Pulse oximeter is widely used in the monitoring of blood oxygen in clinic for its convenience and efficiency. However, synchronizing light source flashing with data collecting is required, otherwise the separation of the data from different LEDs will fail. More importantly, synchronous acquisition makes the pulse oximetry system vulnerable. Meanwhile, the pulse waveform extraction is a crucial procedure in the measurement. Hence, in this paper, an asynchronous acquisition pulse oximetry system based on wavelet transform has been built. PhotoPlethysmoGraph (PPG) and photoelectric detection technology are applied in our homemade system. The adaptive soft-threshold de-noising is realized by Stein's Unbiased Risk Estimate (SURE). The principle and system configuration are described. The preliminary experiment results from wavelet transforms and Fourier transforms are compared. The results show that our homemade system is adaptive, accurate, robust and simple.
© 2013 Optical Society of America
Since 1974 Aoyagi developed an oximeter-like device to measure the arterial hemoglobin saturation , non-invasive oximetry had been taken into practical stage. The theory can be traced back to the 19th century known as the Lambert-Beer law. What should be remarked is that in 1983 Wilber applied 2 solid state LEDs as the light sources which reduced the complexity of oximeter and made it practical for clinic . Nowadays, pulse oximeter has been widely used in clinic for patients monitoring, especially during anesthesia, for its non-invasive, real-time and continuous blood oxygen measurement.
Due to dual-wavelength is used in the pulse oximetry device, synchronous clock controlling in light sources flashing and data collecting is required in conventional pulse oximeter in order to separate each of the wavelength of light, otherwise the detector may collect data at the moment when the state of light source is between on and off which makes the data separation fail. However, the synchronous relationship between light sources and detectors makes the pulse oximetry system vulnerable. On the other hand, motion artifact, baseline drift and background noises are the main interferences which prevent the pulse signals extracted correctly. Hence, signal processing is an important procedure during the measurement. Various signal processing algorithms in pulse oximetry system had been reported [3–7]. Among these, Fourier transform (FT) method and Wavelet transform (WT) method are more capable and commonly used. However, FT method can only filter out specified frequency of signals, meanwhile the pulse waves are non-stationary signals which means the frequency of pulse signals are not stable. Therefore, the results of FT method obtained are just a rough waveform. Furthermore, FT method is short of time resolution which means some details of the pulse waveform maybe lost. For dynamic measurement and non-stationary signals like pulse wave, wavelet transform is a better choice with its superior time-frequency characteristic. Hence, in this paper, an asynchronous acquisition pulse oximeter was built based on wavelet transform. In addition, to reduce the noises, Stein's Unbiased Risk Estimate (SURE) was applied to estimate the thresholds adaptively. Finally, Fourier transform (FT) was also performed as a comparison in signal processing procedure. The preliminary experiment and analysis show that our homemade pulse oximetry system is adaptive, robust, simple and accurate.
2. Description of pulse oximeter
2.1 The theoretical basis of blood oxygen measurements
In clinic, functional hemoglobin saturation which ignores the carboxyhemoglobin (COHb) and methemoglobin (MetHb) is defined asEq. (1) can be expressed as following 
In oximetry measurement, the signal is weak and coupled with various interferences, hence signal processing is crucial. FT and WT are performed respectively to get ΔA1 and ΔA2 from the original data.
2.2 System configuration
The powers of the two LEDs are the same (1 W). Each LED lights up for 10 ms, while one LED distinguishes, the other lights up. The two LEDs are as juxtaposed against each other closely, and covered with the fingertip. To collect the intensity of transmitted light, a Si-biased detector DET36A (U.S. THORLABS Company) is adopted at a sampling frequency of 20000 Hz. The sampling time is 50 s. The detector is placed at the opposite side of the fingertip as shown in Fig. 1.
Unlike conventional pulse oximeter, in which synchronous clock is applied to control LEDs’ flashing and data collection, in our system, asynchronous acquisition is realized by use of the algorithm we developed, which means LEDs flash timing is independent of the data collection.
