Evaluation of Decomposition Analysis on Multi-Models for Digital Volume Pulse Signal

Huang, Shih-Hui; Jan, Hao-Yu; Lin, Wen-Chen; Tsia, Cheng-Lun; Lin, Kang‐Ping

doi:10.1007/978-3-319-19387-8_420

Cited by 3 publications

(8 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Tigges et al (2017a) conducted a comparison of accuracy between several PWD algorithms and found an algorithm based on three Gamma kernels to achieve the best results. Huang et al (2015) examined the benefit of the usage of a mixture of Gamma and Gaussian kernels over pure Gaussian kernel-based algorithms and showed that the algorithm consisting of a Gamma kernel and three Gaussian kernels exhibited the best results in residual analysis. Liu et al (2013) showed that models with three Gaussian kernels achieve a better modeling accuracy than models with fewer kernels.…”

Section: Background On Morphological Ppg Signal Analysis By Pwdmentioning

confidence: 99%

“…We also implemented a combination of a Gamma kernel and a number of Gaussian kernels as proposed by Huang et al (2015), who showed that this combination leads to a better fit than a pure Gaussian model. The kernels are defined as in equation ( 2) and equation ( 4).…”

Section: Gamma Gaussianmentioning

confidence: 99%

“…As table 1 shows, PWD algorithms have been compared before. In contrast to our work, these model comparisons used continuous blood pressure (Liu et al 2013) or PPG recordings (Huang et al 2015, Tigges et al 2017a to evaluate the goodness of fit of the tested algorithms. As these works aimed to compare the goodness of fit of various algorithms to real signals, while our work additionally included the robustness of algorithms against noise, the results are not directly comparable.…”

Section: Model Comparisonmentioning

confidence: 99%

“…It can thus be assumed that algorithms with more than two kernels adapt to the noise. Huang et al (2015) concluded from their comparison that the use of the GammaGaussian kernel combination resulted in a lower mean absolute error than a pure Gaussian model. This corresponds with our observations and can be explained as above.…”

Section: Model Comparisonmentioning

confidence: 99%

See 3 more Smart Citations

Pulse decomposition analysis in photoplethysmography imaging

et al. 2020

View full text Add to dashboard Cite

Objective: Photoplethysmography imaging (PPGI) has gained immense attention over the last few years but only a few works have addressed morphological analysis so far. Pulse wave decomposition (PWD), i.e. the decomposition of a pulse wave by a varying number of kernels, allows for such analyses. This work investigates the applicability of PWD algorithms in the context of PPGI. Approach: We used simulated and experimental data to compare various PWD algorithms from the literature regarding their robustness against noise and motion artifacts while preserving morphological information as well as regarding their ability to reveal physiological changes by PPGI. Main results: Our experiments prove that algorithms that combine Gamma and Gaussian distributions outperform other choices. Further, algorithms with two kernels exhibit the highest robustness against noise and motion artifacts (improvement in N R M S E ‾ of 14.09 %) while preserving the morphology similarly to algorithms using more kernels. Lastly, we showed that PWD can reveal physiological changes upon distal stimuli by PPGI. Significance: This work proves the feasibility of pulse decomposition analysis in PPGI, particularly for algorithms with a low number of kernels, and opens up novel applications for PPGI. Not only for PPGI but for future research on PWD in general, our findings have importance as they elucidate differences between PWD algorithms and emphasize the importance of using initial values. To support such future research, we have released the algorithms and simulated data to the public.

show abstract

Section: Background On Morphological Ppg Signal Analysis By Pwdmentioning

confidence: 99%

Section: Gamma Gaussianmentioning

confidence: 99%

Section: Model Comparisonmentioning

confidence: 99%

Section: Model Comparisonmentioning

confidence: 99%

See 2 more Smart Citations

Pulse decomposition analysis in photoplethysmography imaging

et al. 2020

View full text Add to dashboard Cite

show abstract

“…The model can be synthesized by several single-kernel functions or a mixture of several kernel functions. One to six Gaussian, Lognormal, Rayleigh or Gamma functions in [32], seven Gaussian functions in [35], one Lognormal and two Gaussian functions, one Gamma and two Gaussian functions in [36], one Gamma and four Gaussian in [37], and two Gaussian and three cosine functions in [38] and [39]. There is controversy over the model of the pulsatile wave.…”

Section: Introductionmentioning

confidence: 99%

Bradycardia and Tachycardia Detection Using a Synthesis-by-Analysis Modeling Approach of Pulsatile Signal

Chou

Liu

et al. 2019

IEEE Access

View full text Add to dashboard Cite

Bradycardia and tachycardia reflect abnormalities of the heart that can lead to severe harm to the cardiovascular system. The pulsatile signal is a useful tool to detect these two kinds of arrhythmias. In this study, we present a pulsatile synthesis-by-analysis (PSA) modeling based method to detect bradycardia and tachycardia. A new PSA modeling method was proposed to quantitively describe the changes of pulsatile waves, and we obtained twelve parameters for constructing a feature vector from the PSA model of each wave, by which we trained classifiers of probabilistic neural network (PNN) and random forest (RF). Our experiments were performed on the Fantasia and 2015 PhysioNet/CinC Challenge databases. Some pathological and physiological changes were extracted from the average models of the subjects in different groups. The two-sample ks-test results show that all the parameters between different groups are all markedly different (h = 1, p < 0.05). The classification results show that the performances of RF classifiers are better than that of PNN. The kappa coefficients (KC) of RF classifiers are all over 97%, and that of the classifying among bradycardia, tachycardia, and healthy subjects is 98.652 ± 0.217%. Compared with the performance of some former methods, the obtained results demonstrate that the presented method promotes the classification performance remarkably and has the potential to diagnose bradycardia and tachycardia in m-health. INDEX TERMS Synthesis-by-analysis modeling, pulsatile signal, bradycardia and tachycardia detection, random forest.

show abstract