Heart rate estimation from ballistocardiographic signals using deep learning

Abstract

Objective. Ballistocardiography (BCG) is an unobtrusive approach for cost-effective and patient-friendly health monitoring. In this work, deep learning methods are used for heart rate estimation from BCG signals and are compared against five digital signal processing methods found in literature. Approach. The models are evaluated on a dataset featuring BCG recordings from 42 patients, acquired with a pneumatic system. Several different deep learning architectures, including convolutional, recurrent and a combination of both are investigated. Besides model performance, we are also concerned about model size and specifically investigate less complex and smaller networks. Main results. Deep learning models outperform traditional methods by a large margin. Across 14 patients in a held-out testing set, an architecture with stacked convolutional and recurrent layers achieves an average mean absolute error (MAE) of 2.07 beat min−1, whereas the best-performing traditional method reaches 4.24 beat min−1. Besides smaller errors, deep learning models show more consistent performance across different patients, indicating the ability to better deal with inter-patient variability, a prevalent issue in BCG analysis. In addition, we develop a smaller version of the best-performing architecture, that only features 8283 parameters, yet still achieves an average MAE of 2.32 beat min−1 on the testing set. Significance. This is the first study that applies and compares different deep learning architectures to heart rate estimation from bed-based BCG signals. Compared to signal processing algorithms, deep learning models show dramatically smaller errors and more consistent results across different individuals. The results show that using smaller models instead of excessively large ones can lead to sufficient performance for specific biosignal processing applications. Additionally, we investigate the use of fully convolutional networks for 1D signal processing, which is rarely applied in literature.

Heart pulse demodulation from Emfit mattress sensor using spectral and source separation techniques.

“…The accuracy of the results is similar to some studies that report errors between 3.8-5 bpm [13]. Brüser reports much better results [1].…”

“…Most methods implementation use segmentation methods to discard epochs, reducing the coverage, in search of noise free epochs. The more holistic approach by Pröll [13] reports errors in 100% coverage. Nevertheless, when the SNR is low, usually due to body movements of any kind, the heartbeat is unlikely to be obtained, making näive or statistical methods -such as reporting previous HR the same or interpolating between clean epochs-suitable.…”

“…The 1-minute epochs are divided into 10-second windows with a 50% overlap. This window is chosen for simplicity: several windows have been used in literature, from 8s to one minute [1], [11], [13]. The spectrum of each of the windows is estimated using Lomb-Scargle using an oversampling factor of 4.…”

“…Emfit mattress has been evaluated to estimate heart rate (HR) and heart rate variability (HRV) [1], [11]. Also, recent research has benchmarked some of these algorithms [12], [13].…”

Heart rate estimation from ballistocardiogram signals processing via low-cost telemedicine architectures: a comparative performance evaluation

“…(15)(16)(17),]. In line with this, Pröll proposed two approaches for HR estimation, one algorithm-based (18) and one deploying deep learning networks (19): in both cases, they proved their approach outperforms those coming from the previous contributions. Therefore, recognizing Pröll's work as the field's state of the art, the authors tried to reinterpret the two mentioned approaches for a TM-related, non-controlled environment.…”