2.3 Original signal explanation
In this section a brief explanation on the original signal will be illustrated. A part of the original signals collected by the detector are shown in Fig. 2. Due to the asynchronous acquisition applied in our homemade system, the intensities of light from both LEDs are collected indiscriminately. According to our experiments, it is found that the upper envelop of the original signal contains the data of red LED and the lower contains the data of infrared LED. Meanwhile, these signals are coupled with several types of noise. The data points between upper and lower envelops are collected by the detector at the moment when the LEDs are between on and off. We call this kind of data ‘half-lighted’. Baseline drift which has a low frequency of change and high-frequency noise are also the main influencing factors to each LED’s pulse signal which can be seen in Fig. 2. The above noises are what should be removed to obtain the proper pulse signals.
3. Principles of signal processing
3.1 WT method
Based on the priori knowledge described in section 2.3, the wavelet transform is first used in pulse signal extraction. Assuming s(t) is the signal of a single LED, the discrete wavelet transform (DWT) can be express as
Due to the time-frequency localization and multi-scale features of wavelet transform, the signal energy of pulse fluctuation is centralized in minority wavelet coefficients with larger value than the wavelet coefficients of noises. Therefore, threshold method is applied to screen out the wavelet coefficients of pulse fluctuation and the reconstruction is conducted to obtain the pulse waveform. It is clear that the quality of the threshold setting has a significant impact on the noise reduction. Hence, after careful study, Stein's Unbiased Risk Estimate (SURE)  was found as the suitable and adaptive method to estimate the threshold in our homemade system. The SURE of a single level of the wavelet coefficients can be express asEq. (6), the MSE of p(WT) can be computed by p(WT) independently. Therefore, minimizing the risks in p(WT) to obtain the adaptive threshold value . Hence, WT method becomes an adaptive algorithm as a result of the threshold selection by Stein's Unbiased Risk Estimate. The procedure of selecting the adaptive threshold is shown in Fig. 3.
To conduct the noise reduction, soft-threshold method is applied to the original wavelet coefficients (WT). Assuming ST is the threshold of a single resolution level, the soft-threshold method can be denoted asFigure 4 shows the signal processing procedure by DWT.
3.2 FT method
It is well known that the frequency of normal human pulse fluctuation ranges from 0.5 Hz to 2 Hz approximately. Depending on this priori knowledge, a frequency domain filter was designed to filter out the pulse waveform. According to section 2.3, baseline drift, ‘half-lighted’ data and high-frequency noise are the main noises in pulse signal. Among these noises, baseline drift is a low-frequency noise meanwhile data caused by ‘half-lighted’ is a high-frequency noise. Therefore, a band-pass filter is applied in signal processing procedure with the lower-cut-off frequency 0.5 Hz and higher-cut-off frequency 2 Hz.
As assumed in section 3.1, let s(t) denote the signal of a single LED, the discrete Fourier transform (DFT) can be expressed asFigure 5 shows the signal processing procedure by FT.
The system was configured as described in section 2.2. Due to asynchronous acquisition applied in our homemade system, the intensities of light from both LEDs were collected indiscriminately. In order to separate the two LEDs’ signals, the original data shown in Fig. 2 was divided into two parts through the mean value of the amplitude, and the separated signals are shown in Fig. 6 and Fig. 7 respectively.
The red LED signal (see Fig. 6) is taken for instance to illustrate the comparison of WT and FT method. In WT method, sym8 was selected as the mother wavelet and the wavelet decomposition was conducted into 11 levels to filter out the high-frequency noise and ‘half-lighted’ data. The result is shown in Fig. 8(a). As a comparison, 2 Hz was set in FT method as the higher-cut-off frequency and the result is shown in Fig. 8(b).
To obtain the baseline drift (low-frequency noise), the wavelet decomposition was conducted into 15 levels and 0 – 0.5 Hz was selected in FT method. The performance is shown in Fig. 9(a) and Fig. 9(b).
From Fig. 10, it can be seen that the pulse wave extracted by WT method has a little larger amplitude than the one extracted by FT method. What does it imply will be discussed in next section.
