Appraisal of the Signal Quality of a Contactless ECG-System

Type: Master Thesis
Student: Sebastian Axmann
Advisor: Johannes Schumm, Bert Arnrich
Project: SEAT

As mean age of people has increased during last decades, the demand for health related information that are unobtrusively measured has increased as well. As cardiac incidents represent a large fraction of death causes, continuously monitored electrocardiogram data are of special interest. Contact-less capacitive electrodes, integrated into a seat for example, can be used throughout daily life for an assessment of such information. Signals gained by this method contain movement artifacts and interferences that have to be addressed before further analysis.
A binary quality index whose value indicates signal usability for analysis is presented as a solution to this challenge. Signal features calculated from electrocardiogram signal and additional force sensors are combined as inputs to logistic regression. Output of this algorithm is a set of coefficients for logistic regression formula which models the relation between quality index and signal properties.
Hardware related work contained selection and implementation of pressure sensors as well as their integration into a microprocessor based signal chain. As this work is embedded into the SEAT project (Smart Technologies for Stress Free Air Travel), hardware components were integrated into a real airplane set. An experiment with twelve subjects for data acquisition was conducted. Each subject performed typical passenger activities during one hour, resulting in a database of almost 12 hours of records. Overall system was found to work stable and reliable.
A Pan-Tompkins based QRS detector was applied to the ECG data and a ground truth based quality index was calculated for each person and activity with help of an independent contact based system. Models found for quality index using logistic regression were used for analysis of equal and distinct sets of test andtraining data.
Overall data recorded exhibited in 55.32 % percent of all cases a high quality index. The same number would result if quality index is guessed without using signal features. When taking advantage of signal properties and models derived, quality index of electrocardiogram signal could be modeled using unseen data with an overall accuracy (prediction of good and bad signal parts) of 75.51 % for unsupervised R-peak detection and 83.34 % for hand corrected peaks.