Temporal Modeling of Instantaneous Interbeat Interval Based on Physical Activity

Abstract

Heartbeat serves as a vital sign of health, aiding the diagnosis of various health issues. The autonomic nervous system (ANS) is responsible for regulating heartbeat, and physical activity (PA) can influence the heart rhythm. Changes in heart rate can signal alterations in PA. However, poor measurements or extreme conditions can result in the loss of heartbeat data, which can negatively impact the analysis of heartbeats. To accurately monitor the Heart Rate Variability (HRV) and cardiovascular health, a proper model to compensate for lost data is necessary. This study investigated the effect of different PAs on InterBeat Interval (IBI) prediction and the possibility of using models trained on unrelated activities to predict the next IBI. The IBI series is divided into piecewise stationary sections based on PA, that is, running, walking, and sitting, as verified by a statistical test. Various machine and deep learning methods have been used to model the IBIs related to specific activities. The models were then used to predict the next IBI for the testing sets of related and unrelated activities, and the error changes were compared for each permutation of training and testing. The models were tested using a Physionet archived dataset. The findings suggest that linear models offer the least prediction error, whereas PA-relevant training could minimize errors in most scenarios. However, in cases where specific PA data are not available, the proposed CNN model demonstrated superior generalization capabilities. These findings can improve HRV error correction techniques and enhance cardiovascular health and ANS monitoring.

Keywords

• Machine learning
• piecewise stationarity
• time series prediction
• heart rate
• interbeat interval
• physical activity