For the purpose of blood oxygen measurement and according to Eq. (2), the modulation degree of each incident light (ΔA1 and ΔA2) is desired. ΔA1 and ΔA2 are normalized to eliminate the detector’s different response between each wavelength and the changes of the arterial perfusion caused by pressure variation. The normalization can be expressed asEq. (2) can be re-written as
A comparison was conducted between our homemade system with 2 signal processing methods. Meanwhile an industry-accepted oximeter Prince-100H (Shenzhen Creative Industry Co., Ltd) was used to compare the measuring accuracy of the above two methods. A subject was selected to be measured. Before the measurement, the subject was required to calm down for several minutes in order to reach the resting-state, and then held his breath for a different while before the beginning of each measurement in order to show different SaO2 values . The measurement conducted by our homemade system and Prince-100H at the same time. The results are shown in Table 1 (The decimal parts of measuring results of our homemade system are kept intentionally).
It can be seen in Fig. 10 that the FT method removes noises with a slight reduction of the pulse wave amplitude while the WT method retains the reasonable details at the peak and the valley of the pulse wave. Meanwhile, in Table 1, at each stage of SaO2, FT obtains smaller SaO2 value than WT. This is because the reduction of the amplitude shown in Fig. 10. To figure out the relationship between SaO2 value and amplitude, the total differential of Eq. (13) is performed asEq. (14) can be simplified as
A new pulse oximetry system based on wavelet transform has been built. The system is simple and robust due to omitting the synchronizing data acquiring system of detector and two wave-length LEDs. Moreover, the measurement process is adaptive as a result of the adaptive algorithm is applied in the signal processing procedure which is based on Stein's Unbiased Risk Estimate. The average error is 0.378% at the present time. In addition, SaO2 values computed from WT and FT, respectively, are compared. The preliminary experimental results show that the measuring result with WT is more accurate than that with FT.
The authors acknowledge the finical support from Chinese National Natural Science Foundation 51275033.
References and links
1. T. Aoyagi, M. Kishi, K. Yamaguchi, and S. Watanabe, “Improvement of the earpiece oximeter,” in Abstracts of the Japanese Society of Medical Electronics and Biological Engineering (Japanese Society of Medical Electronics and Biological Engineering, Tokyo, 1974), pp. 90–91.
2. S. A. Wilber, “Blood constituent measuring device and method,” U.S. Patent No. 4,407,290, Washington, DC: U.S. Patent and Trademark Office (1983).
3. P. S. Addison and J. N. Watson, “A novel time–frequency-based 3D Lissajous figure method and its application to the determination of oxygen saturation from the photoplethysmogram,” Meas. Sci. Technol. 15(11), L15–L18 (2004). [CrossRef]
5. F. U. Dowla, P. G. Skokowski, and R. R. Leach, Jr., “Neural networks and wavelet analysis in the computer interpretation of pulse oximetry data,” in Proceedings of IEEE Conference on Neural Networks for Signal Processing (IEEE Signal Processing Society Workshop, Kyoto, 1996), pp. 527–536. [CrossRef]
6. S. Lee, B. L. Ibey, W. Xu, M. A. Wilson, M. N. Ericson, and G. L. Coté, “Processing of pulse oximeter data using discrete wavelet analysis,” IEEE Trans. Biomed. Eng. 52(7), 1350–1352 (2005). [CrossRef] [PubMed]
7. Y. Yong-sheng, C. Y. Poon Carmen, and Z. Yuan-ting, “Reduction of motion artifact in pulse oximetry by smoothed pseudo Wigner-Ville distribution,” J. NeuroEng. Rehabil. 2(3), 9 (2005).
8. X. U. Kexin, G. A. O. Feng, and Z. H. A. O. Huijuan, Biomedical Photonics, 2nd ed. (Science Press, 2010), p. 182.
9. C. M. Stein, “Estimation of the mean of a multivariate normal distribution,” Ann. Stat. 9(6), 1135–1151 (1981). [CrossRef]
10. D. L. Donoho and I. M. Johnstone, “Adapting to unknown smoothness via wavelet shrinkage,” J. Am. Stat. Assoc. 90(432), 1200–1224 (1995). [CrossRef]
11. J. A. Dempsey and P. D. Wagner, “Exercise-induced arterial hypoxemia,” J. Appl. Physiol. 87(6), 1997–2006 (1999). [PubMed]