Cerebral Artery Occlusion Inference Using Pulse Waves: Selection of Dynamic Features by L1 Regularization

Abstract

Cerebral artery occlusion is one of the major causes of serious sequelae and death. It is needed to infer the existence of occlusion at the emergency site for the transportation of patients to appropriate hospitals. With the aim to establish a cerebral artery occlusion inference system that can be used in ambulances, we have developed a pulse wave measurement device using a piezoelectric sensor. In this study, we extracted dynamic features from pulse waves and applied an L1 regularization classifier for occlusion inference. Based on biomedical and physical knowledge, five types of features were extracted considering the waves reflected by a occlusion. Regarding the classifier, a multi-layer perceptron with the L1 regularization on its input layer was used to simultaneously execute occlusion inference and variable selection. The inference performance was estimated using the nested cross validation. As a result, the inference performance depends on the conditions, but it was about 80 [%] considerably higher than the chance level of 53 [%]. Moreover, the particularly important features were identified based on the weights obtained by the L1 regularization.