Pioneering Snore-Enhanced Diagnostics with Low Data Dependency Feature Enhancement Fusing Transfer Learning and Self-Attention Mechanism

This study proposes a model based on YAMnet transfer learning and self-attention mechanism to classify OSA based on snoring sounds. The snoring sound data of 55 OSA patients were selected from the open source PSG-Audio data by the Sleep Research Department of Sismanoglio-Amalia Fleming General Hospital in Athens. 45720 0.96 second sound clips related to snoring and breathing were extracted, including 21489 normal snoring sounds and 24231 apnea or hypopnea snoring sounds. All sound clips were divided into 60% training set, 20% validation set and 20% test set. The sound clips were processed into Mel frequency spectrograms as feature inputs of the transYAMSA-Net model. The model classified snoring sounds into two groups: (1) normal snoring sounds and (2) abnormal (apnea or hypopnea) snoring sounds. The results show that the model can correctly classify 86.7% of the cases, and the model size is 14.7MB. Therefore, the model is expected to be used to screen for OSA using snoring sounds recorded by microphone on mobile devices.