Statistical analysis of neural spike trains for estimation of functional differences in subcortical structures of human brain

In this project we studied spontaneous single unit activity in subthalamic nucleus (STN) of Parkinson’s disease patients. Microelectrode recording (MER) was performed during deep brain stimulation (DBS) stereotactic neurosurgery. The aim of the current study is to find statistical properties of neural spike trains in STN that differ anesthetized and awake states. In our research we analyzed spontaneous activity of 114 cells and 183 cells of 8 Parkinson’s disease patients in anesthetized and awake state respectively.

We calculated multiple statistical features of MER data: Integrated Spectral Power (with ranges 1-3Hz, 3-8Hz, 8-13Hz, 13-30Hz, 30-100Hz), oscillation scores (with ranges 3-8Hz, 8-12Hz, 12-20Hz, 20-30Hz, 30-60Hz, 60-90Hz), several basic descriptions of firing and measures describing bursting. We got 21 features in total.

To find statistically significant features we used logistic regression and calculated p-values for every features and got 5 features with p-value less than 0.05: local variance, burst rate, oscillation score in 3-8Hz, 30-60Hz, ISI (interspike interval) standard deviation and cross-covariance. We checked features and residues for normality using Shapiro test. Also, we built classifier to discriminate spike trains from patients in anesthetized and awake states. We used gradient boosting trees in library xgboost and had achieved 90.6% accuracy with AUC score 0.955. Features with best F-score for classifying model are: oscillation score in range 3-8Hz, 8-12Hz, ISI (interspike interval) mean, modalirity burst, local variance.

We found set of the features that most effectively discriminate the neuronal activity in the awake and anesthetized state in patients with Parkinson’s disease. The reasons for these differences are debatable and require future work.