Implementation of Bagged SVM Ensemble Model for Classification of Epileptic States Using EEG

Machine learning classifiers for electrode selection in the design of closed-loop neuromodulation devices for episodic memory improvement

Successful neuromodulation approaches to alter episodic memory require closed-loop stimulation predicated on the effective classification of brain states. The practical implementation of such strategies requires prior decisions regarding electrode implantation locations. Using a data-driven approach, we employ support vector machine (SVM) classifiers to identify high-yield brain targets on a large data set of 75 human intracranial electroencephalogram subjects performing the free recall (FR) task. Further, we address whether the conserved brain regions provide effective classification in an alternate (associative) memory paradigm along with FR, as well as testing unsupervised classification methods that may be a useful adjunct to clinical device implementation. Finally, we use random forest models to classify functional brain states, differentiating encoding versus retrieval versus non-memory behavior such as rest and mathematical processing. We then test how regions that exhibit good classification for the likelihood of recall success in the SVM models overlap with regions that differentiate functional brain states in the random forest models. Finally, we lay out how these data may be used in the design of neuromodulation devices.

A method for identifying moonlighting proteins based on linear discriminant analysis and bagging-SVM

Moonlighting proteins have at least two independent functions and are widely found in animals, plants and microorganisms. Moonlighting proteins play important roles in signal transduction, cell growth and movement, tumor inhibition, DNA synthesis and repair, and metabolism of biological macromolecules. Moonlighting proteins are difficult to find through biological experiments, so many researchers identify moonlighting proteins through bioinformatics methods, but their accuracies are relatively low. Therefore, we propose a new method. In this study, we select SVMProt-188D as the feature input, and apply a model combining linear discriminant analysis and basic classifiers in machine learning to study moonlighting proteins, and perform bagging ensemble on the best-performing support vector machine. They are identified accurately and efficiently. The model achieves an accuracy of 93.26% and an F-sorce of 0.946 on the MPFit dataset, which is better than the existing MEL-MP model. Meanwhile, it also achieves good results on the other two moonlighting protein datasets.

The relationship between snoring sounds and EEG signals on polysomnography

PURPOSE

The aim of this study was to analyze the relationship of snoring sound signals obtained by polysomnography (PSG) in the sleep laboratory with cortical EEG (6 channel) signals to find answers to two important questions that have been covered to a limited extent in the literature: (1) Would the sounds generated by a snoring individual have an effect on the cerebral electrical waves occurring during sleep (specifically deep restorative sleep)? (2) Would the snoring sounds of an individual being examined by PSG have more of an effect on any one of the EEG electrodes?

METHODS

PSG recordings were obtained from volunteers with primary snoring and those with obstructive sleep apnea syndrome (OSAS) on six different EEG channels (F4-M1, C4-M1, and O2-M1, F3-M2, C3-M2, and O1-M2). The relationship of each of these recordings and snoring sound signals was analyzed by using a computer-based electrophysiological signal analysis method. A three-tier approach was used in this relationship: "Feature extraction, Feature selection, and Classification".

RESULTS

Data were obtained from a total of 40 volunteers (32 men, mean age (± SD) 47.5 ± 3.2 years), 20 with primary snoring and 20 with OSAS. The discrete wavelet transform (DWT) feature extraction method was the most successful method, and by utilizing this method for analyzing EEG channels, snoring sound signals were found to affect the C3-M2 channel the most (Duncan test, p < 0.05). Delta wave frequency levels during snoring were decreased compared to both before snoring (p = 0.160) and after snoring (p = 0.04) periods (paired sample test).

DISCUSSION

When snoring sounds and EEG signals were analyzed for frequency, time, and wave conversion with feature extraction methods, the C3-M2 channel was to be found the most affected channel. The sleep physiologist who made the PSG analyses reported that, among the 6 EEG channels analyzed for periods where there was no apnea or hypopnea events but only snoring, C3-M2 was the channel showing changes in delta wave activity.

CONCLUSION

Our study showed that the monotonous and repetitive snoring sounds of the snorer do not wake the individual, but do affect deep restorative sleep (N3). PSG signal analysis revealed that the most significant changes were in the C3-M2 channel (N3 delta wave amplitude increase and frequency decrease during snoring). Thus, clinicians may be able to monitor the characteristic changes occuring in large cortical delta waves in snoring individuals with innovative single-channel EEG devices without microphones.