Neonatal seizures are critical neurological emergencies requiring prompt diagnosis. Current deep learning methods for electroencephalogram (EEG) analysis often neglect the spatial topology between electrodes. This study aims to develop a novel, lightweight, multi-modal convolutional neural network to address this limitation and improve detection accuracy.

We propose a multimodal convolutional neural network integrating wavelet transform and attention mechanisms (MWAM-CNN). The model introduces three key innovations: (1) a two-dimensional spatial-topographic encoding (2D-STE) strategy to generate a five-dimensional tensor preserving spatial information for 3D convolutions;(2) a learnable wavelet transform convolution (WTConv) module for multi-scale feature extraction;(3) a four-path parallel architecture with a 3D Convolutional Block Attention Module (CBAM-3D) for adaptive feature fusion.

On the Helsinki dataset, balanced using the SMOTE method, the MWAM-CNN model achieved an accuracy of 97.2%, a sensitivity of 98.5%, a specificity of 95.4%, and an AUC of 99.7%. This high performance was achieved with a lightweight design of only 10,110 parameters, demonstrating its efficiency and effectiveness.

The proposed MWAM-CNN provides a robust and efficient solution for neonatal seizure detection. By effectively integrating spatial, temporal, and spectral features through its innovative architecture, the model significantly enhances detection performance while maintaining a lightweight structure suitable for clinical applications.