Lightweight deep learning model for nonconvulsive status epilepticus diagnosis using EEG time-frequency analysis

Nonconvulsive Status Epilepticus (NCSE) is a persistent epileptic seizure state whose detection primarily relies on visual EEG inspection. Automated algorithmic detection of NCSE can effectively increase the detection rate, enabling early intervention and improving prognoses. Previous studies are limited by issues such as not being specifically designed for NCSE, complex feature extraction approaches, suboptimal accuracy, and difficulty in deployment on edge devices. This study proposes an algorithm for the automated detection of NCSE, integrating time-frequency analysis, deep learning, and Exponential Moving Average (EMA). First, EEG data from NCSE patients were collected. Subsequently, features are extracted via Continuous Wavelet Transform (CWT), the multitaper method, and Hilbert-Huang Transform (HHT) to generate time-frequency maps. Finally, the feature maps are input into a lightweight MobileNetV3 network enhanced with a Coordinate Attention (CA) mechanism. Automatic identification of NCSE is then performed under the guidance of medical priors using EMA. Experimental results demonstrate that the proposed HHT + CA-MobileNetV3 + EMA pipeline achieves optimal performance with 97.54% accuracy. Furthermore, its lightweight architecture enables efficient deployment on low-cost edge computing devices. The findings demonstrate that our method provides an efficient and computationally economical solution for the automated diagnosis of NCSE.