EEG fatigue detection stops reading band power alone: LBSNet lets learnable spectral prototypes set each window's weight
Synopsis
EEG fatigue detection has treated multi-band modeling and temporal weighting as separate concerns, without defining how the spectral structure inside a window should shape that window's contribution. LBSNet encodes each window's relative cross-band activity as a normalized descriptor, softly matches it to learnable spectral prototypes, and combines the assignments with prototype importance to weight power aggregation. In subject-independent experiments on SEED-VIG and SADT it beats seven baselines on accuracy and F1 while using the fewest parameters and multiply-accumulate operations among the compared models.
Interpretation
EEG fatigue detection gains a path that links local spectral structure to temporal aggregation: each window's relative cross-band activity is first identified as a latent brain state, and that state decides how much the window counts in the final prediction. Earlier multi-band models and adaptive temporal processing each captured a different aspect of fatigue EEG, but did not define how an interval's spectral structure should influence its temporal contribution. In subject-wise five-fold cross-validation on SEED-VIG and SADT, LBSNet's accuracy and F1-score exceed seven baselines; against the strongest baseline MB-STFormer, accuracy and F1 rise by 1.26 and 1.21 percentage points on SEED-VIG and by 2.68 and 2.77 points on SADT.
The network first decomposes EEG into five canonical bands, applies band-specific temporal convolution and spatial projections shared across bands to obtain compact features, then recalibrates the five bands' relative weights from global spectral activity. The spectral prototypes are a set of learnable cross-band activity templates; each window's normalized spectral descriptor is softly matched to them, and the assignments multiplied by prototype importance become that window's temporal weight. On SADT, LBSNet reaches a mean accuracy of 73.74% with only 2.89 M multiply-accumulate operations and 1.30 K trainable parameters, both the lowest among the compared models; against MB-STFormer it cuts MACs by 89.9% and parameters by 98.3%.
Ablations attribute the gain to prototype-guided temporal weighting, not to learnable temporal weights alone. Replacing the learned temporal weights with uniform pooling, or prototype matching with a parameter-matched MLP attention, both reduce performance. Uniform pooling lowers accuracy and F1 by 2.48 and 2.90 points on SEED-VIG and by 1.78 and 0.81 points on SADT; against the MLP attention variant, LBSNet is higher by 2.10 and 3.50 points on SEED-VIG and by 1.19 and 1.31 points on SADT.
Perspective
The conclusions rest on two driving-fatigue EEG benchmarks, SEED-VIG and SADT, under subject-wise five-fold cross-validation; SEED-VIG keeps 17 channels, downsamples to 200 Hz and segments into 8-second samples, while SADT uses all 30 channels with 3-second samples. Readers should treat these offline benchmark results as the demonstrated scope, not as validation of an in-vehicle real-time system.
The spectral interpretation of the prototypes comes from visualization on SEED-VIG: P1 accounts for 65.6% of assignments in fatigue samples and P2 for 71.6% in alert samples, and the P1 share generally rises with PERCLOS while P2 falls. These are associative observations; the material reports neither cross-dataset prototype consistency nor performance in online settings.