Hardware-efficient EVD processor architecture in FastICA for epileptic seizure detection

Independent component analysis (ICA) is a key signal processing technique to improve the detection accuracy of epileptic seizures. It separate artifacts and epileptic signals, which facilitates the succeeding signal processing for seizure detection. FastICA is an efficient algorithm to compute ICA through proper pre-processing. In the preprocessing stage of the FastICA, eigenvalue decomposition (EVD) is applied to reduce the convergence time of iterative calculation of weights for demultiplexing received multi-channel signals. To calculate EVD efficiently, the Jacobi method is preferable since an array structure is proposed to decompose matrix efficiently by leveraging givens rotations. Multiple diagonal and off-diagonal processing elements run in parallel to calculate EVD. The micro-rotations can be realized efficiently by coordinate rotation digital computer (CORDIC), which calculates trigonometric functions using only addition, shift, and table lookup without dedicated multipliers. In this work, an approximate Jacobi method is adopted instead to reduce the number of iterations significantly. Optimized rotation angles can be calculated efficiently using shift-add operations for multiplications with coefficients of power of 2 in the diagonal processing elements. Normalization operation in the original mathematical formulation can be omitted due to signal re-scaling in both diagonal and off-diagonal processing elements. The number of processing cycles is reduced by 6 times for each sweep due to the reduced number of pipelining stages in the critical path. The approximate Jacobi method provides a 6x speedup (185–252 cycles instead of 1440 cycles) for a 6-channel EVD. An overall 77.2% area reduction is achieved due to arithmetic simplification and hardware reduction. The hardware architecture is verified by testing the human electroencephalogram (EEG) signals from the Freiburg Seizure Prediction EEG (FSPEEG) database.