|
1.IntroductionThe contactless acquisition of vital signs allows a convenient medical assessment and enables clinical and out-of-hospital applications. Various systems and principles for contactless measurements have been introduced in recent years.1,2 Among such approaches, the usage of cameras, referred to as camera-based photoplethysmography or photoplethysmography imaging, is one promising solution to assess the cardiac pulse in a very user-friendly setting. The acquisition of the cardiac pulse using cameras was first demonstrated by Huelsbusch and Blazek.3 Meanwhile, many researchers have addressed the camera-based photoplethysmogram (cbPPG), most often to assess the heart rate.4–10 The most important drawback of the technique is its susceptibility to artifacts induced by movements and changes in illumination. Sophisticated image and signal processing techniques are required to cope with such factors and facilitate the camera-based assessment of cardiac pulse even under real-world conditions. Poh et al.6 were the first to use blind source separation (BSS) algorithms in the context of the cbPPG. Since then BSS algorithms became a core part of signal processing schemes to extract the heart rate from cbPPG recordings. BSS aims at separating the desired signal content (i.e., cardiac pulse) from noise and artifacts by means of decorrelation and utilizing the concept of statistical independence. Principal component analysis (PCA) and independent component analysis (ICA)11 realized as Joint approximation diagonalization of eigen-matrices (JADE)12 or FastICA13 are typical BSS techniques that have been widely applied to cbPPGs. Common approaches6,14–19 use different color channels [typically red, green, and blue (RGB)] extracted from regions of interest (ROI), typically the face, as input to PCA or JADE ICA. FastICA has also been applied to RGB signals8,19,20 and achieved a slightly better performance in comparison to other ICA algorithms.8 Tsouri et al.21 proposed a constrained ICA for RGB information of a face ROI. Other researchers have further developed the idea of applying multispectral cbPPG to PCA/ICA but have used alternatives to RGB, namely combinations of RGB with orange and cyan channels or chrominance as well as hue and infrared-based signals, respectively.9,15,22–25 In addition to wavelength-based considerations, more selective ROI choices, such as reducing the face ROI to a more concise area, have been outlined in the context of PCA/ICA.8,10,16,19,22,26 These approaches seek to exclude regions that are not supposed to contribute with useful signals but can introduce distortions, e.g., by mouth movements during speaking/smiling or blinking eyes.19,27 Approaches, which are described in literature, typically rely on a spatial preselection and use multispectral information (RGB) as input to BSS techniques. Moreover, a monochrome cbPPG, extracted from the forehead, was used as input for spatio-temporal ICA.26 Wang et al.23 alternatively addressed spatial selection without using explicit face detection. The authors utilized the temporal behavior of pixel traces to distinguish skin-like areas showing temporally periodic content from motion-like content. Even Guazzi et al.28 pursued the idea of spatially selecting ROIs according to the local distribution of signal quality. Despite the frequent multispectral BSS use, there is no consensus on performance improvements by using BSS techniques with multispectral inputs. In particular, Kwon et al.17 described a blurred spectral peak after applying RGB ICA as well as an increased heart rate error. Christinaki et al.8 identified only subtle improvements but similar heart rate errors with/without using RGB ICA. Feng et al.10 showed a lack of robustness by applying standard approaches that use ICA with RGB channels. A possible reason for the inadequacy of applied BSS techniques can be the assumption of a linear mixing process13 of available sources in standard PCA/ICA. In particular, the wavelength-dependent penetration depth into human skin3 could introduce nonlinear mixing behavior, which may degrade the performance of BSS algorithms. Other factors may also impact the success of BSS techniques in extracting pulsatile signals. For example, an effect of ROI size was addressed by Mannapperuma et al.29 in the context of multispectral BSS. Based on BSS’s multiple use and high capacity on the one hand, conflicting findings and oppositional statements on the other hand, this contribution investigates the performance of BSS to enhance the cardiac pulse from cbPPGs in dependency to varying input data characteristics. We use standard BSS techniques to compare BSS’s application on multispectral as well as monochrome inputs to identify beneficial conditions for efficient usage of these techniques. To facilitate an appropriate comparison, we further introduce a spatial ROI selection based on cbPPG’s signal quality. Finally, this contribution shall allow for a targeted application of BSS techniques to cbPPG recordings to achieve a higher signal quality and make the technique applicable under real-world conditions. 