Heartbeat detection with dictionary learning from ballistocardiography

Abstract

Non-contact sensor technology, particularly accelerometers and gyroscopes placed on beds during sleep studies, has emerged as a potential method for unobtrusive health monitoring. However, accurately detecting heartbeats from these sensors can be challenging due to factors like motion artefacts and individual variations. This study addresses these challenges by evaluating various dictionary-learning algorithms for heartbeat detection in non-contact sensor data. Dictionary learning algorithms aim to represent the data using a set of basis vectors learned from the data itself. This allows them to efficiently capture the underlying structure of the data, potentially separating heartbeat signals from background noise and motion artefacts. The analysis employed both intra-subject and inter-subject datasets, along with balanced and unbalanced data distributions, reflecting real-world scenarios. By comparing algorithm performance across these diverse conditions, the research sought to identify the most robust and effective dictionary learning methods for accurate heartbeat detection in non-contact sensor applications.

Three different algorithms – DL-FUMI, LC-KSVD1, and LC-KSVD2 – were evaluated alongside a technique for combining information from multiple sensor axes: Principal Component Analysis (PCA) fusion. While PCA fusion aimed to leverage comprehensive data from all axes, it did not outperform results obtained from individual axes. LC-KSVD1 and LC-KSVD2 consistently achieved the highest recall, F1-score, and Matthews Correlation Coefficient (MCC) across most categories. When using a single sensor with all channels, LC-KSVD1 achieved the highest recall (0.782) and MCC (0.214) on the intra-subject balanced dataset using the accelerometer y-axis. On the inter-subject balanced dataset with the accelerometer y-axis, DL-Fumi achieved the highest recall (0.782) and LC-KSVD1 on MCC (0.216). On specificity, DL-Fumi achieved clearly the best performance in intra- (0.628) and in inter-subject (0.638). The evaluation revealed that the algorithms performed at similar performance levels with other algorithms across various metrics, demonstrating their potential for accurate heartbeat detection in non-contact sensor applications.