Vascular waveform analysis using Bayesian pulse deconvolution

Parker S. Ruth, Tommy DeBenedetti, Lily O'Brien, James A. Landay, Todd Coleman, Emily B. Fox2026 · DOI

TL;DR

This paper introduces a novel Bayesian pulse deconvolution method that jointly addresses signal filtering, pulse timing detection, and pulse shape extraction for vascular waveforms. By leveraging a generative model with domain knowledge, it significantly outperforms traditional sequential methods in accuracy and robustness, paving the way for more reliable health monitoring technologies.

What Problem Does This Paper Address?

Analyzing vascular waveforms, such as those from photoplethysmography (PPG), is crucial for monitoring vital signs, diagnosing conditions, and predicting health outcomes. However, extracting meaningful information from these signals is challenging due to noise and the inherent complexity of the underlying physiology. Current approaches typically break down this analysis into three sequential steps: filtering out noise, detecting the timing of each pulse, and extracting the shape of each pulse. The problem is that these steps are not truly independent; errors in one step propagate and negatively impact the others. For instance, a filter might distort the signal, making accurate pulse timing detection difficult, which in turn affects the quality of the extracted pulse shape. Furthermore, obtaining "ground truth" data for training and testing these algorithms is problematic, as sensor measurements are always noisy, and even gold-standard methods like ECG have limitations for precise vascular pulse timing.

Methods

The core innovation is a unified Bayesian pulse deconvolution algorithm. Instead of a sequential pipeline, it models vascular waveforms using a generative process that jointly accounts for:

  • Signal Filtering: It models the observed signal as a convolution of an impulse train (representing pulse timings) with an impulse response (representing pulse shape), plus a residual noise term. This allows for more effective noise removal without introducing the phase distortions common in traditional linear filters.
  • Pulse Timing Detection: It treats pulse timing as a semi-periodic point process, informed by physical and biological domain knowledge. This is encoded through Bayesian priors on pulse intervals, assuming a constant underlying heart rate and modeling relative inter-beat intervals.
  • Pulse Shape Extraction: It models the pulse shape using a weighted combination of Fourier basis elements, allowing for the explicit modeling and handling of overlapping pulses, which is a significant limitation in averaging-based methods.

The algorithm uses Bayesian inference to estimate the most probable (Maximum A Posteriori - MAP) values for these unobservable variables, given the observed signal. Crucially, it also quantifies uncertainty in its estimations, providing credible intervals for the inferred parameters.

Datasets

The paper utilizes two main sources of data:

  • Simulated Waveforms: 100 simulated eleven-second vascular waveforms sampled at 64 Hz with -20dB Gaussian noise were generated. These were crucial for evaluating performance against known "ground truth" (noiseless signals, exact pulse times, and isolated pulse shapes).
  • Human Recordings (Aurora-BP Dataset): 105 recordings from 21 individuals were used to evaluate the algorithm's performance on real-world data. This dataset is notable for its time-synchronized electrocardiogram (ECG) signals, which serve as a proxy ground truth for pulse timing. The authors specifically mention using a subset of tonometry waveforms to fit priors and a non-overlapping subset of photoplethysmography (PPG) waveforms for evaluation to avoid bias.

Key Insights and Ideas

  • Joint Modeling is Key: The fundamental insight is that treating filtering, timing, and shape extraction as interdependent problems and solving them jointly leads to superior performance compared to sequential approaches.
  • Bayesian Priors Encode Domain Knowledge: By encoding physical and biological constraints (like typical heart rate ranges and pulse shape characteristics) into Bayesian priors, the algorithm can infer meaningful information even with noisy or incomplete data, overcoming the lack of perfect ground truth.
  • Handling Overlapping Pulses: The explicit modeling of pulse shape and timing allows the algorithm to disentangle and accurately extract individual pulse shapes even when they overlap, a common issue in high heart rate scenarios.
  • Uncertainty Quantification is Valuable: The Bayesian framework naturally provides uncertainty estimates, which are critical for clinical applications where knowing the confidence in a measurement is as important as the measurement itself.

Key Results

  • Simulation Performance:
    • Filtering: 90% reduction in median filtering error.
  • Pulse Timing: 60% reduction in pulse timing error (median RMSE of 4.9 ms vs. 12.0 ms for the best traditional heuristic).
  • Pulse Shape Extraction: 85% reduction in shape extraction error (median error of 6.9% vs. 48-49% for typical algorithms).
  • Human Recordings Performance:
    • Pulse Interval Estimation: 40% lower pulse interval estimation error (median RMSE of 5.1 ms vs. 8.3 ms for typical algorithms, with p=1e-10).

Limitations and Future Work

  • Computational Complexity: The current implementation faces challenges with multi-modality and scaling for very long signals or high pulse counts. Future work will explore buffered stochastic gradient MCMC or variational inference for improved scalability.
  • Real-time Applications: Inference speed needs further optimization for real-time applications.
  • Model Extensibility: The authors suggest future work could extend the model to:
  • Handle varying heart rates (e.g., arrhythmias, respiratory variations).
  • Incorporate more flexible pulse shape models (e.g., normalizing flows).
  • Accommodate non-Gaussian noise or motion artifacts.
  • Apply to other biosignals like ECG.
  • Personalize priors based on individual demographics.
  • Downstream Applications: While the paper demonstrates improved accuracy, further work is needed to evaluate its impact on specific downstream clinical applications (e.g., blood pressure estimation, blood oxygenation monitoring).

Summary updated

Abstract

Vascular waveforms, which measure bulk flow in blood vessels, are widely used to measure vital signs, diagnose conditions, and predict long-term health outcomes. Analyzing vascular waveforms depends on three fundamentally interdependent tasks: signal filtering, pulse timing detection, and pulse shape extraction. We hypothesized that Bayesian pulse deconvolution can achieve improved performance on all three tasks by solving them jointly. This method uses an analytical, generative model of vascular waveforms with priors informed by physical and biological domain knowledge. In simulations, Bayesian pulse deconvolution achieves better performance on all tasks compared with existing algorithms: 90% reduction of median filtering error, 60% reduction in pulse timing error, and 85% reduction in shape extraction error. The advantages in simulations extend to human recordings of photoplethysmography waveforms. Taking real time-synchronized electrocardiogram R-R intervals as a proxy ground truth, Bayesian pulse deconvolution achieves 40% lower pulse interval estimation error (RMSE =5.1 ms) compared with typical algorithms (RMSE = 8.3 ms, p=1e-10). By extracting more accurate and informative insights from vascular waveforms, Bayesian pulse deconvolution could advance a wide array of health technologies that rely on interpreting signals from blood vessels.
BibTeX
@misc{ruth2026vascular,
  author = {Parker S. Ruth and Tommy DeBenedetti and Lily O'Brien and James A. Landay and Todd Coleman and Emily B. Fox},
  title = {{Vascular waveform analysis using Bayesian pulse deconvolution}},
  year = {2026},
  doi = {10.64898/2026.02.09.699383},
}
Figure