An 8-Channel TDM Spectral Feature Extraction for Neuromodulation SoC

Spectral features extracted from ECoG (Electrocor-ticogram) signals are widely used in on-chip ML-based classifiers for the treatment of neurological disorders like Epileptic seizures and Parkinson's tremors through neuromodulation. Spectral feature extraction requires filtered ECoG signals of different physiological frequency bands which is done using FIR filters because of their stability and linear phase response. This paper compares the performance of the FIR filter in terms of area, power, stopband attenuation, and transition bandwidth for different choices of design windows, order, and coefficient quantization. An 8-channel TDM FIR filter architecture is proposed which also provides the option of feature rejection for compatibility with energy-aware ML classifiers. 4 different configurations of the TDM FIR filter are implemented on 65nm CMOS technology. Kaiser window was found to give the best stop band attenuation and transition bandwidth values in the ECoG signal bandwidth. The 24-order, 10-bit coefficient TDM FIR filter has an area of
$800{\mu \mathrm{m}\mathrm{x}}800{\mu \mathrm{m}}$
and consumes
$27.2\mu \mathrm{W}$
power; 24-order, 12-bit coefficient TDM FIR filter has an area of
$850\mu \text{mx}850\mu \mathrm{m}$
and consumes
$29.8\mu \mathrm{W}$
power; 32-order, 10-bit coefficient TDM FIR filter has an area of
$900\mu \text{mx}900\mu \mathrm{m}$
and consumes
$34.9\mu \mathrm{W}$
power; 32-order, 12-bit coefficient TDM FIR filter has an area of
$950\mu \text{m x}950\mu \mathrm{m}$
and consumes
$36.8\mu \mathrm{W}$
power.