A neural network approach for automatic detection of acoustic alarms

Abstract

Acoustic alarms generated by biomedical equipment are relevant sounds in the noisy Neonatal Intensive Care Unit (NICU) environment both because of their high frequency of occurrence and their possible negative effects on the neurodevelopment of preterm newborns. This work addresses the detection of specific alarms in that difficult environment by using neural network structures. Specifically, both generic and class-specific input models are proposed. The first one does not take advantage of any specific knowledge about alarm classes, while the second one exploits the information about the alarm-specific frequency sub-bands. Two types of partially connected layers were designed to deal with the input information in frequency and in time and reduce the network complexity. The time context was also considered by performing experiments with long short-term memory networks. The database used in this work was acquired in a real-world NICU environment. The reported results show an improvemen t of more than 9% in absolute value for the generic input model and more than 12% for the class-specific input model, when both consider time information using the proposed partially connected layer.