A method for constructing a feature space based on the phase portrait of a biosignal in analytical software subsystems

This is the authors' abstract. We don't add key points for this paper.

The development of software for biomedical monitoring platforms built on internet of medical things architectures requires computational criteria for biosignal analysis that are robust to the heterogeneity of sensor hardware and to random data losses in wireless communication channels, because the conventional geometric characteristics of the phase portrait are highly sensitive to these factors, which complicates the aggregation of data from different devices in distributed systems. The aim of the study is to present a method for constructing the feature space of the predictive software components of a biomedical monitoring platform on the basis of geometric invariants of the biosignal phase portrait that are robust to variations in the sensor sampling rate and to data packet loss. The phase space was reconstructed from the discrete signal using the differential method and the time-delay method based on Takens' embedding theorem, while feature extraction employed three newly developed computational criteria: the dynamic phase asymmetry index, the normalised fractal density of the trajectory and the phase coefficient of shape variation. The properties of the criteria were verified on a synthetic sample of two hundred pulse wave segments at three sampling rates and three levels of artificial packet loss, while their contribution to the classification quality of critical physiological states was assessed on a multisource clinical sample of one hundred and thirty patients by means of recurrent neural networks, support vector machines and random forests, with statistical significance tested by the Wilcoxon signed-rank test. The dynamic phase asymmetry index proved to be asymptotically invariant to the sampling rate: the coefficient of variation of its values was 2.89% against 63.28% for the integral attractor area. The normalised fractal density of the trajectory significantly outperformed the box-counting method in robustness to packet loss from ten per cent of lost samples onwards. The superiority of the combined phase feature vector over time-domain heart rate variability indices was not confirmed on the synthetic sample, which indicates the need for clinical validation of the diagnostic value of the proposed criteria. The developed criteria have a computational complexity suitable for stream processing on the edge nodes of distributed monitoring platforms and ensure the robustness of the feature vector to sensor hardware heterogeneity, which makes it possible to aggregate data from different devices without a systematic bias in the predictions of predictive models