Title: CLSP-REQA: A Real-Time Quality-Aware Closed-Loop Seizure Prediction Framework with Mamba-BiLSTM and Confidence-Gated Intervention

Abstract:

While reliable seizure prediction is essential for closed-loop neurostimulation therapies, current methodologies often fail to address the fluctuating quality of EEG signals in practical applications. Furthermore, most existing approaches rely on lenient evaluation standards that artificially inflate estimates of generalization capability. To address these limitations, we introduce CLSP-REQA (Closed-Loop Seizure Prediction with Real-time EEG Quality Assessment), a cohesive framework that integrates a lightweight signal quality estimator directly into the prediction workflow. This system features a Real-time EEG Quality Assessment (REQA) module operating in parallel with a Mamba-BiLSTM backbone. The REQA module generates a scalar quality score, $q$, ranging from 0 to 1, which adjusts the model’s output confidence via a tiered non-linear fusion function known as ECLO.

In rigorous cross-patient evaluations using the CHB-MIT Scalp EEG Database (comprising 23 subjects and 198 seizures), CLSP-REQA attained an AUC-ROC of 0.7426 ± 0.0199. This performance surpasses the unadapted cross-patient baseline of 0.69 established by Jemal et al., despite utilizing only 16 EEG channels compared to the 23 employed in previous studies, and without the need for target-patient data or domain adaptation. Additionally, on the SIENA Scalp EEG Database (14 subjects, 47 seizures), the framework achieved an AUC of 0.7012 ± 0.0249, significantly outperforming the top domain-adapted cross-patient result of 0.61 recorded on this dataset, thereby highlighting its robust generalization across different datasets. The framework delivers a structured four-tuple $(p, q, c, \Phi_{SHAP})$ that is immediately compatible with closed-loop neurostimulator interfaces.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC