Inter-subject prediction of pediatric emergence delirium using feature selection and classification from spontaneous EEG signals

Pediatric emergence delirium is a troublesome clinical phenomenon, which may cause serious safety problems and sequelae. Predicting pediatric emergence delirium before pediatric patients regain consciousness can provide early warning for timely treatment, which helps to reduce the occurrence of adverse events. In this paper, we proposed a feature selection and classification method fusing marine predator algorithm and k-nearest neighbors, named MPA-KNN-FSC, to achieve inter-subject pediatric emergence delirium prediction from spontaneous electroencephalogram (EEG) signals. Firstly, 64-channel spontaneous EEG signals were collected from 20 pediatric patients after anesthetic surgery, and processed to extract 1792 features containing time domain, frequency domain, time–frequency domain and nonlinear features. The marine predator algorithm was then exploited to search the optimal feature subset and hyperparameters of the k-nearest neighbors, which was employed for predicting pediatric emergence delirium. To resolve the sample imbalance problem, the fitness function of the marine predator algorithm was constructed with the area under the receiver operating characteristic curve (AUROC) obtained by the k-nearest neighbors classifier. The proposed AUROC-based MPA-KNN-FSC method achieved the best inter-subject pediatric emergence delirium prediction accuracy (77.90 ± 2.59 %) and AUROC (0.871 ± 0.017), compared to the feature selection and classification methods with ten metaheuristic algorithms and the accuracy-based MPA-KNN-FSC method. These results demonstrate the feasibility, effectiveness and generalizability of the proposed AUROC-based MPA-KNN-FSC method for automatic and unsupervised physiological monitoring of pediatric patients.