Sensor Sound Classification in Neonatal Intensive Care Units Based on Multiple Features and Neural Networks

Newborns with health complications frequently need to be treated in specialized units called Neonatal Intensive Care Units (NICUs). These environments require efficient monitoring and analysis. However, many factors can influence treatment phases, including sound sources and noise levels. Inadequate acoustic conditions and infrastructure can damage babies' health. Our work proposes a valuable method to enable proper monitoring and feedback to medical staff through correctly classifying the main hospital sounds. We performed sound classification in NICUs using Convolutional and Long Short-Term Memory (LSTM) Neural Networks. We focus on three audio classes: cry, human talks, and alerts from hospital machines (beep sounds). The results include extracting relevant sound features and comparing classifiers considering the main NICU sound classes. The CNN and LSTM approaches performed cry sound classification with a precision of 83.5% and 84.0%, respectively. To the alerts and talks, the LSTM approach increased CNN recall by 5.8% and 5.6%.