A practical noise stress test to assess performance of automated photo-plethysmogram analysis

Abstract

Due to a significant spectral overlap between the motion artifact and underlying photoplethysmogram (PPG), reliable automated PPG analysis in real-life environment may be challenging. To evaluate the impact of motion artifact on the accuracy of automated PPG pulse detection, we designed a noise stress test (NST) in which artifact-bearing (noise-added) recordings are assembled from actual recordings by selecting intervals that contain predominantly motion artifact. To assemble the NST database, we analyzed 2000 synchronized electrocardiogram (ECG) and PPG recordings from MIMIC-II database. One-minute segments with the highest and lowest agreement between the ECG beats and the PPG pulses were selected using a semi-automated protocol. The resulting NST database included 52 artifact-free base recordings by visually selecting clean segments with normal pulse rate and rhythm, and 10 pure artifact recordings by selecting segments with negligible spectral content from the base signal. Cross combination of the base and artifact recordings, by calibrating the level of added artifact, generated 520 one-minute PPG signals for each desired signal-to-noise ratio (SNR). For each combined signal, the performance of automatic pulse detection and time-domain pulse rate variability analysis was evaluated by using the annotations from artifact-free base recordings as reference.