Reconstruction of Corrupted and Lost Segments from Photoplethysmographic Data Using Recurrent Neural Network

TLDR: An approach to predict the lost and highly corrupted data segments from short history (immediate proceeding four beats) of the time series PPG data based on recurrent neural network (RNN) based on RNN based data prediction model is described.

Abstract: Finger pulse signal, commonly known as Photoplethysmogram (PPG) is an important physiological signal used in intensive care unit (ICU) for heart rate and blood oxygen saturation measurement. In ICU monitoring for long-term analysis, there may be occasional clinical data corruption or loss due to either patient’s hand movement or sensor detachment from the measurement site. In this paper, we describe an approach to predict the lost and highly corrupted data segments from short history (immediate proceeding four beats) of the time series PPG data based on recurrent neural network (RNN). For identification of corrupted data segments, a support vector machine (SVM) in conjunction with Kernel radial basis function was used. The reconstruction of the lost segments and the corrupted segments from PPG data were carried out on a beat-by-beat basis, by using a joint principal component analysis (PCA) based feature extraction and RNN based data prediction model with recursive feeding of outputs to the PCA unit. Using finger PPG records of 40 volunteers, PPG beat classification sensitivity and specificity were found as of 98.1% and 91.78% respectively, with maximum absolute error (MAE) for single, consecutive five, and consecutive ten lost beats segments were 0.38%, 2.24% and 5.98% respectively.