A lightweight deep convolutional neural network for detecting artifacts in continuous EEG signals

Develop and validate artifact-specific lightweight convolutional neural networks (CNNs) for automated detection of eye movement, muscle-related, and non-physiological artifacts in clinical EEG, and determine the optimal temporal window for each class. Three binary CNN detectors were trained on the Temple University Hospital EEG artifact corpus with patient-level 60/20/20 splits. Signals were standardized to 250 Hz and a 22-channel bipolar montage. Non-overlapping segments of 1–30 s were evaluated. Operating points were fixed by Youden’s J on validation and applied unchanged to the test set. Rule-based clinical comparators were implemented for each class. CNNs outperformed rule-based baselines. Optimal windows differed by artifact type: 20 s for eye movements (ROC AUC 0.975; F1 0.905), 5 s for muscle (accuracy 93.2%, specificity 96.0%, F1 0.855), and 1 s for non-physiological artifacts (F1 0.774; specificity 98.2%). Lightweight artifact-specific CNNs with class-tailored windows provide reliable EEG artifact detection and exceed rule-based performance at fixed operating points. The work offers practical guidance on per-class windowing (20 s eye, 5 s muscle, 1 s non-physiological) and transparent threshold selection for clinically oriented EEG quality control. • Deep learning outperforms rule-based artifact detection (F1 gain +11%–45%). • Optimal windows (seconds): 20 (eye), 5 (muscle), 1 (non-physiological). • Lightweight architecture achieves 96%–98% specificity for clinical use.