Automated and precise heartbeat detection in ballistocardiography signals using bidirectional LSTM

Abstract

Ballistocardiography (BCG) provides an unobtrusive, mechanical measurement of heartbeat for early detection and longitudinal monitoring of cardiovascular conditions, although it is plagued by poor signal to noise ratio. Despite tremendous efforts over the last decade in related advanced signal processing and deep learning research, automated and precise heartbeat detection in BCG remains a major technical challenge. In this work, we design and study a sequential deep neural network based approach that addresses three key issues in automated BCG learning and detection: auto-labeling of BCG samples in the raw training data; design of a recursive neural network that can learn the association between the continuous BCG waves and the label sequence in the form of sample-by-sample predictions; and finally, heartbeat detection from the output sequence. We evaluate the proposed method using our data set comprising 8 human subjects under varying cardiac output conditions: pre-exercise and post-exercise. Using multiple performance metrics, we report that the proposed method can achieve a heart rate accuracy of 98.68% and root-mean-square-error of 1.37 bpm with 98.85% coverage, which compare favorably to two state-of-the-art methods in the same test.

Keywords

• Ballistocardiography
• Recursive deep learning
• Heartbeat detection