2.Materials and Methods2.1.Patients and Clinical SettingThe data for this study were gathered within the scientific project “CardioVisio—Contactless aquisition of vital parameters.” Measurements were carried out at the cardiac surgical intensive care unit at the Heart Center Dresden, University Hospital Carl Gustav Carus, TU Dresden, Dresden, Germany. The project was approved by the Institutional Review Board of the TU Dresden (IRB00001473, EK168052013). Patients after elective cardiac surgery were included if they gave written, informed consent prior to surgery. Video data were recorded during the immediate recovery from surgery after admission at the intensive care unit.30 Postoperative care followed clinical standards including mechanical ventilation and external cardiac pacing by temporal atrial and ventricular wires adjusted to intrinsic cardiac rhythm and haemodynamic needs. A four-lead electrocardiogram (ECG) and finger photoplethysmogram (PPG) were simultaneously recorded at 300 Hz and used to derive the reference heart rate. 2.2.Video Recording and SelectionVideo data were recorded using an industrial camera (IDS UI-3370CP-C-HQ, IDS Imaging Development Systems GmbH, Obersulm, Germany, 100 fps, , RGB ). The camera was placed at a distance of to 100 cm to patients’ faces. Clinical ceiling fluorescent lamps served as the primary illumination source. However, the luminous color, intensity, and homogeneity of the illumination varied across the measurements due to varying patient positions with respect to the illumination, varying room geometries, and entering daylight. Therefore, a broad range of illumination characteristics is covered by the used data. To use only suitable data for further analysis, we restricted our analysis to data segments that showed high-quality reference PPGs to correctly identify the true heart rate. Furthermore, data segments with severe cardiac disorders were excluded. Only continuous segments with a minimum length of 500 s (one per patient) were considered for further processing. Based on such criteria, overall recordings of 18 patients (13 males and 5 females; 30 min per recording) were selected from a larger collective of 76 patients. The selected material included a total of about 6 h video data (average length per patient). The selection did not consider video quality, i.e., slight patient motion as well as illumination inadequacies (changes or insufficient lightning) persisted in the dataset. 2.3.Signal ExtractionThe selected video data were processed in windows of 10 s resulting in 2197 windows ( per patient). The cbPPGs were extracted in three ways as depicted in Fig. 1. To allow a spatial selection of desired ROIs, every video frame was covered by overlapping square ROIs (50% overlap at each direction) of . The ROI placement is indicated in Fig. 1(a). The ROI size was chosen since own prior investigations addressing the relationship between ROI size and signal quality [signal-to-noise ratio (SNR)] using a comparable technical setup (i.e., camera sensor, image resolution, and camera distance to subject) showed that no higher signal quality can be obtained by further increasing the ROI size. [Other investigations,31 which showed an appropriate ROI size for pulse wave extraction to be larger (100 to 150 pixels ROI side length), used a higher image resolution.] To compare the spatial ROI selection to standard ROI selection prior to BSS,6,8,22 manually annotated ROIs were used to extract the cbPPG of the complete face () and the forehead-cheeks region (), respectively. See Figs. 1(b) and 1(c) for exemplary ROI annotations. cbPPG was extracted from each (with ) as well as from and at every wavelength by averaging its pixels values4 for each frame. See Fig. 2 for exemplary signals together with the reference PPG. 2.4.Signal Processing2.4.1.PreprocessingEach 10-s cbPPG signal (from , , and ) was normalized by a three-step procedure. The signal was linearly detrended followed by 0.5 Hz high-pass filtering (fifth-order Butterworth) to limit low-frequency content below an expected heart rate.32 Furthermore, the signal amplitude was normalized by subtracting its mean and dividing the result by signal’s standard deviation. Suchlike preprocessed (and , ) were used for further processing. 2.4.2.Definition of Inputs to Blind Source SeparationTo evaluate BSSs’ benefit in consideration of varying inputs, different input sets were defined. Each input set contained three input signals to reflect the common number of input channels when RGB videos are used. The input sets differed regarding the wavelength(s) to be used and the frequency content of chosen ROIs. Regarding the wavelength, we distinguished using cbPPGs from the green channel (monochrome approach) or using RGB channels (multispectral approach). Regarding the frequency content, we distinguished using ROIs, which showed equal dominant frequencies and using ROIs, which showed differing dominant frequencies (“dominant frequency” refers to the location of the global maximum in the fast Fourier transform of the cbPPG signal from a ROI after applying a Hanning window and zero padding to 4096 points). To select three ROIs from , we further distinguished between a deterministic choice and a random choice. The deterministic choice selects three ROIs that showed the highest signal quality. For the random selection, three ROIs, which possess the desired frequency content, were chosen randomly (independently of its signal quality). To give an example, the selection “equal dominant frequency + deterministic choice” means that ROIs were first ordered according to their dominant frequency. Afterward, the ROIs that showed the highest SNRs within the desired dominant frequencies were selected. Note that it has to be defined which dominant frequency is the desired one. As there is no unambiguous answer to this question, we decided to create input sets for the three most often occurring dominant frequencies (i.e., for “equal dominant frequency” always three different input sets were used). A detailed mathematical definition of equal/differing dominant frequency and the definition of signal quality are depicted in Appendix A (a graphical overview on the selection process is also provided in Fig. 8). Table 1 summarizes the resulting input sets. Table 1Definition of input sets S for BSS.
