Comparison of the Utility of Amplitude-Spectral and Coherence Features of Psychotropic Drugs' Action on ECoG Signal...

A naive Bayesian classifier (NBC) combined with principal component analysis (PCA) effectively differentiates the dose-dependent effects of certain groups of psychoactive drugs based on their impact on the amplitude-spectral characteristics of electrocorticograms (ECoG) in rats. This approach has been shown to be useful for pharmacological screening of agents with unknown or poorly understood activity. Despite previously obtained optimistic results, classification determination for some drugs was inaccurate, necessitating the search for possible ways to improve the predictive effectiveness of the proposed algorithm. One possible approach would be to use as input quantitative data not only the impact of the psychoactive drugs studied on the amplitude-spectral characteristics of ECoG but also connectivity changes, including the average coherence power of different pairs of leads. The aim of this study was to compare the accuracy of NBC in classifying the pharmacological mechanism of action of agents with well-known mechanisms (test set) using pharmaco-EEG data on changes in the amplitude-spectral characteristics of ECoG, coherence, and the combined use of two data sets. Materials and methods. Experiments were performed on Wistar rats with chronically implanted ECoG electrodes. The training set, relative to which the effects of the pharmacological agents from the test set were classified, were the matrices of effects of 12 pharmacological agents: the NMDA antagonist dizocilpine, the D2/D3 antagonists haloperidol and sulpiride, the M-anticholinergic tropicamide, the H1/5HT2A receptor blocker hydroxyzine, the acetylcholinesterase inhibitor galantamine, the alpha-2 adrenergic agonist dexmedetomidine, the alpha-2 adrenergic antagonist atipamezole, the adenosine receptor blocker caffeine and the GABA-mimetics aminophenylbutyric acid (phenibut), bromdihydrochlorophenylbenzodiazepine (phenazepam) and 5-ethyl-5-phenyl-2,4,6(1H,3H,5H)-pyrimidinetrione. The test set included various drugs with tropism for the targets of the training set drugs: dopamine receptor antagonists chlorpromazine, droperidol, tiapride and raclopride, H1-histamine blockers diphenhydramine and promethazine, 5-HT2-receptor blockers ritanserin and glemenserin, acetylcholinesterase inhibitor ipidacrine, alpha2-adrenergic receptor antagonist yohimbine, alpha2-adrenergic agonists medetomidine and xylazine, GABA-mimetics 5-ethyl-5-(1-methylbutyl)-2,4,6(1H,3H,5H)-pyrimidinetrione and chloral hydrate. The analysis of the ECoG signal included the calculation of 132 amplitude-spectral characteristics and 75 coherence indicators, which, using the PCA, led to new integrative indicators used for further classification of the NBC. Results and discussion. For each drug in the test set, the median similarity probability with a particular group from the training set was calculated, which was used to assess the classification quality. It was found that, when using the amplitude-spectral characteristics of ECoG, the proposed methodological approach allows for the identification of the ECoG effects of several groups of psychoactive drugs, including D2/D3-dopamine, M-cholinergic, H1-histamine, and 5-HT2-serotonin receptor blockers, AChE inhibitors, GABA-mimetics, and alpha-2-adrenergic receptor agonists and antagonists. This approach enabled the correct classification of 18 of 24 groups in the test set. When using changes in coherence indices as the initial data, the classification accuracy also amounted to 18 of 24 groups. When combining the two data sets, the number of correctly identified NBC groups was 20 of 24 groups. When comparing the classification during training (confusion matrix), it was found that coherence data or adding coherence data to the data based on changes in amplitude-spectral characteristics leads to a statistically significant (p < 0.01 in both cases) increase in accuracy. Conclusions. The obtained data demonstrated high accuracy in classifying the pharmacological activity of the test sample drugs using any of the three compared approaches. Despite the lack of statistically significant differences between them, classification based on the combined dataset demonstrated a higher number of "correct" similarities. This allows us to recommend the approach based on combined data of drug effects on amplitude-spectral characteristics and coherence as the most promising for further studies using pharmaco-EEG screening.