E-Stethoscope: Preliminary Classification of Chest Sound for Proper Intubation in Paediatrics

Tracheal intubation is a common procedure in the Neonatal Intensive Care Unit (NICU) and delivery rooms as an airway management strategy for infants especially premature babies that have respiratory complications such as pneumonia, and respiratory distress syndrome (RDS). Chest auscultation is a necessary method used to confirm the correct placement of the endotracheal tube. However, this method is highly subjective and skill-dependent, and requires a breach of personal protective equipment, especially during a pandemic. The advancement in automated lung sound analysis using artificial intelligence offers increased safety and performance of respiratory diagnosis in healthcare, particularly in resource-limited settings. However, none of the works presented so far are intubation specific for infants. The goal of this study is to fill a research gap by creating a reliable AI model that can be employed as part of an electronic stethoscope. This study used a dataset of breath sounds acquired from an infant simulator and a public dataset. To classify the lung sounds, this work employed the MFCCs summary statistics and the Mel-spectrogram as characteristic features of the audio signal to train a support vector machine and a convolutional neural network model, respectively. The study found that the support vector machine algorithm has a higher accuracy of 95.8% and overall has better performance than the CNN model. This algorithm is more reliable and protective against overfitting which allows it to provide an accurate classification for intubation-specific lung sounds.