An experimental PPG system is used to isolate the arterial signal because commercial systems contain unknown analogue and digital processing elements that may favour one approach to artefact recognition (or reduction) above another. The experimental PPG system is a high-resolution two channel analogue device with electronic ambient light compensation and containing a novel motion-artefact reduction system that will be used later. The probe used in these experiments is a conventional, commercially available clip device of the type routinely used in pulse oximetry. Artefacts are induced by four methods designed to explore the range of artefacts encountered in reality, namely: finger bending (at distal and mid-phalangic joints), pressure application to the probe body and probe clip, and finally pseudo-random motions produced by waving the hand. All of these artefacts are induced across a range of amplitudes and frequency bands. 60 periods of artefact corrupted arterial PPG signals each of duration 20s are used in this study as sample sets for subsequent artefact analysis, recognition and reduction. Two distinct approaches to artefact recognition will be investigated, namely spectral and feature (or shape) based.

When the artefact is induced at frequencies predominantly out of the spectral bandwidth of the PPG signal it is possible to estimate the signal to artefact relative power by analysis of the Fourier spectrum of the corrupted signal. By sliding a window across the signal and evaluating the spectral density of artefact and signal in each window it is possible to produce an analogue representation of the proportion of artefact (relative to signal) [10]. An example of that representation is shown in Figure 1.

This technique is itself an artefact recognition method, although clearly it will not be accurate for in-band artefact. Another approach to artefact recognition attempts to identify principal characteristics or features of canonical PPG signals [7]. Generalisation of such methods is likely to be achieved if the feature sets are relatively simple. One such method for feature recognition compares the rising and falling edges of the pulsatile signal and sets thresholds on their ratio. Figure 2 shows the derivative of an artefact corrupted PPG signal (original signal is shown in Figure 1) and the peaks are identified with the help of an estimate of the pulse rate, derived from uncorrupted elements of the sequence. When the ratio of successive peaks lies outside predetermined limits (arbitrarily set) the signal is deemed to be artefact corrupted. A binary output is therefore obtained (illustrated in Figure 1) that allows suppression of the output. The performance of this method is analysed statistically, on 60 independent sample periods, in the next section.

Artefact reduction by a method that is independent of the PPG feature set and the bandwidth of artefact has recently been proposed by the authors [8]. This method, described in detail elsewhere, involves a non-linear transformation of two PPG signals separated in the optical spectrum and is optimised for artefact arising from changes in the coupling of light between the tissue and probe. The artefact reduction technique is integrated with the experimental system used to acquire artefact corrupted PPG signals. The system has the capability of synchronous output of raw signals and artefact reduced signals. It is stressed that no digital processing of the raw signals is involved with the artefact reduction method. The spectral artefact recognition technique has been applied to each sample of artefact corrupted PPG and artefact reduced PPG, resulting in a quantitative comparison between signals obtained with and without the reduction process. Application of this analysis to the whole set of 60 PPG signals with a broad range of artefact amplitude and origin results in the statistical measure of the performance.

The artefact recognition and reduction techniques were applied to 60 sample sequences of artefact corrupted PPG data and the results presented in Figures 3 and 4 were obtained.

Figure 3 plots the percentage of power in the artefact relative to the signal as recognised by the shape (or feature) based algorithm versus the artefact to signal percent power recognised by the spectral method. Note that there are twice as many points on figure 3 as there are on figure 4, because in the former the artefact recognition methods are applied to both equalised and raw signals. The artefact was intentionally induced to be spectrally separate from expected PPG signals, so that the results of spectral artefact quantification may be taken to be accurate. Although the performance of the two recognition methods will be determined by different criteria, the conditions for successful spectral identification are relatively well defined. This enables us to examine the applicability of the feature-based approach. It is clear from figure 3 that a highly non-linear relationship is indicated between the two artefact recognition algorithms. This implies that, whilst the shape-based method may be over-sensitive to artefact-induced derivatives that are large by comparison to the artefact magnitude, this technique consistently fails to correctly identify all periods of artefact corruption when the artefact magnitude is high (greater than 20% of signal power as defined by spectral quantification). Figure 4 shows the statistical results of the artefact recognition algorithm, as quantified by the spectral technique. In all cases the artefact is significantly reduced by the methodology. In addition, the residual artefact after equalisation is consistently reduced to below 20% of signal power. This implies that artefact reduction could be used to improve the success of shape-based artefact recognition, which can fail at high levels of artefact corruption. Obviously, the high degree of equalisation shown suggests that useable signals may be obtained in many cases that would otherwise be precluded by the severity of artefact. Artefact recognition should still be utilised, however, to identify cases where the reduction technique is only partially successful.

