Breathing sound detector as a means to identify possible apneic periods from tracheal sound recordings

Tracheal sound represents an easily acquired and processed signal that is particularly popular in the evolution of smartphone-based systems for Sleep Apnea Syndrome (SAS) diagnosis. SAS is characterized by partial or complete breath cessation for at least 10 s. The developed algorithms mainly rely on neural networks focusing on the Apnea/Hypopnea Index (AHI) extraction; this index corresponds to the apneic episodes’ count per sleeping hour. Though reported accurate in AHI estimation, neural networks are severely affected by the inter- and intra-patient breathing sound variability. Alternatively, breathing detection algorithms can contribute in identifying the dominant sound patterns within the apnea event. Similar works propose silence detection to locate apnea onset, eventually neglecting hypopnea, often accompanied by snoring. In this work, we employ four features: zero-crossing rate, signal power, Tsallis entropy and Shannon information to discriminate breathing from silent frames. These features are extracted independently by tracheal sound recordings from 178 patients undergoing a sleep study. A candidate apnea corresponds to silence detected by at least one feature for a minimum duration of 5 s. Additionally, a reduction of the mean signal power, in the detected breathing frames, is indicative of hypopnea. The algorithm presents a maximum sensitivity per patient 98.75 % and it can locate the end of 80.45 % of all annotated episodes (32824 out of 40800) with an error less than 10 s. Despite the non-negligible number of false positive detections, the proposed algorithm proves the dominance of the described sound pattern in the majority of the apnea/hypopnea episodes.