Classification of Epileptic Motor Manifestations and Detection of Tonic-Clonic Seizures With Acceleration Norm Entropy
Résumé
In this paper, three triaxis accelerometers positioned on the wrists and the head of epileptic patients submitted to long-term video electroencephalographic monitoring as part of presurgical investigation are evaluated to characterize the different classes of motor manifestations observed during seizures. Quadratic discriminant classifiers are trained on features extracted from 1 or 4 s windows. It is shown that a simple rule applied to the acceleration norm entropy HnA produces the best performances compared to other classifiers trained on other feature sets. The simple rule is as follows with values given in bits: (0 <; HnA <; 1.34), no movement; (1.34 <; HnA <; 3.87), tonic manifestations; (3.87 <; HnA), tonic-clonic manifestations. For this classifier, features are extracted from 1 s windows and the misclassification rate is 11% evaluated on 5 607 s of epileptic motor manifestations obtained from 58 seizures in 30 patients. A quantile normalization can improve the results with features based on absolute power spectral density but performances are not as good as the ones obtained with HnA. Based on the classifier using only HnA, a simple tonic-clonic seizure detector is proposed and produces a 80% sensitivity with a 95% specificity.