A Neural Network Approach for Automatic Detection of Equipment Alarms in a Neonatal Intensive Care Unit Environment

The acoustic environment of a typical Neonatal Intensive Care Unit (NICU) is highly diverse and may contain a large number of sounds coming from numerous sources, such as alarm sounds generated by di erent biomedical equipment, noisy mechanical ventilation, telephone ring sound and people conversations. Various acoustic events are usually taking place simultaneously in a NICU and the maximum recommended sound pressure level limits are exceeded frequently, a fact that is of great concern according to the medical literature. In particular, equipment alarms are extensively present in a NICU environment. They are used in monitoring or supporting equipment to alert of situations requiring medical attention. This thesis work is focused on the detection of these equipment alarms by using a Neural Network (NN) architecture for binary classi cation between speci c alarm classes and background noise. As acoustic alarms have a clear frequency distribution and duration, this information can be used to train models. The database used in this work was acquired in a real-world environment in previous works on automatic event detection in NICU. It contains more than 1.5 hours of recorded audio data, and about half of it was manually annotated. In total, 7 di erent alarm classes have been selected due to their importance according to the medical sta , and due to being the most represented in the database. The features used are the spectral bins extracted from each frame and the classi cation has been performed at a frame level. Further, a post-processing step has been included after the classi cation output to evaluate the system at the period level. Two types of models are presented to address the acoustic alarm detection problem: Generic Model (GM) and Particular Model (PM). First one does not take advantage of any speci c alarm class knowledge, that is to say, the input is the whole spectrum representation of the signal and the model has to learn the alarm localization in frequency. On the other hand, the input of the second model is a concatenation of the speci c frequency sub-bands where the alarm event is located, which makes a smaller and more informative input for the speci c alarm class. Both models have been applied rst to detect the most populated alarm class from the database, and then the best architecture from each model has been used to train and detect the rest of the alarm classes also in a binary class vs. non-class way. Conventional single-layered NN has been explored in this work rst. But due to the low amount of available data, other alternatives have been considered in order to exploit the spectral information while avoiding over tting during training. Hence, partially connected hidden layers have been