A prototype deep learning system for the acoustic monitoring of intensive care patients

We present a prototype system for the acoustic monitoring of artificially ventilated patients in intensive care. A device placed in the patient room detects sounds indicating an emergency situation and notifies a pager of the care staff. The staff can react more quickly and take appropriate action, as well as provide feedback on the prediction for continual learning. A microphone array with adaptive beamforming and an integrated microcomputer is employed, autonomously performing recording, audio preprocessing as well as deep learning based inference. The training dataset originates from a variety of patients and spatial and sonic environments, accommodating for different patterns of background noise and distortions. Mel spectrograms of short length are extracted and used for training a convolutional neural network. An initial evaluation of the system shows an accuracy of 80% for a binary, balanced dataset. The system is deployed in several intensive care facilities and can easily be adapted to other types of medically relevant sounds.