Prediction of oxygen saturation by pulse oximetory from image and sound data with long short-term memory recurrent neural network

Abstract

An Internet of Things (IoT) communication function was attached to inexpensive sensors such as cameras and microphones and was used for data acquisition and analysis. In this study, the value of saturation of blood oxygen measured by pulse oximetry (SpO2), which is used for sleep apnea syndrome (SAS) detection, was estimated from the data obtained from the camera and microphone. SpO2 was recorded by a pulse oximeter worn by subjects suspected of having SAS when sleeping overnight. The camera and microphone were located on the side of the bed to record the data. The SpO2 value was learned using long short-term memory (LSTM), which is one of the deep neural network methods that have shown excellent results as a method of analyzing time series data. When evaluated by leave-one-out cross validation using the data of four persons, it was found that the amplitude of the estimated SpO2 was about half. The cause seems to be individual differences in time from apnea occurrence to SpO2 declination. By doubling the estimation result, it was confirmed that the SpO2 value was well estimated.