The Optimal Morphological Model for Arterial Blood Pressure Wave Related Classification: Comparison of Two Types of Kernel Function Mixtures

• Published in: IEEE access Vol. 8; pp. 4133 - 4148
• Main Authors: Chou, Yongxin, Wang, Ping, Feng, Yufeng
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Arterial blood pressure; Arterial blood pressure signal; Blood pressure; Bradycardia; Classification; Classifiers; Computational modeling; Data models; Elastic waves; Gaussian function; Indexes; Kernel; Kernel functions; Kolmogorov-Smirnov test; lognormal function; Mathematical models; Modelling; Morphology; Neural networks; optimal morphological model; Parameters; Performance indices; Redundancy; Solid modeling; Tachycardia
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2019.2958304

The morphological modeling methods are efficient in quantifying the change of arterial blood pressure (ABP) waves. The related works focus on minimizing the modeling error but ignore the classification related modeling expression in practical applications. In this study, we explored the optimal modeling method for ABP wave related classifications. Two types of conventional models, Gaussian or Lognormal kernel function mixtures, were employed to quantitively describe the change of ABP signals, and the parameters of different models were engaged to train the different classifiers by probabilistic neural network (PNN) and random forest (RF) for identifying the ABP waves by age, gender, and whether belonging to extreme bradycardia (EB) or extreme tachycardia (ET). Then, we defined some indexes about the performance of modeling and classifications as the references to compare the different models. The ABP signals of Fantasia and 2015 PhysioNet/CinC Challenge databases were exploited as the experimental data to select the optimal model. The modeling results show that the Lognormal kernel function mixtures have a lower error in ABP wave modeling. The two-sample Kolmogorov-Smirnov test (ks-test) results indicate that the parameters of all models are markedly different at a highly significant level (h = 1, p <; 0.05) between different groups. The classification results show that the classifiers based on the four-Gaussian function model have the best performance with the average Kappa coefficients (KC) of 99.160 ± 0.123%, while the average KC for the classifiers of two-Lognormal function models is 97.585 ± 0.172%, which means there is excessive information redundancy in the classifications by the three and four kernel functions models. Considering some other indexes such as time consumption and RAM space, the 2 Lognormal function model has more potential in practical applications.