Signal-Level Fusion With Convolutional Neural Networks for Capacitively Coupled ECG in the Car

Abstract

Unobtrusive measurement technologies for vital signs, such as capacitively coupled electrocardiography (cECG), allow for health monitoring outside the clinical domain, for example in automotive applications. While cECG has the potential to deliver accurate information about the driver's heart, the signal quality is very volatile compared to standard ECG and can be corrupted easily by motion artifacts. In this work, we present a signal-level fusion algorithm based on a convolutional neural network (CNN) to locate individual heartbeats in three-channel cECG signals. To design and optimize the network's structure, data from the PhysioNet / CinC challenge "Robust Detection of Heart Beats in Multimodal Data" was used as an independent source. To train and test the algorithm, we used cECG data from six subjects and three different driving scenarios (highway, city, and proving ground) that is freely available as part of our UnoViS-database. Data consisted of 31 recordings with a total duration of 13.4 hours. Leaveone-subject-out cross validation was performed to assess the algorithm's performance. We achieved a sensitivity of 88.0% and a positive predictive value of 95.2% compared to the reference ECG, with a root-mean-square R-Peak localization error of 20.81 ms. The developed algorithm is available for download via the UnoViS-Website.

Physiological Driver Monitoring Using Capacitively Coupled and Radar Sensors

“…The motion artefact algorithms improve the median R peak detection sensitivity from 92.74% to 98.28% and the median positive predictive value (PPV) from 92.34% to 99.03%, city and highway combined. This compares favorably to [30] (average sensitivity 91.75%, PPV 84.47%) and [31] (average sensitivity 88.02%, PPV 95.17%). The overall mean absolute error on the heart rate when compared to the reference system is 1.48 bpm, calculated over 60 s windows.…”

Evaluation of Capacitive ECG for Unobtrusive Atrial Fibrillation Monitoring

“…Recent work has attempted to overcome the volatile nature of the cECG modality through methods such as denoising [11] and signal fusion [12]. Current cECG-based comparisons are also limited by the dearth of publicly available unobtrusive signal databases as compared with large clinical ECG databases such as Physionet [13] and MIT-BIH [14].…”

Multiple Channel Electrocardiogram QRS Detection by Temporal Pattern Search

“…Frequently encountered noise sources confound electrocardiogram heart beat detection methods that rely solely on size or shape criteria to distinguish QRS complexes from noise peaks.8 A single channel convolutional neural network (CNN) described by Cai and Hu has shown some promise in overcoming the limitations of conventional peak size/shape based QRS detectors by implicitly exploiting temporal patterns in single channel normal sinus rhythm cardiac signals. 3 However, CNNs often don't generalize well, and the Cai and Hu detector has relatively poor temporal resolution, with each output node corresponding to 16 ms. 3 Antink 1 and Ravichandran 10 independently extended neural networks to QRS detection in 3 channel capacitive electrode recordings from the UnoViS automobile database 14 (described further below). Antink's CNN architecture achieved a sensitivity of 88.0% with a false discovery rate of 4.8% based on a 150ms peak matching window.…”