Dynamic Heartbeat Modeling with Recurrent Neural Networks and Inverse Gaussian Point Process

TL;DR

This paper explores using recurrent neural networks to improve inverse Gaussian process-based heartbeat modeling, enhancing probabilistic analysis of heart rate variability.

Contribution

It demonstrates that RNN architectures can effectively identify parameters in IGP models, advancing probabilistic HRV modeling.

Findings

RNNs can accurately estimate IGP parameters for heartbeat data.

Combining neural networks with IGP enhances probabilistic modeling of R-R intervals.

The approach allows for future integration of complex physiological mechanisms.

Abstract

Heart rate variability (HRV) analysis is important for the assessment of autonomic cardiovascular regulation. The inverse Gaussian process (IGP) has been widely used for beat-to-beat HRV modeling, as it gives a physiological relevant interpretation of heart depolarization process. A key challenge in IGP-based heartbeat modeling is the accurate estimation of time-varying parameters. In this study, we investigated whether recurrent neural networks (RNNs) can be used for IGP parameter identification and thereby enhance probabilistic modeling of R-R dynamics. Specifically, four representative RNN architectures, namely, GRU, LSTM, Structured State Space sequence model (S4), and Mamba, were evaluated using the Kolmogorov-Smirnov statistics. The results demonstrate the possibility of combining neural sequence models with the IGP framework for beat-wise R-R series modeling. This approach provides a flexible basis for probabilistic HRV modeling and for future incorporation of more complex physiological mechanisms and dynamic conditions.