Frequency Band Separability Feature Extraction Method With Weighted Haar Wavelet Implementation for Implantable Spike Sorting

Hardware-efficient feature extraction is an important step for real-time and on-chip spike sorting. Based on an analysis of spike energy spectrum, a new feature set is developed using the positive and negative spike peaks in low and high frequency bands. A separability metric that evaluates the informativeness and noise sensitivity of features is introduced to optimize the cutoff frequency of each band. Haar-based discrete wavelet transform was chosen to implement memory- and hardware-efficient filters for extracting frequency band separability features. Specifically, peaks from the first level detail and the fourth level approximation were used to represent a spike. To improve clustering performance, the detail features were weighted into the same dynamic range as the approximation features. The new feature extraction method was tested at different signal-to-noise ratios using synthesized datasets consisting of considerable and various spike shapes extracted from real neural recordings. The results show that the new method has 3%-10% better spike sorting performance than other hardware-efficient methods while consuming comparable hardware resources.