Single Sensor Techniques for Sleep Apnea Diagnosis Using Deep Learning

Abstract

A large number of obstructive sleep apnea (OSA) cases are under-diagnosed due unavailability, inconvenience or expense of sleep labs. Hence, an automated detection by applying computational techniques to multivariate signals has already become a well-researched subject. However, the best-known techniques that use various features have not achieved the gold standard of polysomnography (PSG) tests. In this paper, we substantiate the medical conjecture that OSA directly impacts body parameters such as Instantaneous Heart Rate (IHR) and blood oxygen saturation (SpO2). We then use a deep learning technique called LSTM-RNN (long short-term memory recurrent neural networks) to experimentally prove that OSA severity detection can be solely based on either IHR or SpO2 signals, which can be easily, obtained using off-the-shelf non-intrusive wearable single sensors. The results obtained from LSTM-RNN model shows an area under curve (AUC) of 0.98 associated with very high accuracy on a dataset of more than 16,000 apnea non-apnea minutes. These results have encouraged our collaborating doctors to further come up with a diagnostic protocol that is based on LSTM-RNN, SpO2, and IHR, thereby increasing the chances of larger adoption among medical community.