Note: See Appendix A for a detailed mathematical description.*Note that in case of equal dominant frequency, three different input sets were evaluated for the three most occurring dominant frequencies.Set IDs refer to MC: monochrome, MS: multispectral, R: random SNR and standard ROI sets from F: face and FhC: forehead and cheeks. 2.4.3.Blind Source Separation ProcessingEvery set was further processed with PCA and ICA, respectively. The PCA was computed without dimension reduction by conducting the singular value composition of the covariance matrix of each set.11 Let us consider a set where denotes a preprocessed cbPPG of the dimension (window length ) extracted using . A transformation was computed that satisfies with as a diagonal matrix with squared singular values (equivalent to the eigenvalues of the covariance matrix) as diagonal entries and as covariance matrix of . Accordingly, the orthogonal transformation matrix consists of the covariance matrix’s eigenvectors of and diagonalizes . The PCA output was obtained byThe output was further normalized (according to the preprocessing) prior to evaluation. As representative ICA method, the FastICA algorithm13 was chosen to compute a transformation , which aims at statistical independence of the output because Christinaki et al.8 have shown a superior performance of FastICA compared to JADE for processing the cbPPG to extract the heart rate. Prewhitening of was used. Prewhitening works according to the PCA while ensuring the identity matrix, which addresses an orthonormal output. The FastICA was initialized with a fixed random demixing matrix , which was used as starting point for every processed segment. The FastICA was symmetrically conducted for dimension preservation between and . The standard tanh-nonlinearity was applied as contrast function, which supports super-Gaussian source extraction,13 as indicated for the PPG signal by Tsouri et al.21 Simultaneously, it does not aim at highly super-Gaussian signals, which is a consequence of Morris et al.33 selecting the PPG component after ICA by using the lowest kurtosis of the components.2.5.Evaluation MetricsThe signal quality of inputs, i.e., of each single cbPPG, and outputs, i.e., each independent component/principal component, was assessed by a spectral SNR, which was proposed by de Haan and Jeanne.14 The SNR considers the true heart rate , which is gained from reference recordings. Based on the true heart rate, a binary mask was defined according to sustains the spectral indices of the heart rate as well as its first harmonic. The precision refers to the accuracy demanded for heart rate meters specified in ANSI/AAMI EC13:2002.32 The SNR was calculated from a given amplitude spectrum by For this contribution, the true heart rate was estimated by first averaging manually annotated beat-to-beat intervals from the reference ECG to obtain . Afterward the closest frequency peak in the amplitude spectrum of the reference PPG was searched (). Input sets were typically formed of cbPPGs with three different dominant frequencies, respectively, signal qualities (see Appendix A). Accordingly, one obtains input signals as well as output signals after BSS with respective values ( and only contain input and output signals). Since the automated selection of an appropriate output component is not in the scope of this contribution (but has been considered by our group),34 we always selected the best possible SNR of inputs or BSS outputs for the further evaluation. This procedure comprises assessing only the highest input and output SNR out of three dominant frequencies. 2.6.Statistical AssessmentThis contribution investigates the performance of BSS to enhance the cardiac pulse from cbPPG in dependency to varying input data characteristics. To that end, the results and respective statistical analyses are broken down into three aspects:
The first aspect was addressed by a comparison of input SNRs using one-way analysis of variance (ANOVA). As posthoc tests, a selection of 13 pairwise t-tests with Bonferroni–Holm correction was applied (see Table 2 for the test selection). We selected pairwise comparisons of the deterministic small-sized ROIs (, , , and ) plus the tests of the random small-sized ROI selection matched with and (i.e., and ). Moreover, pairwise comparisons of all deterministic multispectral sets were conducted (, , , and ). To avoid the large sample size to determine the statistical results,35 we calculated subjects’ means and applied the statistical analysis to the mean values (i.e., for ANOVA and posthoc analysis). Furthermore, the effect size measure Hedges ,36 including the 95% confidence interval (CI) of ,37 is used as standardized mean difference between groups. The interpretation of is straight forward: given a comparable CI, the larger the effect size, the bigger the impact of an experimental variable.36,38 As contextual information is required to interpret effect sizes in terms of absolute values,38 we abstain from interpreting ’s absolute value. To define which effects are relevant for further discussion, we instead introduce the concept of CI consistency: an effect is regarded as consistent, if the CI of a given is completely positive or negative, respectively (denoted as consistent effect). Table 2Results of the comparison of input sets.
Note: Pairwise results: Bonferroni–Holm corrected p values from posthoc pairwise t-tests, effect size g and 95% CIs of g37 in brackets. For the second aspect, the comparison of output SNRs does not suffice the needs for a statement on the BSS performance, because the BSS performance (difference between output and input SNR) and the output SNR must be assumed to be heavily dependent on the input SNR. A statement, which bases solely on output SNRs, thus, might favor the output featuring the highest input SNR and will not provide a meaningful statement on BSSs’ performance. A dependence on the input SNR implies using the input SNR as covariate, i.e., analysis of covariance (ANCOVA). A suchlike analysis provides a meaningful statement on the benefit of applying BSS. However, a poor input SNR will be favored by this analysis as a large improvement can be obtained while the outcome still might be worse than using another input. For high input SNRs, on the other hand, the potential improvement, which can be gained by BSS, is limited as the output SNR is bounded. For such reasons, we decided to combine both analyses, ANOVA and ANCOVA. An ANOVA to the output SNRs and respective posthoc tests could be used as described before. For ANCOVA, the input SNR served as covariate. As posthoc tests for ANCOVA, -tests with centered mean39 were applied and Hedges was calculated (again subjects means were used to avoid that large sample sizes determine the statistical results). The selection of posthoc tests applied for the first aspect was also used for the second aspect. For the pairwise comparison of two settings, one of them was regarded as superior and relevant for further discussion if ANOVAs and ANCOVAs posthoc test show consistent effects with the same sign (both CI entirely positive or negative, for a detailed example, see the description of results in Sec. 3.2). Additionally, the question if a significant could be achieved by applying a BSS algorithm to a single input was answered by pairwise -tests of the differences between output and input SNRs. The third aspect was assessed by pairwise -tests of ICA and PCA outputs. Again, subject means were used and additionally Hedges was considered. 3.Results3.1.Signal Quality of Input SetsThe ANOVA yields a highly significant difference between inputs (). Table 2 shows the results of the pairwise comparison between input sets (i.e., significance from posthoc test and results on Hedge’s , the input SNR values can be found in Table 3). shows consistent effects between various inputs. posthoc -tests confirm significant differences between various inputs. Table 3Pairwise comparison between ICA and PCA outputs.
