Signal quality measures for pulse oximetry through waveform morphology analysis

Abstract

Pulse oximetry has been extensively used to estimate oxygen saturation in blood, a vital physiological parameter commonly used when monitoring a subject's health status. However, accurate estimation of this parameter is difficult to achieve when the fundamental signal from which it is derived, the photoplethysmograph (PPG), is contaminated with noise artifact induced by movement of the subject or the measurement apparatus. This study presents a novel method for automatic rejection of artifact contaminated pulse oximetry waveforms, based on waveform morphology analysis. The performance of the proposed algorithm is compared to a manually annotated gold standard. The creation of the gold standard involved two experts identifying sections of the PPG signal containing good quality PPG pulses and/or noise, in 104 fingertip PPG signals, using a simultaneous electrocardiograph (ECG) signal as a reference signal. The fingertip PPG signals were each 1 min in duration and were acquired from 13 healthy subjects (10 males and 3 females). Each signal contained approximately 20 s of purposely induced artifact noise from a variety of activities involving subject movement. Some unique waveform morphology features were extracted from the PPG signals, which were believed to be correlated with signal quality. A simple decision-tree classifier was employed to arrive at a classification decision, at a pulse-by-pulse resolution, of whether a pulse was of acceptable quality for use or not. The performance of the algorithm was assessed using Cohen's kappa coefficient (κ), sensitivity, specificity and accuracy measures. A mean κ of 0.64 ± 0.22 was obtained, while the mean sensitivity, specificity and accuracy were 89 ± 10%, 77 ± 19% and 83 ± 11%, respectively. Furthermore, a heart rate estimate, extracted from uncontaminated sections of PPG, as identified by the algorithm, was compared with the heart rate derived from an uncontaminated simultaneous ECG signal. The mean error between both heart rate readings was 0.49 ± 0.66 beats per minute (BPM), in comparison to an error value observed without using the artifact detection algorithm of 7.23 ± 5.78 BPM. These results demonstrate that automated identification of signal artifact in the PPG signal through waveform morphology analysis is achievable. In addition, a clear improvement in the accuracy of the derived heart rate is also evident when such methods are employed.

Signal Quality Measures on Pulse Oximetry and Blood Pressure Signals Acquired from Self-Measurement in a Home Environment

“…1. Note that the ECG was simultaneously recorded in order to be used as a reference signal for pulse oximetry and BP signals as described in detail in [13] and [14].…”

“…The GS development was based on the same methods applied in [13] and [14], for pulse oximetry and BP, respectively. In the development processes, two human experts (known as Rater 1 and Rater 2) for each signal first manually annotated the recorded signal to classify any noise section in the signals.…”

“…Pulse oximetry and BP signal quality algorithms have been developed by our research group to detect and eliminate artifact in noise contaminated pulse oximetry signals and to assess the quality of an NIBPM measurement, respectively [13], [14].…”

“…The main objective of this paper is to assess the performance of the developed algorithms as in [13] and [14] when measuring the quality of the respective signals acquired from selfmeasurement performs by the subjects in their homes, which involved the presence of a trainer nearby but not actually intervening during the subject's signal acquisition process. This is a preliminary study before the developed algorithms are tested with the actual data recorded in a telehealth system.…”

What Does Big Data Mean for Wearable Sensor Systems?

“…Some notable mentions involve poor device affixation or failure to wear the sensor when recording inertial signals [15], and electrode movement or detachment during ECG recording [16]. Verifying the quality of data acquired during unsupervised monitoring is currently an active research area which employs both hardware and algorithmic solutions to ensure that poor quality data is identified and rejected [17][18][19]. It remains to be seen whether the complex relationships and context present in big data sets will allow the negative impact of these noise artifacts to be further mitigated.…”