| Autor: |
Aghagolzadeh M, Gerhard F, Truccolo W |
| Rok vydání: |
2015 |
| Popis: |
The detection of neural events over long time periods introduces new challenges to an already complicated signal detection problem e.g. detecting spike wave discharges from local field potentials (LFPs) or action potentials from broadband recordings in human neocortex. In relatively artifact free data the background noise tends to be stable over hours/days due to the stability of resistance/capacitance features of electrodes in these time scales. However current state of the art methods for estimating detection thresholds (e.g. methods based on standard deviation (SD) of the noise) are not robust to fluctuations in ongoing neural activity and thus cannot reliably detect discrete neural events. Here we examine an approach for estimating the detection threshold for individual electrodes based on maximizing the log likelihood of a probabilistic mixture model in which a separate model is assigned for the background noise and neural signal. In particular we consider the case of a mixture of an Exponential Gaussian density function to fit the highly skewed noise distribution and a Gaussian density function to fit the target neural signal distribution. Under significant fluctuations in the level of neural activity in our datasets the parameters of the Gaussian density function (mean and variance) and the mixture weight varied substantially over time while the parameters of the Exponential Gaussian density function which identifies the distribution of the background noise tended to remain almost constant. A simpler two class Gaussian mixture model (GMM) failed to provide reasonable estimates of the detection threshold due to the skewness of the noise distribution. In comparison with the two class GMM the proposed mixture model significantly reduced the Kullback Leibler divergence between the empirical and estimated signal distribution. Visual inspection of the data indicated that the new proposed approach led to more stable estimation of the detection threshold and fewer false positives in comparison to previous SD and two class GMM methods. |
| Databáze: |
OpenAIRE |
| Externí odkaz: |
|