Note: SNR in dB shown as mean±standard deviation, p values from pairwise t-tests and effect size g and 95% CIs of g37 in brackets. The deterministic automated selection of input and respective according to Appendix A provides higher quality cbPPGs compared to random selection of of the same size or standard and , respectively. Figure 3 confirms this finding graphically. Omitting the random selection, an increased ROI size comes along with a decreased input SNR [for example, on a multispectral input ROI ()]. Regarding the random selection of , the input signal quality in both cases, the monochrome and multispectral random case, is significantly worse than the one achieved by the deterministic selection. 3.2.Blind Source Separation Performance on Different Input SetsFigure 3 shows boxplots of individually taken BSS performances, namely the patient-wise averaged output and input SNRs including the statistical measure of the pairwise difference. Especially for a low input SNR, statistically significant SNR improvements are obtained for both PCA and ICA. PCA moreover significantly improves the SNR of the monochrome set MC1 with equal dominant frequencies whereas ICA significantly improves the SNR of the monochrome set MC2 with different dominant frequencies. ICA also shows a highly significant SNR improvement on MS2. Figure 4 illustrates a distinct dependence of the BSS performance, i.e., the obtained SNR difference , on the input SNR. ANCOVA proves that there are no significant differences in the strength of that dependence (i.e., no differences in the slope of separate regression lines with for PCA and for ICA). ANCOVA further proves highly significant differences in terms of adjusted means (i.e., significant differences in the intercepts of parallel regression lines with for PCA and ICA). ANCOVA posthoc tests confirm significant differences between various outputs (see Fig. 5). ANOVA for the output SNR yields a highly significant difference between outputs ( for PCA and ICA). ANOVA posthoc tests confirm significant differences between various outputs (see also Fig. 5). Figure 5 gives a comprehensive overview on the posthoc results of ANCOVA and ANOVA together with Hedges including its 95% CIs. As stated before, BSS’s application can be considered as superior in a pairwise comparison if and its CI are consistent and show the same direction (sign of ) for both ANOVA and ANCOVA posthoc tests. As can be seen in Fig. 5, not every pairwise comparison shows relevant differences. The comparison of ICA results for ANCOVA and ANOVA for the standard approaches of sets FhC and F may serve as an example for the interpretation of Fig. 5. Both sets show significant differences in their input SNR (-axis) with a negative effect size (input SNR of is smaller than input SNR of ). Considering the comparison of adjusted means, i.e., ANCOVA’s posthoc test, a consistent effect in favor of is found (indicated by the entirely positive CI of ). Concerning the comparison of the output SNR without adjustment, i.e., ANOVA’s posthoc test, also a consistent effect, is obtained. However, this time the CI is entirely negative, which indicates that provides a significantly higher output SNR. Apparently, the lower input SNR of , together with a bounded SNR, favors within ANCOVA’s posthoc test. As example for a pairwise comparison, which shows a relevant difference, readers may be referred to MC1 versus MCR using ICA: besides significantly differing input SNRs (with , input SNR of is higher than input SNR of ), both ANOVA and ANCOVA show consistent effects, either of them is positive. It indicates the better performance of ICA on homogeneous (frequency) inputs of best available SNR compared to homogeneous (frequency) input of random SNR regardless of the input SNR. The results depicted in Fig. 5 can be summarized as follows. Consistent effect size measures, according to the definition (i.e., fully positive, respective negative CI of ), can be found for both PCA and ICA, for example, assessing the adjusted means (ANCOVA) of when compared to high SNR inputs and . Thus, in case of poor input SNRs, PCA and ICA both can be efficiently applied. However, relevant effects, according to the definition, i.e., considering ANCOVA and ANOVA at the same time, are found for PCA with and as well as for ICA with , , and . Accordingly, PCA is considered performing worse on inhomogeneous frequency inputs compared to homogeneous frequencies. Moreover, PCA and ICA are considered performing better on monochrome inputs compared to multispectral inputs of the same ROI size. Increasing the ROI size (from to ) also favors ICA performance. 3.3.Blind Source Separation Performance Given an Input SetTable 3 shows the input and output SNR for all input sets and BSS techniques (PCA and ICA). Note that for equal dominant frequencies, originally three different input sets were available but only the highest SNR is shown (independently from the used dominant frequency; in 44%/36%/20% the first/second/third dominant frequency yielded the highest SNR). Additionally, Fig. 3 shows the according boxplots of input and output SNRs including pairwise -test results in case of significance (). Segment-wise is also shown as a function of the respective input SNR in Fig. 4. Comparing the performance of PCA and ICA on a given input (see Table 3), PCA works significantly better on homogeneous inputs (dominant frequency and wavelength). ICA works significantly better on inhomogeneous inputs (dominant frequency) as well as for inhomogeneous ROIs (). 