BCG Signal Quality Assessment Based on Time-Series Imaging Methods

Abstract

This paper describes a signal quality classification method for arm ballistocardiogram (BCG), which has the potential for non-invasive and continuous blood pressure measurement. An advantage of the BCG signal for wearable devices is that it can easily be measured using accelerometers. However, the BCG signal is also susceptible to noise caused by motion artifacts. This distortion leads to errors in blood pressure estimation, thereby lowering the performance of blood pressure measurement based on BCG. In this study, to prevent such performance degradation, a binary classification model was created to distinguish between high-quality versus low-quality BCG signals. To estimate the most accurate model, four time-series imaging methods (recurrence plot, the Gramain angular summation field, the Gramain angular difference field, and the Markov transition field) were studied to convert the temporal BCG signal associated with each heartbeat into a 448 × 448 pixel image, and the image was classified using CNN models such as ResNet, SqueezeNet, DenseNet, and LeNet. A total of 9626 BCG beats were used for training, validation, and testing. The experimental results showed that the ResNet and SqueezeNet models with the Gramain angular difference field method achieved a binary classification accuracy of up to 87.5%.

Keywords: ballistocardiogram; classification; convolutional neural network; signal quality assessment; time-series imaging.

Figure 1

BCG dataset. (a) Bio-signals used in the study. (b) Interventions performed to perturb blood pressure. The green arrow indicates the procedure of the interventions.

Figure 2

ECG, PPG, and BCG after pre-processing and segmentation procedures.

Figure 3

Examples of high- and low-quality BCG beats.

Figure 4

Examples of imaging a BCG time-series beat by recurrence plot (RP), Gramian angular summation field (GASF), Gramian angular difference field (GADF), and Markov transition field (MTF).

Figure 5

Confusion matrices of SqueezeNet with GADF (87.5% in accuracy), DenseNet with GADF (87.3%), LeNet_Tanh with MTF (75.6%), and the baseline model, FCN, (78.1%). Each experiment had 2888 samples for the test set, and the experiment was repeated five times by five-fold CV. In total, 14,440 samples were used for the confusion matrices. The number in the confusion matrices is the number of samples.

Figure 6

Plots of training/validation accuracy, [0, 1], and loss, [0, ~), of SqueezeNet with GADF (87.5% in accuracy), DenseNet with GADF (87.3%), LeNet_Tanh with MTF (75.6%), and the baseline model, FCN, (78.1%).