A Bayesian Approach to Modeling Variance of Intensive Longitudinal Biomarker Data as a Predictor of Health Outcomes

Mingyan Yu, Zhenke Wu, Margaret T. Hicken, Michael R. Elliott

TL;DR

This paper introduces a Bayesian hierarchical joint model that treats the variability of high-frequency biomarker data—rather than just the average—as a primary predictor for health outcomes. By decomposing heart rate data into "long-term" (minute-to-minute) and "short-term" (second-to-second) components, the researchers demonstrate that the way a biological signal fluctuates can be more informative than its mean value.

What Problem Does This Paper Address?

In the era of wearables and Ecological Momentary Assessment (EMA), researchers are drowning in "intensive longitudinal data"—think hertz-level heart rate readings or minute-by-minute glucose levels. Traditionally, statisticians treat the fluctuations in this data as "nuisance parameters" or noise to be smoothed away to find the average trajectory (the mean).

However, in many health contexts, the "noise" is actually the signal. For example, Heart Rate Variability (HRV) is a known indicator of stress and autonomic health. The challenge is that standard methods for analyzing this variability often rely on simple averages (like the standard deviation of intervals) and fail to account for the uncertainty inherent in measuring these fluctuations. Furthermore, most existing models use a "two-stage" approach: first estimate the variance, then plug it into a second model to predict an outcome. This often leads to biased results because the second stage ignores the estimation errors from the first.

Methods

The authors propose a Bayesian hierarchical joint model consisting of two interconnected sub-models:

  1. The Longitudinal Sub-model: To handle the high intensity of data, they use cubic B-splines to model the biomarker trajectory. This allows for a flexible, non-linear "mean" path. Crucially, they decompose the variance into two levels:
  2. Long-term variability ($\sigma_{bi}$): The "wiggliness" of the spline itself (minute-to-minute changes).
  3. Short-term variability ($\sigma_{ei}$): The residual "second-to-second" jitter around that spline.
  4. They also incorporate an AR1 (Autoregressive) process to account for the fact that heartbeats are naturally correlated with the ones immediately preceding them.
  5. The Outcome Sub-model: This is a probit regression that uses the estimated variances ($\sigma_{bi}$ and $\sigma_{ei}$) from the first sub-model as predictors for a binary health outcome.

By using a Bayesian framework (implemented via JAGS), the model estimates everything simultaneously. This "joint" approach ensures that the uncertainty in the variance estimates is correctly propagated through to the final prediction, leading to more robust and less biased results than traditional two-stage methods.

Datasets

  • Simulation Data: The authors generated synthetic datasets to compare their joint model against traditional two-stage approaches, specifically testing scenarios with high measurement error.
  • Michigan Work-Life Study: The primary real-world application involves hertz-level heart rate data from 165 Black women participating in a social stress study. Participants underwent the Trier Social Stress Test (TSST), where they were exposed to different audience types (e.g., racially hostile vs. friendly) to elicit stress responses.

Key Insights and Ideas

  • Variance as a Predictor, Not a Nuisance: The core "aha moment" is that the structure of biological variability is a rich source of information. A person might have the same average heart rate in two different states, but the "texture" of the fluctuations (long-term vs. short-term) changes significantly under stress.
  • The Scale of Variability Matters: The model distinguishes between "wiggliness" (the slower, physiological shifts) and "jitter" (the rapid, residual noise). The study found these two scales can move in opposite directions or have different predictive powers.
  • Joint Modeling vs. Two-Stage: The simulation results highlight a critical statistical trap: if your measurement error is high, a two-stage model will give you high-confidence but wrong answers. The joint model is "honest" about its uncertainty, maintaining high coverage rates even in noisy conditions.

Key Results

  • Simulation Superiority: In high-error scenarios, the joint model maintained a 94.5% coverage rate for parameters, while the two-stage model plummeted to 11.5%, showing severe bias.
  • Stress and Variability: In the Michigan Work-Life Study, the model revealed that participants in the "Control" group (lowest stress) actually exhibited greater minute-to-minute (long-term) variability compared to those in the stress-inducing audience groups.
  • Short-term Jitter: Conversely, second-to-second variability tended to be higher during the actual stress-test phases compared to anticipation phases, suggesting that acute stress "tightens" the physiological rhythm at one scale while potentially increasing "noise" at another.

Limitations and Future Work

  • Sample Size: The real-world application was limited by a relatively small sample size (165 participants), particularly in the control group (18 participants), which made it difficult for some complex models to converge quickly.
  • Single Biomarker: The current model only looks at one biomarker (heart rate) at a time. The authors suggest that future iterations could incorporate multiple simultaneous signals (like skin conductance and temperature) to create a more holistic "stress profile."
  • Time-Varying Variance: While the model allows variance to differ between experimental blocks (e.g., "Anticipation" vs. "Stress"), it assumes variance is constant within those blocks. Future work could allow these variances to evolve continuously over time.

Summary updated

Abstract

Intensive longitudinal biomarker data are increasingly common in scientific studies that seek temporally granular understanding of the role of behavioral and physiological factors in relation to outcomes of interest. Intensive longitudinal biomarker data, such as those obtained from wearable devices, are often obtained at a high frequency typically resulting in several hundred to thousand observations per individual measured over minutes, hours, or days. Often in longitudinal studies, the primary focus is on relating the means of biomarker trajectories to an outcome, and the variances are treated as nuisance parameters, although they may also be informative for the outcomes. In this paper, we propose a Bayesian hierarchical model to jointly model a cross-sectional outcome and the intensive longitudinal biomarkers. To model the variability of biomarkers and deal with the high intensity of data, we develop subject-level cubic B-splines and allow the sharing of information across individuals for both the residual variability and the random effects variability. Then different levels of variability are extracted and incorporated into an outcome submodel for inferential and predictive purposes. We demonstrate the utility of the proposed model via an application involving bio-monitoring of hertz-level heart rate information from a study on social stress.
BibTeX
@inproceedings{yu2024bayesian,
  author = {Mingyan Yu and Zhenke Wu and Margaret T. Hicken and Michael R. Elliott},
  title = {{A Bayesian Approach to Modeling Variance of Intensive Longitudinal Biomarker Data as a Predictor of Health Outcomes}},
  booktitle = {Statistics},
  year = {2024},
  doi = {10.1002/sim.10281},
}
Figure