4.Discussion4.1.Spatial Selection versus Standard ApproachesIn real applications, ROIs for cbPPG extraction have to be selected by using automated video processing algorithms. Typically, rectangular ROIs using frontal face classifiers for detection and (stabilized) tracking of the complete face6 or selective rectangular ROIs of smaller face parts further using facial landmarks10,22 are used. Also, nonrectangular and, thus, more specific ROIs based on facial landmarks27 have been used. To evaluate our spatial ROI selection, we simulate a perfect functioning face/face part detection by manually selecting standard ROIs (see Fig. 1). As can be seen from the results (see Tables 2 and 3), our automated spatial selection among small-sized ( and ) does not require any face detection and outperforms the input SNR of selecting ROIs of the whole face or parts of it (forehead and cheeks) including consistent effect sizes . Together with BSS, the output SNR (Fig. 5) shows inconsistent effects. Thereby, ANCOVA posthoc tests are affected by the highly differing input SNR while ANOVA posthoc tests show similar output SNRs between small-sized and , and . Given comparable output SNRs, an automated estimation of an and would require a higher effort for stable automated landmark detection in different subjects. Our automated spatial selection shows to compete successfully against even manually annotated ROIs achieving higher input SNRs and comparable output SNRs. However, the comparison of deterministic and random selection of clearly motivates that assembling homogeneous input characteristics (wavelength and dominant frequency) for BSS processing is not sufficient to yield the best BSS outcome. Both, the deterministic as well as the random selection of ROIs make use of periodicity in terms of frequently occurring dominant frequencies. However, the random selection does not force the highest possible SNR of a periodic component, like it is conducted in the deterministic selection. Accordingly, this comparison serves as a test, if always the best available signal quality needs to be selected in context of BSS. According to that, input and output SNR of the sets MCR and MSR of both BSS algorithms show absolute values below 0 dB (see Tables 2, 3, and Fig. 3). In addition, the BSS performance as a function of the input SNR (Fig. 4) shows an inferior performance compared to deterministic selection of equally sized ROIs in terms of a broadened area of negative performance (performance below -axis), which widely lasts into the range of negative input SNRs. This finding underlines the necessity to process the best available input. Consequently, BSS is not necessarily able to compensate for lower input SNR under comparable conditions (ROI size and dominant frequency). 4.2.Blind Source Separation PerformanceFigures 6 and 7 show examples of selecting according to the proposed BSS input selection and the further processing of these signals with PCA and ICA for the monochrome (MC1) and multispectral (MS1) case, respectively. In both cases, PCA and ICA are able to extract a distinct pulsatile component in the time domain. Focusing on the markedness of the spectral peak related to the heart rate in the spectra , it is worth noting that in the monochrome set (Fig. 6) PCA and ICA are performing similar. On the contrary, in the multispectral set (Fig. 7), PCA shows a decrease in the spectral power of the cardiac pulse compared to ICA. However, both examples and both BSS algorithms show at least one output component of proper quality regarding common postprocessing tasks (e.g., heart rate estimation). Accordingly, even a decrease in signal quality by application of BSS not necessarily renders postprocessing impossible. Nevertheless, morphology retention through BSS should be considered carefully.40 Several researchers have reported that BSS not necessarily improves the cbPPG quality and outcome. ICA was found to, if any, only subtly decrease the heart rate error for a small-sized cheek ROI8 and even (slightly) increase for rectangular face ROI16,17 compared to the BSS inputs. Moreover, PCA was found to perform worse on multispectral inputs compared to FastICA.16 Since movements affect the input signal quality, it is worth relating these results with the movement conditions during recording. Christinaki et al.8 allowed for small movements (facial expression) while extracting cheek ROIs, whereas face ROIs were extracted from subjects who were asked not to move.16,17 On the other hand, improvements of Bland–Altman heart rate measures were shown for facial ROIs,6 while these improvements turned out to be higher for movement phases compared to no movement. Accordingly, the beneficial usage of BSS for cbPPG is rather found in conditions of less signal quality. Such findings are in accordance with the results obtained by our investigations. Considering Fig. 4, which indicates an inverse relationship between input SNR and BSS performance (negative ), the usage of BSS can even decrease the signal quality; its application, thus, should be considered with care. Particularly for small-sized ROIs after deterministic spatial selection (MC1, MC2, MS1, and MS2), an SNR decrease mainly appears in case of high-quality inputs (to the right of the -axis) and is differently pronounced for different sets. On the contrary, the standard approaches using the face ROI and the forehead-cheeks ROI (Fig. 4) are not showing this marked negative BSS performance, whereas the increase of SNR for good quality inputs is also limited. Accordingly, the performance of the compared methods is limited for high-quality inputs. This behavior might be attributed to the used contrast. Focusing for instance on MC1 that consists of inputs, which are as homogeneous as possible, the decorrelative transformation conducted by PCA is mostly able to preserve the SNR and shows the lowest number of segments with negative for high (positive) input SNRs (see Fig. 4). In comparison, the additional rotation introduced by ICA decreases the SNR for this set. However, the exclusive usage of PCA for high-quality inputs is not sufficient if the input is not as homogeneous as assembled by MC1 as the results of MC2, MS1, and MS2 show. So far, we used a standard tanh-contrast for FastICA as well as symmetric optimization for uniformity of the amount of output signals. One should further test optimized contrasts for example rather abstaining from super-Gaussian source optimization as indicated by the cbPPG component selection of Morris et al.33 or alternatively let the demixing be guided by an expected cardiac pulse composition.15 Furthermore, deflationary ICA could be applied for avoiding a model order violation. Another possibility to avoid undesired SNR decrease, if no SNR preserving contrast for high input quality is available, could be an adaptive decision, whether a BSS algorithm should be applied or not. This decision could be based on the prior SNR estimate based on peak frequency detection, which is done during the selection process of inputs. 