Home

Sleep Apnea Detection System

End-to-end signal-processing pipeline for automated sleep apnea detection from audio using Mel spectrograms and a custom CNN.

Stack

PyTorch, CNN, Mel Spectrograms, Apple MPS

Sleep Apnea Detection System screenshot

The Process

Developed an end-to-end signal-processing pipeline for automated sleep apnea detection from audio. The system handles segmentation, Mel spectrogram extraction, and CNN inference. It automates 30-second epoch segmentation and event merging to compute a clinical Apnea-Hypopnea Index.

The Outcome

Successfully deployed on Apple Silicon MPS for real-time inference. For example, it correctly classified a recording with 42 events over 7.19 hours resulting in an AHI of 5.83 (Mild). Tuned via F1/ROC-AUC.