Abstract
Purpose: :
To determine if adaptive averaging algorithms can be used to improve the quality and speed of clinical VEP recording.
Methods: :
Visual Evoked Potentials from 5 compliant adults were recorded at 3 electrode positions (O1, Oz & O2). 200 trials were recorded in response to a reversing checkerboard (15 minutes arc @ 2Hz) stimulus (signal epochs). In addition 200 trials were recorded with the stimulus screen occluded (noise epochs). Signal and Noise epochs for each patient were randomly mixed to create a single VEP session of 400 epochs. Each VEP session was analysed using 4 different algorithms: 1)Traditional ensemble averaging assigns an equal weight to all epochs in the session; 2)A template matching algorithm, where epochs closely matching the template of an ideal VEP recording were given a high weight in the final average; 3) A coherence matching algorithm, where epochs closely matching the running average were given a higher weight; 4)A neural network algorithm was trained to assign a probability to each epoch describing the likelyhood of it originating from the signal or from the noise set. Each algorithm was analysed to determine if higher weights were assigned to epochs originating from the signal set of epochs. The final weighted average waveform was also analysed using a signal to noise ratio (SNR) paradigm.
Results: :
All non-naive algorithms assigned significantly higher weights to epochs originating from the signal set (ANOVA p<0.01). The average increase in SNR of the final VEP was 26%. Method 2 showed the greatest increase in performance with an SNR increase of 30% (+/- 5%).
Conclusions: :
The algorithms examined in this paper can be implemented inline, during recording sessions. All the methods studied increased the SNR of the final VEP however those that utilised prior knowledge had the best performance. Adaptive averaging algorithms show the potential to improve the quality of clinical VEP recording.
Keywords: electrophysiology: clinical • infant vision