Low-Power Feature Extraction for BCI Applications Using Approximate Continuous Wavelet Transform
Joe Saad, Adrian Evans, Ivan Miro-Panades, Tetiana Aksenova, Lorena Anghel · 2025
The wavelet transform is commonly used as a feature extraction method for brain signal decoding, but extracting this type of feature has a high energy cost as the wavelet transform relies on computing a convolution operation. Certain portable or implantable medical applications for Brain-Computer Interfaces (BCIs) require low-power signal decoding approaches. We investigate two methods to compute approximate wavelet transforms using fewer operations than existing techniques, which is attractive for embedded implementations. Both methods are validated on Electrocorticography (ECoG) signal decoding, reducing the energy consumption of feature extraction by up to$4\times$, with a negligible loss in decoding accuracy.