Scholarlink

Obstructive sleep apnea (OSA) is a common breathing disorder that is rarely detected early because it appears normal but is associated with serious long-term health risks. While polysomnography (PSG) is paramount for diagnosis, it is also very expensive, invasive, and impractical on a large scale. To address these limitations, this study analyzes a noninvasive approach for OSA detection using acoustic analysis. A deep learning model trained with tracheal audio recordings collected during PSG sessions is proposed. The recordings were segmented into 10-s clips, transformed into Mel spectrograms, and classified using a convolutional neural network (CNN) based on the EfficientNetV2B1 architecture. The model achieved an accuracy greater than 90%, an F1 score of 91%, and an AUC-ROC of 0.96 on the test set. These results indicate the potential of audio-based deep learning systems as cost-effective and scalable tools for early detection of OSA both in clinics and remotely.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Detection of Obstructive Sleep Apnea Using Deep Learning on Audio Signals

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bryan Darquea,
  • Paulina Vizcaíno-Imacaña,
  • Solange Criollo,
  • Angie Inga-Ipiales,
  • Luis Zhinin-Vera,
  • Lenin Ramírez-Cando,
  • Carolina Cadena-Morejón,
  • Diego Almeida-Galárraga,
  • Andrés Tirado-Espín,
  • Fernando Villalba Meneses

摘要

Obstructive sleep apnea (OSA) is a common breathing disorder that is rarely detected early because it appears normal but is associated with serious long-term health risks. While polysomnography (PSG) is paramount for diagnosis, it is also very expensive, invasive, and impractical on a large scale. To address these limitations, this study analyzes a noninvasive approach for OSA detection using acoustic analysis. A deep learning model trained with tracheal audio recordings collected during PSG sessions is proposed. The recordings were segmented into 10-s clips, transformed into Mel spectrograms, and classified using a convolutional neural network (CNN) based on the EfficientNetV2B1 architecture. The model achieved an accuracy greater than 90%, an F1 score of 91%, and an AUC-ROC of 0.96 on the test set. These results indicate the potential of audio-based deep learning systems as cost-effective and scalable tools for early detection of OSA both in clinics and remotely.