A Real-Time Brain–Computer Interface With FPGA-Accelerated Neural Signal Processing and Enhanced EEG Classification Using Adaptive SVM

Abstract

Electroencephalography (EEG)-based brain–computer interfaces require accurate classification of noisy, non-stationary signals with low processing latency. This work presents the Neural Signal Processing and Machine Learning System (NSPMS), an open-toolchain hardware/software platform combining RISC-V control with FPGA acceleration for multi-channel EEG acquisition, signal processing, and classification. The proposed Enhanced SVM with Dynamic Feature Weighting and Kernel Adaptation (ESVM-DFWKA) is evaluated with conventional classifiers and EEGNet on the Movement Dataset and externally checked on BCI Competition IV-2a. ESVM-DFWKA achieved <inline-formula> <tex-math notation="LaTeX">$91.4\pm 5.2$ </tex-math></inline-formula>% LOSO-CV accuracy on the Movement Dataset, comparable to EEGNet without a statistically significant superiority claim, and <inline-formula> <tex-math notation="LaTeX">$82.2\pm 20.1$ </tex-math></inline-formula>% aggregate accuracy on the public benchmark. For the evaluated ICA/FFT/SSA signal-processing workload, NSPMS achieved an aggregate <inline-formula> <tex-math notation="LaTeX">$2.21\times $ </tex-math></inline-formula> execution speedup and 71.1% lower energy consumption than the PBS baseline. Real-time performance is reported as a 4.0 s EEG observation window followed by 150.44 ms mean post-window processing latency. These results support NSPMS as an energy-efficient platform for low-latency post-window EEG processing and classification.