and dropout regularization. Loss convergence is fast and stable without evidence of overfitting, supported by simultaneous evaluation of precision, recall, accuracy, F1-score, and AUC-ROC. On the held-out test set, ConvMambaNet achieves 99% accuracy, with precision, recall, and F1-score of 0.99, and an AUC of 0.97. The model demonstrates extremely high sensitivity and specificity for both seizure and non-seizure classes, with near-perfect detection in the negative class—a key clinical requirement to suppress false alarms.

Figure 3: Training and validation loss curves for ConvMambaNet, showing stable convergence and absence of overfitting.

Comparative Performance Analysis

Comprehensive benchmarking is conducted against baseline and state-of-the-art models, including pure CNNs, RNNs (LSTM/GRU), Transformers, and recent SSM-based and spiking models (SGSTAN, EEGMamba, SlimSeizNet, Single-Channel DL, Spiking Conformer, and SOUL Online). ConvMambaNet outperforms on all metrics, including precision, recall, F1, and AUC, achieving superior results even relative to SOTA SSM and graph-based architectures. Select results for ConvMambaNet versus strong baselines include:

• CNN: 96% accuracy, F1-score 0.94
• Transformer: 94% accuracy, F1-score 0.97
• SGSTAN: 98% accuracy, F1-score 0.97
• EEGMamba: 97% accuracy, F1-score 0.79
• ConvMambaNet: 99% accuracy, F1-score 0.99

This demonstrates not only technical superiority in terms of predictive performance but also the robustness of the architecture for broad clinical translation.

Figure 4: (a) Confusion matrix highlighting balanced classification; (b) ROC curve visualizing model discrimination ($\mathrm{AUC} = 0.97$).

Computational Efficiency

ConvMambaNet's integration of Mamba-SSM yields significant computational benefits: training time is reduced to 48 seconds per epoch (on an RTX 2070 GPU), which is substantially less than that of CNNs (over two minutes), RNNs (1:43 min), and Transformers (four minutes per epoch). The architecture's linear scaling with respect to sequence length, combined with hardware friendliness, makes it suitable for low-latency, high-throughput settings, a critical requirement for real-time bedside seizure detection and ICU monitoring.

Practical and Theoretical Implications

Pragmatically, ConvMambaNet's strong performance under class imbalance and across subjects supports its direct applicability to clinical EEG monitoring, with the potential for integration into wearable or bedside diagnostic workflows. The architecture's real-time suitability and low false-positive rate address barriers that have hampered prior deep models. Theoretically, the paper validates the hypothesis that hybrid convolutional/state-space models systematically outperform both pure convolutional and pure sequence-based models for biomedical time series involving complex spatiotemporal dependencies. The approach generalizes to other domains where spatial and long-range sequential dependencies co-occur. The selective gating in Mamba-SSM opens new directions for interpretable sequence modeling and efficient deployment in biomedical signal processing.

Future Directions

The success of ConvMambaNet suggests several avenues for future research:

• Extension of the hybrid SSM-CNN framework to multi-task EEG problems (prediction vs. detection, localization)
• Personalized seizure detection with transfer and meta-learning strategies
• Channel pruning and sample-efficient learning for ultra-low-power deployment
• Model interpretability via attention/memory visualizations for supporting clinical decision-making
• Exploration of the hybrid paradigm in other dense physiological time series modalities (ECG, MEG, fNIRS)

Conclusion

ConvMambaNet demonstrates the efficacy of fusing CNN-based spatial encoders with Mamba-based SSM temporal modules for high-performance, low-latency seizure detection on clinical EEG data. The model achieves strong numerical results on the CHB-MIT dataset, returning 99% accuracy and 0.99 F1-scores with pronounced training efficiency. The proposed paradigm not only advances the technical frontier in biomedical sequence modeling but also directly facilitates the deployment of real-time, scalable, and robust neurodiagnostic solutions for epilepsy and similar neurological disorders.