Feature selection for electrical brain activity classification in newborns in case of painful events

The understanding of pain mechanisms in infants is critically important because newborns lack verbal communication abilities to report their pain experiences. This study focuses on analyzing electrical brain activity features in time and time-frequency domains using electroencephalographic (EEG) signals collected during clinically required noxious stimuli in newborns. Different feature extraction techniques are explored by applying a combined feature selection approach with forward feature selection and statistical measures involved. Six machine learning algorithms, namely Logistic Regression, Linear Discriminant, K-Nearest Neighbors, Support Vector Machines, Random Forest, and Gaussian Naive Bayes, were used and compared with the purpose of painful events recognition in newborns. Two binary classification tasks were considered: the first task was to distinguish between EEG signals before painful stimulus and after it (painless and painful state of the patient) and the second task was to distinguish between EEG signals on the background of the painful event (heel lance for blood sampling in neonates) and signals without painful events (audio stimulation). In the task of EEG signals classification into pre- and post-painful stimulus segments, the support vector machines classifier showed the best accuracy estimate of 93.5% with the pre-painful EEG segments classification accuracy of 100% and post-painful segment classification accuracy of 86.9%. In the task of distinguishing between EEG signals containing painful events as heel lance and signals without painful events, the linear discriminant algorithm showed the best accuracy estimate of 84% with 76.9% correctly determined EEG segments containing painful events and 91.6% correctly determined EEG segments without painful events. Results demonstrate the potential of using features that focus on spectral power in alpha, beta, and gamma frequency bands and machine learning techniques for advancing pain detection in neonates.