Benchmark evaluation demonstrates efficient motor imagery classification using edge-cloud Bayesian architecture, indicating scalable, privacy-preserving brain interfaces.
Key Points
To develop and evaluate BELT, a modular Bayesian edge-cloud architecture designed to improve portability, computational speed, and data privacy in brain-computer interfaces.
Implemented BELT-lite using linear time-invariant operations and evaluated it on BCI Competition IV-2a and IV-2b motor imagery datasets (N=18 subjects total) with ten-fold cross-validation.
Assessed on-device latency on ARM Cortex-A7 hardware against EEGNet and evaluated privacy-preserving data compression via task-unaware autoencoders.
BELT-lite achieved mean posterior accuracies of 87.9% ± 6.8% on Dataset B and 80.6% ± 8.6% on Dataset A with data augmentation.
Hardware testing on ARM Cortex-A7 showed a latency of 6.75 ms per sample for BELT-lite versus 8.36 ms for EEGNet (p < 10^-17), reflecting a 21% speed improvement.
Task-unaware autoencoder compression reduced data size by 3.3× with only a ~1% accuracy drop during full posterior fine-tuning.