Abstract:To address high clinical false positive rates in wearable continuous long-term electroencephalogram (EEG) monitoring caused by insufficient broadband weak signal parsing and blurred classification boundaries, a lightweight multi-scale contrastive EEGNet (MC-EEGNet) parsing and prediction model is constructed. Firstly, a multi-scale temporal convolutional block attention module (MS-CBAM) is designed to decouple complex EEG signals in the frequency domain via parallel multi-scale branches, utilizing a dual attention mechanism to adaptively recalibrate focal channels and key temporal features; secondly, a margin-supervised contrastive learning loss (Margin-SupCon) is introduced to establish a rigid inter-class isolation band within the high-dimensional feature manifold to sharpen classification decision boundaries; lastly, a two-stage data augmentation strategy cascading offline preprocessing and online dynamic training is proposed to compensate for signal interferences under non-ideal physical measurement constraints such as transient signal interruption in EEG acquisition links and poor sensor contact via wavelet frequency mixing, spatiotemporal masking, and a physics-prior-based impedance noise simulation module. The MC-EEGNet is evaluated using leave-one-out cross-validation on the public CHB-MIT scalp EEG dataset. Results demonstrate that under a 30-minute prediction window, the model achieves an average sensitivity of 95.58%, the hourly false positive rate (FPR/h) is reduced to 0.013 (a 46% reduction compared to the baseline model), and the average prediction time is extended to 22.46 minutes. The model effectively overcomes the perceptual limitations of traditional single-scale convolutions and achieves an optimal balance between sensitivity and false positive suppression under low computational overhead, providing robust algorithmic support for the edge-embedded deployment of portable smart EEG monitoring instruments.