Link to paper

The full paper is available here.

You can also find the paper on PapersWithCode here.

Abstract

• Recent research is exploring how cognitive processes are affected by the brain and body.
• Physiological features such as breathing rhythms, heart rate, and skin conductance are being analyzed.
• Heart rate dynamics are of particular interest as they provide insight into the autonomic nervous system.
• Extracting useful information from heartbeats is challenging due to noisy estimates.
• A new modelling framework views heart rate as a hidden stochastic process.

Paper Content

Introduction

• Autonomic nervous system (ANS) is responsible for physiological adjustments and responding to environmental demands
• ANS is composed of afferent and efferent pathways involving multiple neural structures
• ANS is involved in cognitive processes such as attention, memory, decision-making, and emotional processing
• Measurement of autonomic outflow can provide insights into physiological substrates that support human cognition
• Popular approach to investigate autonomic outflow is by analysing properties of how heart rate changes over time
• Heart rate variability (HRV) is thought to reflect changes in sympathetic and parasympathetic branches of ANS
• Conventional metrics of HRV are based on spectral decomposition or other linear methods, unable to capture non-linear fluctuations
• Growing literature provides insights on HRV patterns by investigating fractality and signal complexity
• Need for more powerful and flexible estimation methods to better characterise richness of heart rate dynamics
• Proposed framework to model and estimate heart rate dynamics based on state-space modelling and Bayesian statistics

Results

A bayesian approach to model heart rate dynamics

• Heart rate is estimated by counting the number of beats over a period of time.
• A point estimate is used to calculate the average number of beats.
• An alternative approach uses a generative model to estimate the heart rate dynamics.
• The model involves two time series, one counting the number of heart beats and one representing the heart rate.
• The model uses a probability distribution to describe the likelihood of observing a given sequence of heart beats.
• The model estimates the most likely heart rate trajectory and also estimates uncertainty.

Method validation

• Tested proposed method by analyzing autonomic outflow of healthy subjects
• Used public dataset of 10 healthy subjects
• Calculated inter-beat intervals via template-based matching of QRS waves
• Generated smoother and more plausible trajectories than frequentist approach
• Used computational modelling to assess consistency of estimated autonomic outflow
• Synthetic heart rate matches measured heart rate in supine and standing positions
• Values of mean heart rate significantly correlated among subjects

Comparable discriminative power of spectral properties

• Generative modelling can be used to build Bayesian estimators of properties of heart rate dynamics
• Bayesian estimators weight the value of a property for each possible trajectory by the likelihood of the trajectory given the observed data
• For linear properties, the equation reduces to the average trajectory under the posterior
• Spectral properties of heart rate dynamics were studied using Bayesian estimators
• Two different values for hyperparameter θ were used: 0.01 and 1
• Mean heart rate effects were similar for all estimation approaches
• Model with θ = 0.01 had better performance distinguishing between different states than the model with θ = 1
• Discriminative power of the model with θ = 0.01 was comparable to the frequentist method

Better performance for measures of dynamical complexity

• Analytic techniques inspired by complexity science can describe biological processes
• If a dynamical property of heart rate dynamics is non-linear, its expected value is different from the property evaluated on the average trajectory
• Bayesian estimators can assess this difference
• Hurst exponent and permutation entropy are two metrics of dynamical complexity
• Bayesian model with θ = 1 reveals strong effects in both metrics

Discussion

• Introduces a new conceptual and modelling paradigm for autonomic processes
• Uses state-space modelling and Bayesian statistics
• Sampled trajectories can be used to estimate posterior distribution of heart rate dynamics
• Bayesian estimators of dynamical complexity have more discrimative power than frequentist approach
• Future work could model sympathetic and parasympathetic activities separately and model interactions between multiple physiological signals

Materials and methods

Overall modelling approach

• Model of heart rate dynamics based on two time series
• Time series correspond to values of dynamical processes
• Assume heart beats are observed, heart rate follows hidden stochastic process
• Joint probability distribution of heart rate and heart beats expressed as Markovian
• Model requires three ingredients: heart rate dynamics, link between heart rate and heart beats, and distribution of initial condition
• Dynamics of hidden process modeled by Gamma Markov Chain
• Link between heart rate and heart beats modeled by Poisson distribution
• Model parameters are α1, β1, and γ
• Maximum likelihood estimation used to choose α1 and β1
• Prior distribution used to choose γ
• Posterior distributions have simple expressions
• Samples of trajectories computed using Gibbs sampler
• Hyperparameters θ and τ used to control smoothness and connectivity strength
• Sampling frequency f s = 1 Hz

Building bayesian estimators of properties of heart

• Leverage trajectory sampling technology to build Bayesian estimators of properties of heart rate dynamics
• Estimate various properties of heart rate dynamics as the empirical mean value of the posterior
• Suggest that the proposed model for heart rate dynamics is generally non-ergodic
• Run Gibbs sampler for Nr = 2 x 10^4 and discard the first Nd

Ecg data and preprocessing

• Analyzed estimation of heart rate on data set of postural changes
• Data set comprised one-lead ECG series from 10 healthy subjects
• Subjects underwent tilt-table test
• Six sessions sorted randomly between resting periods
• ECG preprocessing involved frequency filtering, R-peak detection, and correction of misdetections
• Used IPFM model to generate synthetic series of heartbeat dynamics
• IPFM model considers combination of heart rate components of sympathetic-vagal control as inputs to an integrator
• GMC model used for Bayesian inference

Code availability

• Code to run Bayesian estimation of heart rate dynamics can be found on github
• Observable data (heart beats) are driven by hidden stochastic process (heart rate)
• Figure 2 compares frequentist and Bayesian estimation of heart rate
• Figure 3 validates the method using synthetic heart rate series
• Figure 4 shows effects of physiological states on properties of heart rate dynamics
• Figure 5 is a graphical model representation of the proposed state-space model for heart rate dynamics