Performance of real time separation of multi-neuron recordings with a DSP32C microprocessor

The performance of a method for sorting of waveforms in multi-neuron data (Gädicke and Albus, 1995) is evaluated by using artificial spike patterns generated by the computer and by adding to these spikes noise or free running sine waves of varying frequency and amplitude to simulate EEG-waves. The DSP32C is capable of continuously processing spikes at 183.106 Hz. In addition to real-time sorting the DSP32C also performs a running average of the spikes sorted into each class and transfers data to the host computer. The ability of the system to analyse burst of activity is determined by the FIFO memory buffer (2048 samples, or 32.768 ms at 62.5 kHz sampling rate). Adding a 50 Hz sine wave discrimination worked correctly with sine wave amplitudes of up to 2.5 times that of the smallest spike. Combining spikes with noise revealed errors of inclusion and/or exclusion of less than 0.1% provided the models spikes were determined from noiseless spikes and the spike threshold was set above the noise peak level. When noisy spikes were used to define model spikes about 4% of the smallest amplitude spikes (signal to noise ratio 3.3) were incorrectly classified. For higher amplitude spikes (signal to noise ratio > or = 5) the classification error was on average less than 1%. The artificial patterns used for performance testing are exactly defined and could be used to standardize the comparison between different sorting techniques.