Real-time Ventilation Detection for Feedback and Guidance During Newborn Resuscitation

For a healthcare provider treating a non-breathing newborn baby, timely ventilation of the lungs is the most critical action to save the newborn’s life. However, the guidelines for newborn resuscitation are challenging to follow and a real-time feedback and guidance system may contribute to more optimal treatment. In this study we utilized deep neural networks for detecting ventilation in the audio signal extracted from video recordings of 25 newborn resuscitation episodes. The proposed ventilation detection system utilizes Log Mel Spectrograms for feature generation, audio augmentation for dataset expansion, and 2D Convolutional Neural Networks (CNNs) for ventilation classification. For the task of image classification, we compared two different pretrained backbones suited for real-time inference and a custom-built efficient CNN, called VD-CNN. We found that using MobilenNetV3Large as the feature extractor yielded the best results, achieving a F1-score of 97.21%. Our efficient VD-CNN, with a parameter count one fifth that of MobilenNetV3Large, still achieved a robust F1-score of 95.37%. To demonstrate that the proposed system can perform ventilation detection and provide immediate feedback during newborn resuscitation in real-time, we integrated the ventilation detection system into a newborn resuscitation station called Liveborn Station.