An assessment of algorithms to estimate respiratory rate from the electrocardiogram and photoplethysmogram

Peter Charlton, Timothy Bonnici, Lionel Tarassenko, David A. Clifton, Richard Beale, Peter WatkinsonPhysiological 2016 · DOI

TL;DR

This study systematically assesses over 300 algorithms for estimating respiratory rate (RR) from ECG and PPG signals, revealing that novel combinations of time-domain estimation and smart fusion techniques, particularly using ECG, outperform traditional methods. The authors provide a valuable open-source toolbox to facilitate reproducible research in this critical area of physiological monitoring.

What Problem Does This Paper Address?

Respiratory rate (RR) is a crucial indicator of physiological status, yet its continuous monitoring, especially in non-clinical settings, remains challenging. While numerous algorithms exist to derive RR from readily available signals like the electrocardiogram (ECG) and photoplethysmogram (PPG), there's been no comprehensive, systematic comparison to determine which ones perform best. This lack of standardized evaluation makes it difficult for researchers and developers to choose the most effective algorithms for wearable sensors or clinical applications. The paper aims to fill this gap by rigorously comparing a vast array of algorithms under controlled conditions and providing a benchmark for future work.

Methods

The researchers adopted a modular approach, breaking down RR estimation into three stages:

  1. Extraction of Respiratory Signals: This stage involves deriving a time series that reflects respiratory modulations from the raw ECG or PPG signal. They implemented both filter-based and feature-based techniques, including band-pass filtering, wavelet transforms, and extracting specific signal features like peak amplitudes and time intervals.
  2. RR Estimation: This is where the actual respiratory rate is calculated from the extracted signal. Techniques included frequency-based methods (like Fast Fourier Transforms and auto-regressive spectral analysis) and time-domain methods (primarily breath detection based on peak and trough analysis).
  3. Fusion of RR Estimates (Optional): This stage combines estimates from different respiratory modulations (baseline wander, amplitude modulation, frequency modulation) or smooths successive estimates over time. They implemented "smart fusion" (combining estimates from different modulations with quality checks) and "temporal fusion" (smoothing estimates over time).

Crucially, they assembled algorithms by combining all possible permutations of these techniques, creating many novel algorithms. These implementations were first verified on simulated data before being tested on real-world data.

Datasets

The study utilized two types of datasets:

  • Simulated Data: Generated to verify the correct implementation of individual techniques. These signals mimicked ECG and PPG with various respiratory modulations (BW, AM, FM) across a range of heart rates and respiratory rates.
  • Real-World Data: Collected from 45 healthy adult participants (aged 18-40) at rest and during exercise. This dataset included simultaneous recordings of:
  • Lead II ECG
    • PPG (from a finger probe)
    • Oral-nasal pressure (as a gold standard for RR)
    • Thoracic Impedance Pneumography (IP) (a common clinical standard)

The real-world data was segmented into 32-second windows, and signal quality was assessed to exclude low-quality segments.

Key Insights and Ideas

  • ECG Outperforms PPG: Algorithms generally performed better and were more precise when using ECG signals compared to PPG signals. This suggests that the physiological modulations in ECG are more robustly captured or easier to extract RR from in healthy individuals.
  • Time-Domain Dominance: Time-domain breath detection techniques consistently outperformed frequency-domain methods in the top-ranked algorithms. This is attributed to their ability to handle non-stationary signals, which are common in physiological data.
  • Fusion is Key: The vast majority of high-performing algorithms incorporated a fusion component, particularly "smart fusion" (FM1), which combines information from different respiratory modulations. This suggests that integrating multiple sources of respiratory information leads to more robust estimates.
  • Novel Combinations Shine: The best-performing algorithms were often novel combinations of techniques, particularly combining time-domain estimation (ET1-5) with smart fusion (FM1). This highlights the value of exploring a wide range of algorithmic architectures.
  • ECG Algorithms Can Beat Clinical Standard: Four ECG-based algorithms were found to be more precise than Impedance Pneumography (IP), the current clinical standard for continuous RR monitoring in spontaneously breathing patients.