4.3.Homogeneity of the Blind Source Separation InputIn this work, different factors of (in-)homogeneity of input signal sets to BSS are assessed in a controlled fashion. First, the sensitivity of BSS algorithms regarding input of different dominant frequencies is analyzed. The results show that the ICA can take advantage of input signals comprising content with different dominant frequencies (MC2) while PCA shows a significantly worse performance (see Table 3). One might infer that the concept of statistical independence as applied for ICA is rather suited to such content compared to the concept of decorrelation utilized for PCA. However, in case of a perfect homogeneous BSS input (MC1) comprising only one uniform wavelength and dominant frequency, PCA performs significantly better than ICA. As we used prewhitening prior to ICA that is similar to PCA, again one might deduce that the contrast applied for ICA, which is used to additionally transform the prewhithened data, is not well chosen to extract the cardiac pulse component. Second, the question of sensitivity of BSS algorithms regarding wavelength homogeneity could clearly be answered for same ROI sizes and equal dominant frequencies (MC1 and MS2) in support of the monochrome approach. Both, PCA and ICA showed a significantly higher output SNR (see Table 3) by using the monochrome input. Also, consistent effects are found for both, ANCOVA and ANCOVA on the outputs (Fig. 5). Such findings support the idea that wavelength-dependent penetration depth into human skin imposes a nonlinear problem,3 which BSS cannot consistently handle properly. However, our results also confirm the better suitability of ICA compared to PCA for multispectral face ROI input.16 Another result giving insight into BSS input homogeneity is stated by the comparison between the multispectral sets MS1, MS2, FhC, and F. While ROI size and input SNR show an inverse relationship, the output SNR of multispectral PCA and ICA shows proportionality to the ROI size except for the rectangular face ROI (see Tables 2 and 3). Up to , one might assume a homogeneous ROI augmentation since mostly homogeneous skin regions without marked edges and regions, which do not necessarily contribute to a distinct cardiac pulse, are consolidated with . The same skin regions are principally addressed by the sets MS1 and MS2. The only exception gives where also less-suited regions, such as mouth and nose, are included in the ROI, thus, serving an inhomogeneous ROI augmentation. Consequently, homogeneous ROI augmentations seem to be beneficial for multispectral BSS, whereas inhomogeneous areas inside ROIs should be omitted to optimize the extraction of the cardiac pulse. This behavior could also be found regarding heart rate error measures after FastICA on 15-s cbPPGs of different ROIs.29 Despite that investigation neglected the input quality, the FastICA output showed decreasing heart rate error measures while step-wise excluding the face surrounding and face borders from the ROI. On the other hand, heart rate errors of outputs increased again while assessing ROIs with highly edged face regions mostly containing nose and mouth structures. Nevertheless, the positive effect of homogeneity of the input seems to be limited especially considering monochrome inputs of high signal quality, which links to the previous discussion on appropriate BSS contrasts. A future application of the result that only PCA is able to preserve the SNR in very homogeneous (wavelength and dominant frequency) high-quality inputs of small-sized ROIs could be the evaluation of of spatially distributed ROIs to address the spatial homogeneity of the cutaneous microcirculation. Two-dimensional statements of the microcirculation may provide clinical significance in critical care patients.30 5.ConclusionIn conclusion, we investigated the performance of BSS to enhance the cardiac pulse from the cbPPG in dependency to varying input data characteristics. To that end, we developed an automated spatial selection of small-sized ROIs to locate the cardiac pulse in video recordings and control the input characteristics. The cbPPGs obtained from this spatial ROI selection significantly increased the SNR of the cardiac pulse compared to standard approaches, which select the whole face or anatomically defined subregions as ROIs. Subsequent BSS application did not show an unambiguous effect, rather input characteristics and particular BSS techniques had to be considered for future BSS usage on cbPPGs. While PCA developed the better performance compared to ICA using a very homogeneous input of same (green) wavelength and dominant frequency, ICA showed a significantly better performance compared to PCA on inhomogeneous inputs. Both, PCA and ICA performed better on monochrome inputs compared to multispectral inputs for the same ROI size. Algorithms, such as the proposed automated spatial ROI selection, can help to ensure passing appropriate inputs to respective BSS routines. Our results indicate, that, regardless of a subsequent BSS application, the usage of signal characteristics, such as simple frequency-domain features, can help identifying beneficial ROI locations to obtain superior cbPPGs compared to using classical ROI definitions based on face features. Furthermore, it turned out that BSS application might suffer from an inverse relationship between input signal quality and BSS performance, which even can cause a decrease of signal quality compared to the input data. Future research should address adaptive cbPPG processing schemes, which invoke BSS optionally or search for BSS contrasts capable of separating cbPPGs from noise independently from the input quality. Notwithstanding this, unsupervised signal processing techniques, such as BSS, should be carefully characterized in context with the measurement techniques and respective signal characteristics to which they are applied. The introduction of in the context of quality assessment and our results allow comparisons of other algorithms and datasets in the future. AppendicesAppendix:Algorithm for Selection of BSS InputAlgorithm InputAll available signals with and serve as input to the input set selection algorithm. The selection is mainly based on evaluating peak frequencies and its grouping clusters considered as the dominant frequency, respectively, periodic component of the amplitude spectrum of a .
