Lungprints: An alternative view in multiple-stream time-series analysis of bioimpedance signals

In this paper we introduce Lungprints, which are 2D barcode-pattern like images representing ventilation characteristics of a patient. Recent advances in impedance imaging, namely electrical impedance tomography (EIT), allow for the development of an inexpensive, non-ionizing and non-invasive way of monitoring lung ventilation. Various methods have been proposed for the characterization of temporal lung tissue behaviour during ventilation using EIT. Impedance images are built in real time (13 frames/sec) by processing the measured voltage-streams on the thoracic surface. However, impedance recovery is a highly non-linear ill-posed inverse problem, thus the generated images are highly fuzzy and blurry. The latter characteristics impede robust and precise feature extraction and quantification, which could be used as suitable features for lung signal analysis and diagnosis. In this light, we propose an automatic pipeline for lung bioimpedance signal representation and analysis using rapid lungprint classification. We train a single-class vector machine (SVM) classifier, which estimates the parameters of the class from a dictionary of features, composed of moments extracted from a multi-scale representation of healthy patient lungprints onto a translation invariant set of atoms, namely stationary wavelets. Results show the appropriateness of the proposed method in automatic lung signal representation and classification and its potential advantages not only as a bed-side lung monitoring tool, but in other areas such as multiple stream time-series analysis where data streams exhibit cyclic behavior.