Detection-Guided Artifact Removal for Clinical EEG: A Deep Learning Framework

Abstract Objective We developed and validated a detection-guided artifact removal framework for clinical electroencephalography (EEG). It corrects only the contaminated segments and preserves artifact-free data. Approach The framework employs convolutional neural network (CNN) detectors trained on the Temple University Hospital (TUH) Artifact Corpus of 150 recordings from 105 patients. For eye movement artifacts (20 second windows), it uses independent component analysis (ICA) and canonical correlation analysis (CCA). For muscle artifacts (5 second windows), it employs wavelet thresholding and empirical mode decomposition (EMD). For non-physiological artifacts (1 second windows), it utilizes spherical spline interpolation and artifact subspace reconstruction (ASR). Removal is applied exclusively to detector-flagged windows, and unflagged windows remain unchanged. In a held-out test set of 21 patients and 30 recordings, we compared selective and global removal using correlation, root mean squared error (RMSE), and peak signal-to-noise ratio (PSNR). Main results Selective removal outperformed global removal in all six methods and 18 metric comparisons, with a p -value of less than 10 −105 . Selective processing maintained a clean-segment correlation above 0.987, whereas global removal reduced the correlation to values as low as 0.39 for CCA and 0.47 for ASR. CCA removed 74.6% of the eye movement artifact amplitude, EMD removed 99.8% of the high-frequency (30–40 Hz) muscle contamination, and ASR reduced the non-physiological artifact amplitude by 37.1%. The preservation of artifact-free windows remained high for all methods and indicated minimal distortion of the clean EEG. Significance Detection-guided selective removal addresses a significant limitation of global correction pipelines that can remove clean EEG signals. This framework automates artifact removal without manual review and preserves the signal fidelity for clinical interpretation. Its modular design facilitates integration into real-time monitoring systems for acute and perioperative care.