Fig.4: The effect of artefact reduction on spectral artefact quantification.

The two methods investigated in this paper for artefact recognition are very useful techniques for diagnosing the amplitude and type of artefact present. The spectral method is limited by the frequency distribution of artefact, whereas the shape based method is limited by the need for generality (simplest possible feature set) and accuracy (most complete possible feature set). It is significant that when the spectral method recognises either low or high amplitudes of artefact the shape-based method may not. Unfortunately these methods do not generally complement one another in the sense that one is more suitable when the other is less suitable, however their combination could be summarised in the following simple argument. When the spectral recognition method shows unacceptably high levels of artefact power the signal should be rejected and when the output indicates acceptable levels of artefact the feature based algorithm should be used to check the possibility of the presence of large artefact; bearing in mind that this latter algorithm may produce false negatives and positives.

Clearly artefact reduction should always be attempted prior to any recognition algorithm unless the reduction scheme has the potential to increase the level of corruption. The reduction algorithm investigated here has been tested on a wide range of artefact and has consistently improved the signal to noise (artefact) ratio. Probe coupling artefact is reduced significantly, to the extent that artefact recognition and suppression is not required. In addition, successful recognition of severe artefact that cannot be completely removed by this methodology will still simplified by the resultant improvement in signal quality.

Artefact recognition and reduction are essential tools for PPG and NIRS measurements in which motion of the subject is involved. These methods are required in combination if the performance of such systems is to be maximised and if the accuracy of the fundamental signals and hence their derived properties are to be accurate. Current technology will not perform at all in periods of medium and severe artefact, and the accuracy of derived results must be questioned during periods of mild artefact.

[1] Kumar G and Schmitt JM, "Optimal probe geometry for near-infrared spectroscopy of biological tissue", App. Opt. 36(10), pp. 2286-2293 (1997)

[2] Hertzman A.B, "Photoelectric plethysmograph of the fingers and toes in man.", Proc. Soc. Exp. Biol. Med., 37, pp. 529 (1937)

[3] Aoyagi T, Kishi M, Yamaguchi K, Wantanabe S, "Improvement of the ear piece oximeter", Abstracts Jpn. Soc. Med. Electronics Biol. Engineer, pp. 90 - 91 (1974)

[4] Blazek V, Schultz-Ehrenburg U, "Quantitative Photoplethysmography Basic facts and examination tests for evaluating peripheral vascular functions", VDI (Verein Deutscher Ingenieure) Verlag, 20(192) (1996)

[5] Jawahar D, Rachamalla HR, Rafalowski A, Ilkhani R, Bharathan T and Anandarao N, "Pulse oximetry in the evaluation of peripheral vascular disease", J Vascular Diseases 48, pp. 721-724 (1997)

[6] Norton L.H, Squires B, Craig N.P, McLeay G, McGrath P, Norton K.I, "Accuracy of pulse oximetry during exercise stress testing", Int. J. Sports Med., 13, pp. 523-527 (1992)

[7] Swedlow D.B, "Apparatus for the detection of motion transients", United States Patent No:5226417 (1993)

[8] Hayes M.J, Smith P.R, "Artifact Reduction in Photoplethysmography", App. Opt., 37(31), pp. 7437-7446 (1998).

[9] Diab, Mohamed K, "Signal Processing Apparatus", International Patent Application WO 96/12435 (1996)

[10] Hayes M.J, Smith P.R., "Quantitative evaluation of photoplethysmographic artefact reduction for pulse oximetry", in Proceedings BiOS '98 (Biomedical Optics), Stockholm, ELA / EOS / SPIE (1998).