Monochrome Approach (Homogenous Frequency Content)Assemble the green channel from appropriate ROIs inside a cluster to three input sets according to ( Monochrome Approach (Heterogeneous Frequency Content)Assemble the green channel from appropriate ROIs inside an SNR positioning (of absolute SNRs) for all three clusters with to three input sets according to ( Multispectral Approach (Homogenous Frequency Content)Assemble the color channels from appropriate ROIs inside a cluster to three input sets according to ( Multispectral Approach (ROI Area Adaption to Monochrome Approach)Assemble the color channels from appropriate ROIs inside a cluster containing the area () of the three best SNR ROIs obtained by frame-wise averaging the respective ROIs located as in for three wavelengths each to three input sets according to ( Random ApproachesFor testing against choosing only the highest SNRs, assemble analogous monochrome and multispectral sets with respective random selection out of available subject to and . AcknowledgmentsThe authors thank the Saxon State Ministry of Science and Culture (SMWK) for funding the project CardioVisio—Contactless acquisition of vital parameters. All authors report grants from Saxon State Ministry of Science and Culture (SMWK) during the conduct of the study. D. Wedekind and A. Trumpp report grants from Steinbeis Innovation Center Applied Medical Technology outside the submitted work. ReferencesY. L. Zheng et al.,
“Unobtrusive sensing and wearable devices for health informatics,”
IEEE Trans. Biomed. Eng., 61
(5), 1538
–1554
(2014). http://dx.doi.org/10.1109/TBME.2014.2309951 IEBEAX 0018-9294 Google Scholar
C. Brueser et al.,
“Ambient and unobtrusive cardiorespiratory monitoring techniques,”
IEEE Rev. Biomed. Eng., 8 30
–43
(2015). http://dx.doi.org/10.1109/RBME.2015.2414661 Google Scholar
M. Huelsbusch and V. Blazek,
“Contactless mapping of rhythmical phenomena in tissue perfusion using PPGI,”
Proc. SPIE, 4683 110
–117
(2002). http://dx.doi.org/10.1117/12.463573 PSISDG 0277-786X Google Scholar
W. Verkruysse, L. Svaasand and J. Stuart Nelson,
“Remote plethysmographic imaging using ambient light,”
Opt. Express, 16
(26), 21434
–21445
(2008). http://dx.doi.org/10.1364/OE.16.021434 OPEXFF 1094-4087 Google Scholar
G. Balakrishnan, F. Durand and J. Guttag,
“Detecting pulse from head motions in video,”
in Proc. of the 26th IEEE Computer Vision and Pattern Recognition Conf. (CVPR),
3430
–3437
(2013). Google Scholar
M. Poh, D. McDuff and R. Picard,
“Non-contact, automated cardiac pulse measurements using video imaging and blind source separation,”
Opt. Express, 18
(10), 10762
–10774
(2010). http://dx.doi.org/10.1364/OE.18.010762 OPEXFF 1094-4087 Google Scholar
H. Wu et al.,
“Eulerian video magnification for revealing subtle changes in the world,”
ACM Trans. Graphics, 31
(4), 1
–8
(2012). http://dx.doi.org/10.1145/2185520 ATGRDF 0730-0301 Google Scholar
E. Christinaki et al.,
“Comparison of blind source separation algorithms for optical heart rate monitoring,”
in Proc. of the 4th Mobihealth,
339
–342
(2014). Google Scholar
Y. Hsu, Y. Lin and W. Hsu,
“Learning-based heart rate detection from remote photoplethysmography features,”
in Proc. of the 39th Int. Conf. on Acoustics, Speech, and Signal Processing (ICASSP),
4433
–4437
(2014). Google Scholar
L. Feng et al.,
“Motion artifacts suppression for remote imaging photoplethysmography,”
in Proc. of the 19th Int. Conf. on Digital Signal Processing (DSP),
18
–23
(2014). Google Scholar
P. Comon,
“Independent component analysis, a new concept?,”
Signal Process., 36
(3), 287
–314
(1994). http://dx.doi.org/10.1016/0165-1684(94)90029-9 Google Scholar
J. Cardoso,
“High-order contrasts for independent component analysis,”
Neural Comput., 11
(1), 157
–192
(1999). http://dx.doi.org/10.1162/089976699300016863 NEUCEB 0899-7667 Google Scholar
A. Hyvärinen,
“Fast and robust fixed-point algorithms for independent component analysis,”
IEEE Trans. Neural. Networks, 10
(3), 626
–634
(1999). http://dx.doi.org/10.1109/72.761722 ITNNEP 1045-9227 Google Scholar
G. de Haan and V. Jeanne,
“Robust pulse-rate from chrominance-based rPPG,”
IEEE Trans. Biomed. Eng., 60
(10), 2878
–2886
(2013). http://dx.doi.org/10.1109/TBME.2013.2266196 IEBEAX 0018-9294 Google Scholar
G. de Haan and A. van Leest,
“Improved motion robustness of remote-PPG by using the blood volume pulse signature,”
Physiol. Meas., 35
(9), 1913
–1926
(2014). http://dx.doi.org/10.1088/0967-3334/35/9/1913 PMEAE3 0967-3334 Google Scholar
B. Holton et al.,
“Signal recovery in imaging photoplethysmography,”
Physiol. Meas., 34
(11), 1499
–1511
(2013). http://dx.doi.org/10.1088/0967-3334/34/11/1499 PMEAE3 0967-3334 Google Scholar
S. Kwon, H. Kim and K. Suk Park,
“Validation of heart rate extraction using video imaging on a built-in camera system of a smartphone,”
in Proc. of the 34th Annual Conf. of the IEEE Engineering in Medicine and Biology Society (EMBC),
2174
–2177
(2012). Google Scholar
M. Poh, D. McDuff and R. Picard,
“Advancements in noncontact, multiparameter physiological measurements using a webcam,”
IEEE Trans. Biomed. Eng., 58
(1), 7
–11
(2011). http://dx.doi.org/10.1109/TBME.2010.2086456 IEBEAX 0018-9294 Google Scholar
M. Lewandowska et al.,
“Measuring pulse rate with a webcam: a non-contact method for evaluating cardiac activity,”
in Proc. of the Federate Conf. on Computer Science and Information Systems (FedCSIS),
405