Key Results

  • 314 Algorithms Assessed: A total of 314 algorithms were evaluated, with 270 capable of using either ECG or PPG.
  • Top Algorithm Performance (ECG): The best algorithm (XB1,2,3 ET4 FM1) achieved a bias of 0.0 bpm and 95% Limits of Agreement (LOA) of -4.7 to 4.7 bpm, with a coverage probability (CP2) of 80.5%.
  • IP Performance: The clinical standard, IP, had a bias of -0.2 bpm and 95% LOA of -5.6 to 5.2 bpm, with a CP2 of 76.0%.
  • ECG vs. PPG Precision: 63.6% of algorithms were more precise using ECG compared to PPG.
  • Four ECG Algorithms Better Than IP: Four ECG-based algorithms demonstrated superior precision compared to IP.
  • Open-Source Toolbox: The study provides a publicly available toolbox of algorithms and data to enable reproducible research.

Limitations and Future Work

  • Implementation Matching: The authors acknowledge that their implementations might not perfectly match original descriptions for all algorithms due to variations in detail or omitted tuning steps. However, they verified their implementations using simulated data.
  • Generalizability: The study was conducted under "ideal conditions" with healthy volunteers at rest. The results may not directly extrapolate to scenarios with significant artifact, different recording equipment, or diverse populations (e.g., elderly, patients with comorbidities, or during exercise).
  • Future Directions:
    • Investigate the impact of subject population, artifact, and recording equipment on algorithm performance.
  • Assess the utility of RR algorithms in wider monitoring scenarios, including clinical settings.
  • Explore weighting fusion techniques based on the strength of individual respiratory modulations, as these can vary with age and comorbidities.
  • Consider the computational requirements of algorithms for real-time applications on power-constrained devices.
  • Welcome contributions of new techniques to the open-source toolkit.

Summary updated

Abstract

Over 100 algorithms have been proposed to estimate respiratory rate (RR) from the electrocardiogram (ECG) and photoplethysmogram (PPG). As they have never been compared systematically it is unclear which algorithm performs the best. Our primary aim was to determine how closely algorithms agreed with a gold standard RR measure when operating under ideal conditions. Secondary aims were: (i) to compare algorithm performance with IP, the clinical standard for continuous respiratory rate measurement in spontaneously breathing patients; (ii) to compare algorithm performance when using ECG and PPG; and (iii) to provide a toolbox of algorithms and data to allow future researchers to conduct reproducible comparisons of algorithms. Algorithms were divided into three stages: extraction of respiratory signals, estimation of RR, and fusion of estimates. Several interchangeable techniques were implemented for each stage. Algorithms were assembled using all possible combinations of techniques, many of which were novel. After verification on simulated data, algorithms were tested on data from healthy participants. RRs derived from ECG, PPG and IP were compared to reference RRs obtained using a nasal-oral pressure sensor using the limits of agreement (LOA) technique. 314 algorithms were assessed. Of these, 270 could operate on either ECG or PPG, and 44 on only ECG. The best algorithm had 95% LOAs of -4.7 to 4.7 bpm and a bias of 0.0 bpm when using the ECG, and -5.1 to 7.2 bpm and 1.0 bpm when using PPG. IP had 95% LOAs of -5.6 to 5.2 bpm and a bias of -0.2 bpm. Four algorithms operating on ECG performed better than IP. All high-performing algorithms consisted of novel combinations of time domain RR estimation and modulation fusion techniques. Algorithms performed better when using ECG than PPG. The toolbox of algorithms and data used in this study are publicly available.
BibTeX
@inproceedings{charlton2016assessment,
  author = {Peter Charlton and Timothy Bonnici and Lionel Tarassenko and David A. Clifton and Richard Beale and Peter Watkinson},
  title = {{An assessment of algorithms to estimate respiratory rate from the electrocardiogram and photoplethysmogram}},
  booktitle = {Physiological},
  year = {2016},
  doi = {10.1088/0967-3334/37/4/610},
}
Figure