Cyclic Alternating Pattern Estimation from One EEG Monopolar Derivation Using a Long Short-Term Memory

The cyclic alternating pattern is a characteristic phasic event present in the electroencephalogram signals and is commonly scored by experts through a visual examination. This pattern is considered to be a marker of sleep instability and can be used for the assessment of sleep quality. However, in manual scoring, each one second epoch of the signal is considered to be a monotonous and time-consuming task that is propitious to produce errors. Therefore, an automatic scoring algorithm is desired. The developed method uses an electroencephalogram monopolar deviation signal as input to a long short-term memory neural network to estimate the CAP phases, without the need to handcraft features. This information was then fed to a finite state machine to determine the CAP cycles occurrence. Multiple configurations of the neural network were tested and the best accuracy for the CAP phase estimation was 70%, with an area under the receiver operating characteristic curve of 0.663. Regarding the CAP cycles detection the best accuracy was 68% with an area under the receiver operating characteristic curve of 0.703. These values are in the range of what is considered to be the mutual agreement between two clinicians, analyzing the same signals. Therefore, the developed method could possibly be employed for clinical analysis.