Estimation of Continuous Blood Pressure from PPG via a Federated Learning Approach

Eoin Brophy, Maarten De Vos, Geraldine B. Boylan, Tomás Ward

TL;DR

This paper introduces a novel federated learning framework that uses a single photoplethysmogram (PPG) sensor to estimate continuous blood pressure (BP). This approach allows for privacy-preserving, large-scale deployment on wearable devices, paving the way for more accessible and continuous BP monitoring.

What Problem Does This Paper Address?

Ischemic heart disease is a leading cause of global mortality, and understanding its dynamics often relies on continuous arterial blood pressure (ABP) monitoring. However, current methods for obtaining ABP are either invasive (arterial catheters), provide only spot measurements (sphygmomanometers), or are too bulky and expensive for widespread personal use (finapres). While photoplethysmography (PPG) sensors, commonly found in wearables, can provide heart rate, extracting continuous ABP from PPG alone has been a significant challenge. Existing machine learning approaches often require domain expertise for signal processing or are limited to discrete BP predictions. This paper aims to bridge this gap by developing a machine learning framework that can infer continuous ABP from PPG signals, with a focus on real-world deployability and patient privacy.

Methods

The core of this work is a Time series to Time series Generative Adversarial Network (T2T-GAN), inspired by the CycleGAN architecture. This T2T-GAN is designed to translate between different time-series modalities, specifically PPG and ABP. The framework consists of two main generators: GPA (PPG to ABP) and GAP (ABP to PPG), and two discriminators, DA and DP, to distinguish between real and generated signals.

What makes this approach novel is the integration of Federated Learning. Instead of centralizing sensitive patient data, this method trains individual client models on their local data. Only the model weights are shared and aggregated on a central server to create a global model. This preserves patient privacy while enabling collaborative learning across a large number of devices, mimicking how such a system would operate on smartwatches. The T2T-GAN architecture itself utilizes stacked LSTMs for generators and 1D CNNs for discriminators.

Datasets

The study utilized two datasets:

  • "Cuff-Less Blood Pressure Estimation" dataset: This publicly available dataset (from Kaggle and UCI Machine Learning Repository) contains preprocessed and validated ECG, PPG, and invasive ABP signals sampled at 125 Hz. For training, the first 5 records were used, segmented into 8-second intervals, yielding approximately 320 hours of data. The last 2 records were used for validation, providing about 122 hours of data.
  • "University of Queensland vital signs dataset: development of an accessible repository of anesthesia patient monitoring data for research": This dataset, focusing on patients under anesthesia, was used as an independent test set to evaluate the model's generalizability. The PPG and ABP signals from one patient (Case 5) were used, sampled at 100 Hz and segmented into 10-second intervals, resulting in 150 minutes of data.

Key Insights and Ideas

  • PPG as a sole source for continuous ABP: The research demonstrates that it's feasible to estimate continuous ABP solely from PPG signals, eliminating the need for more invasive or cumbersome measurement devices.
  • Federated Learning for privacy-preserving BP monitoring: The adoption of federated learning is a critical step towards real-world deployment. It allows for the development of robust models without compromising user privacy, a significant hurdle for health-related data.
  • T2T-GAN for time-series translation: The use of a T2T-GAN architecture effectively captures the complex, non-linear relationships between PPG and ABP time series, enabling accurate waveform generation.
  • Calibration's impact on accuracy: A short, one-minute calibration period significantly improved the accuracy of the mean arterial pressure (MAP) estimation, bringing it closer to clinical standards.

Key Results

The paper reports the following key results:

  • Mean Arterial Pressure (MAP) Estimation:
    • Without calibration: Mean error of -4.23 mmHg and a standard deviation of 23.7 mmHg.
  • With a 1-minute calibration period: Mean error of 2.54 mmHg and a standard deviation of 23.7 mmHg.
  • Time Series Similarity Metrics (on test and validation datasets):
    • The federated framework achieved comparable results to the non-federated approach, with DTW values around 62, RMSE around 0.24, and PCC around -0.33. While slightly degraded compared to the non-federated model, these results indicate good performance.
  • Qualitative Assessment: Visual inspection of generated ABP waveforms from PPG signals showed successful modeling of the underlying dynamics.

Limitations and Future Work

  • AAMI Standard: While the calibrated results show a mean error of 2.54 mmHg, the standard deviation of 23.7 mmHg still exceeds the AAMI standard of 8 mmHg for BP measurement devices. The authors attribute this to the "real-world-ness" and inherent variability of the test dataset, which includes patients under anesthesia with marked cardiovascular changes.
  • Dataset Specificity: The performance on the anesthesia patient dataset, while demonstrating generalizability, might be influenced by the specific physiological state of these patients. Further testing on more diverse datasets is warranted.
  • Future Work: The authors suggest that further work on data cleaning and preprocessing could lead to improved results. They also highlight the potential for this framework to enable lower-cost, commoditized wearable devices for continuous BP monitoring, aiding clinicians in managing cardiovascular diseases. The ability to infer one physiological waveform from another (e.g., ABP2PPG) also opens up further research avenues.

Summary updated

Abstract

Ischemic heart disease is the highest cause of mortality globally each year. This not only puts a massive strain on the lives of those affected but also on the public healthcare systems. To understand the dynamics of the healthy and unhealthy heart doctors commonly use electrocardiogram (ECG) and blood pressure (BP) readings. These methods are often quite invasive, in particular when continuous arterial blood pressure (ABP) readings are taken and not to mention very costly. Using machine learning methods we seek to develop a framework that is capable of inferring ABP from a single optical photoplethysmogram (PPG) sensor alone. We train our framework across distributed models and data sources to mimic a large-scale distributed collaborative learning experiment that could be implemented across low-cost wearables. Our time series-to-time series generative adversarial network (T2TGAN) is capable of high-quality continuous ABP generation from a PPG signal with a mean error of 2.54 mmHg and a standard deviation of 23.7 mmHg when estimating mean arterial pressure on a previously unseen, noisy, independent dataset. To our knowledge, this framework is the first example of a GAN capable of continuous ABP generation from an input PPG signal that also uses a federated learning methodology.
BibTeX
@article{brophy2021estimation,
  author = {Eoin Brophy and Maarten De Vos and Geraldine B. Boylan and Tomás Ward},
  title = {{Estimation of Continuous Blood Pressure from PPG via a Federated Learning Approach}},
  journal = {Sensors},
  year = {2021},
  doi = {10.3390/s21186311},
}
Figure