Hybrid classical and deep learning preprocessing improves consistency of scalp EEG signals

Abstract Raw scalp EEG signals are often degraded by noise, low signal-to-noise ratio (SNR), and the high epoch-to-epoch amplitude variability, all of which can negatively affect the downstream classifier performance. The existing preprocessing pipelines typically rely either on the classical digital signal processing or learned autoencoder-based denoising but rarely integrate both while clearly distinguishing their respective roles. This paper presents a six-stage hybrid preprocessing pipeline evaluated on 56.5 hours of EEG data from 11 subjects (36 recordings) in the CHB-MIT Scalp EEG database. The classical preprocessing block—comprising common average referencing, zero-phase FIR bandpass filtering (0.5–40 Hz), a 50 Hz notch filter, and per-channel z-score normalization—improves broadband SNR by +13.99 ± 6.32 dB (from 9.55 ± 5.40 dB to 23.54 ± 6.32 dB). A per-subject fully connected denoising autoencoder is subsequently applied to overlapping 5-second epochs. While this stage reduces the broadband SNR (final: 18.32 ± 4.75 dB; Δ = − 5.22 ± 2.75 dB), it reduces inter-epoch RMS variance by 91.16 ± 5.96%, suggesting a shift in function from denoising to consistency regularization. The main contributions of this work are: (1) a hybrid preprocessing pipeline that decouples classical signal enhancement from learned representation regularization, (2) stage-wise quantification of SNR and variance to attribute the effect of each component, (3) an empirical reframing of the autoencoder as a consistency regularizer, and (4) a fully reproducible implementation using public datasets and standard tools.