Abstract
Purpose: :
To develop an optimal method for estimating latency of mfVEP using Gaussian wavelet transform and a classifier to first discard noisy signals. The classifier avoids the need for manual inspection. Reproducibly identifying mfVEP peaks has previously posed a problem which limits clinical applications.
Methods: :
mfVEPs were recorded from 10 normal subjects in 2 sessions 1 week apart. Two bipolar electrodes were placed 2.5 cm above the inion and 4.5 cm below. The stimulus was cortically scaled pattern-reversal checkerboard of 24 sectors (3 rings). Initially, each mfVEP sectoral trace was inspected and noisy sectors were discarded. The first 280ms was then cross-correlated with 2nd order Gaussian wavelets centered at 120ms. Latency of the mfVEP corresponded to timing of the largest peak from the cross-correlation, offset by 120ms. Secondly an algorithm to classify noisy mfVEP was developed. Ten leave-one-out (LOO) cases were created according to subjects. In each LOO case, a classifier to detect noisy signals was trained on nine subjects (training) and classified the left-out subject (test). Each mfVEP was transformed to 35 features and various combinations of 3 features were generated yielding 6545 feature sets. For each set, the training data was split into training and validation for LOO classification. One-way MANOVA was applied on each feature set from the training data yielding canonical eigenvectors to classify the validation data. ROC curve and its area-under-the-curve were evaluated. Feature sets that corresponded to the top 10% area-under-the-curve were selected. The final classifier selected the top 3 features. The test mfVEPs that were classified as not noisy in both recording sessions were fed to the wavelet-based algorithm to measure their latency.
Results: :
Using the wavelet-based algorithm on the manually inspected mfVEPs, the smallest inter-session difference in latency is 2.6+/-2.5ms which corresponds to the peak at 112+/-6.8ms. The LOO classifications identify signal-noise-ratio, magnitude at 8-12 Hz and noise magnitude as the top 3 features. The mean classification success rate is 92%. The resulting inter-session difference of the test mfVEPs is -0.4+/-7.6ms. The corresponding peak is at 109+/-10.5ms.
Conclusions: :
The algorithm based on cross-correlation with 2nd order Gaussian wavelets can estimate latency of mfVEP with very low inter-session variability. The proposed classifier algorithm can detect noisy mfVEP successfully. The resulting latency also has low inter-session variability.
Keywords: electrophysiology: non-clinical • neuro-ophthalmology: optic nerve • visual cortex