–410
(2011). Google Scholar
F. Zhao et al.,
“Remote measurements of heart and respiration rates for telemedicine,”
PLoS One, 8
(10), e71384
(2013). http://dx.doi.org/10.1371/journal.pone.0071384 POLNCL 1932-6203 Google Scholar
G. Tsouri et al.,
“Constrained independent component analysis approach to nonobtrusive pulse rate measurements,”
J. Biomed. Opt., 17
(7), 077011
(2012). http://dx.doi.org/10.1117/1.JBO.17.7.077011 JBOPFO 1083-3668 Google Scholar
D. McDuff, S. Gontarek and R. Picard,
“Remote detection of photoplethysmographic systolic and diastolic peaks using a digital camera,”
IEEE Trans. Biomed. Eng., 61
(12), 2948
–2954
(2014). http://dx.doi.org/10.1109/TBME.2014.2340991 IEBEAX 0018-9294 Google Scholar
W. Wang, S. Stuijk and G. de Haan,
“Exploiting spatial-redundancy of image sensor for motion robust rPPG,”
IEEE Trans. Biomed. Eng., 62
(2), 415
–425
(2015). http://dx.doi.org/10.1109/TBME.2014.2356291 IEBEAX 0018-9294 Google Scholar
C. Lueangwattana, T. Kondo and H. Haneishi,
“A comparative study of video signals for non- contact heart rate measurement,”
in Proc. of the 12th Int. Conf. on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology (ECTI-CON),
1
–5
(2015). Google Scholar
L. Yang et al.,
“Motion-compensated non-contact detection of heart rate,”
Opt. Commun., 357 161
–168
(2015). http://dx.doi.org/10.1016/j.optcom.2015.08.017 OPCOB8 0030-4018 Google Scholar
Y. Sun et al.,
“Motion-compensated noncontact imaging photoplethysmography to monitor cardiorespiratory status during exercise,”
J. Biomed. Opt., 16
(7), 077010
(2011). http://dx.doi.org/10.1117/1.3602852 JBOPFO 1083-3668 Google Scholar
H. Tasli, A. Gudi and M. Uyl,
“Remote PPG based vital sign measurement using adaptive facial regions,”
in Proc. of the 16th Int. Conf. on Image Processing (ICIP),
1
–5
(2014). Google Scholar
A. R. Guazzi et al.,
“Non-contact measurement of oxygen saturation with an RGB camera,”
Biomed. Opt. Express., 6
(9), 3320
–3338
(2015). http://dx.doi.org/10.1364/BOE.6.003320 JBOPFO 1083-3668 Google Scholar
K. Mannapperuma et al.,
“Performance limits of ICA-based heart rate identification techniques in imaging photoplethysmography,”
Physiol. Meas., 36
(1), 67
–83
(2015). http://dx.doi.org/10.1088/0967-3334/36/1/67 PMEAE3 0967-3334 Google Scholar
S. Rasche et al.,
“Camera-based photoplethysmography in critical care patients,”
Clin. Hemorheol. Microcirc., 64
(1), 77
–90
(2016). http://dx.doi.org/10.3233/CH-162048 Google Scholar
L. Tarassenko et al.,
“Non-contact video-based vital sign monitoring using ambient light and auto-regressive models,”
Physiol. Meas., 35
(5), 807
–831
(2014). http://dx.doi.org/10.1088/0967-3334/35/5/807 PMEAE3 0967-3334 Google Scholar
“Cardiac monitors, heart rate meters, and alarms,”
American National Standard,
(2002). Google Scholar
D. Wedekind et al.,
“Automated identification of cardiac signals after blind source separation for camera-based photoplethysmography,”
in Proc. of the 35th Int. Conf. on Electronics and Nanotechnology (ELNANO),
422
–427
(2015). Google Scholar
B. Lantz,
“The large sample size fallacy,”
Scand. J. Caring Sci., 27
(2), 487
–492
(2013). http://dx.doi.org/10.1111/scs.2013.27.issue-2 SJSCEN Google Scholar
C. O. Fritz, P. E. Morris and J. J. Richler,
“Effect size estimates: current use, calculations, and interpretation,”
J. Exp. Psychol. Gen., 141
(1), 2
–18
(2012). http://dx.doi.org/10.1037/a0024338 JPGEDD 1939-2222 Google Scholar
H. Hentschke and M. Stüttgen,
“Computation of measures of effect size for neuroscience data sets,”
Eur. J. Neurosci., 34
(12), 1887
–1894
(2011). http://dx.doi.org/10.1111/j.1460-9568.2011.07902.x EJONEI 0953-816X Google Scholar
S. Nakagawa and I. C. Cuthill,
“Effect size, confidence interval and statistical significance: a practical guide for biologists,”
Biol. Rev., 82
(4), 591
–605
(2007). http://dx.doi.org/10.1111/brv.2007.82.issue-4 Google Scholar
E. C. Hedberg and S. Ayers,
“The power of a paired t-test with a covariate,”
Social Sci. Res., 50 277
–291
(2015). http://dx.doi.org/10.1016/j.ssresearch.2014.12.004 SSREBG Google Scholar
F. Andreotti et al.,
“An open-source framework for stress-testing non-invasive foetal ECG extraction algorithms,”
Physiol. Meas., 37
(5), 627
–648
(2016). http://dx.doi.org/10.1088/0967-3334/37/5/627 PMEAE3 0967-3334 Google Scholar
BiographyDaniel Wedekind is a PhD student at the Institute of Biomedical Engineering, TU Dresden. He received his diploma degree in electrical engineering from the TU Dresden in 2013. His current research interests include practically relevant questions of the application of blind source separation for contactless vital signs monitoring. Alexander Trumpp is a PhD student at the Institute of Biomedical Engineering, TU Dresden. He received his diploma degree in mechatronics engineering from the TU Dresden in 2013. His current studies focus on camera-based vital sign monitoring, both from a research and application point of view. Sebastian Zaunseder received his PhD in Electrical Engineering from TU Dresden, Germany in 2011. Subsequently, he joined the Institute of Biomedical Engineering of TU Dresden, where he is currently head of the group Biosignals. His research interests include contact-free measurement systems, processing of biomedical signals and images to acquire robust information on vital signs, investigations on the cardio-respiratory autonomic modulation and research related to sleep. |