Source-linked AI summary
Analysis of Respiratory Sinus Arrhythmia with Neural Networks
Julian Szymanski, Patryk Orkisz, Higinio Mora
TL;DR
Respiratory waveform regression from ECG is rarely addressed with end-to-end deep learning, despite RSA linking heart rhythm and breathing. The paper uses ECG-derived features and a neural regressor to reconstruct respiratory waveforms; it outperforms a classical algorithm on three of four test subsets, with lower accuracy on high-frequency waveforms.
Problem
End-to-end deep-learning respiratory waveform regression is rarely studied, and most neural models predict respiratory rate rather than waveforms.
Source: §2 · p. 3Method
The paper reconstructs respiratory waveforms from ECG using RSA-related information, including recurrent-network R-wave detection and ECG-derived features.
Source: §1 · p. 2, §5 · p. 10Results
The neural network achieved better average L1 and NCC metrics on 3 of 4 test subsets than the classical algorithm, but lower regression accuracy on high-frequency waveforms.
Source: §4 · p. 9Takeaways & Limitations
The experiments support the method’s effectiveness for respiratory waveform regression within the medium breathing-rate range.
Source: §5 · p. 10Training covered 6–9 breaths per minute, and performance deteriorated at extreme test rates, especially at lower respiratory frequencies.
Source: §5 · p. 10
Abstract
from arXiv · showhide
The paper introduces a neural network-based approach for analyzing ECG signals to estimate respiratory rate by leveraging the phe- nomenon of Respiratory Sinus Arrhythmia (RSA). Our method employs a deep learning model trained to predict respiratory waveforms directly from ECG input data. To achieve this, we developed and evaluated three different neural network architectures capable of automatically extract- ing relevant features from ECG signals without the need for manual preprocessing. The proposed approach offers a robust and scalable solu- tion for non-invasive respiratory monitoring, with potential applications in healthcare and wearable technology
1 Introduction
ECG can reveal respiratory activity through RSA, the respiration-linked variation in intervals between heartbeats. The paper proposes neural-network ECG analysis for estimating HRV and RSA, addressing limitations of traditional signal-processing methods.
1 Introduction
- Variability in heart and respiratory rates is presented as an indicator of overall health and the body’s adaptation to external stimuli and surrounding conditions.Source: §1 · p. 1
- ECG supports physical and mental health assessment because autonomic nervous system activity influences the heart and other organs, including the respiratory system.HRV metrics can be derived from QRS or R-wave detection, with frequency-domain measures obtained by spectral analysis.Source: §1 · pp. 1–2
- RSA links respiration to fluctuating RR intervals: inhalation decreases vagal activity and accelerates the heart, while exhalation increases vagal activity and slows it.ECG-derived respiration uses this interaction to monitor respiratory activity when direct respiratory signals are not recorded.Source: §1 · p. 2
- The authors propose neural-network ECG analysis to estimate HRV and RSA, with claimed advantages over traditional methods in noisy conditions and through patient-data personalization.Traditional approaches are described as task-specific, dependent on manual feature extraction, and sensitive to noise and patient variability.Source: §1 · p. 2
2 ECG Signal Analysis Methods
The review groups ECG analysis into traditional, machine-learning, and deep-learning approaches, then focuses on neural methods for HRV and respiratory analysis. It identifies end-to-end respiratory waveform regression as rarely addressed, with most neural models predicting respiratory rate instead.
2 ECG Signal Analysis Methods
- ECG analysis methods are grouped as traditional, machine-learning, and deep-learning algorithms, used for tasks including anomaly detection, classification, and signal delineation.Source: §2 · p. 3
- Traditional methods typically filter signals, extract time-domain or morphological features, and then classify using threshold-based algorithms.Source: §2 · p. 3
- Prior neural ECG work spans dense, recurrent, convolutional, and Transformer architectures, including stress classification, QRS detection, R-wave detection, and camera-based heart-rate measurement.A review of QRS and ST-T localization categorized traditional and deep-learning methods and proposed ensemble training with convolutional and recurrent U-NET models.Source: §2 · p. 3
- The literature rarely addresses respiratory waveform regression with end-to-end deep learning, and existing neural models are generally used to predict respiratory rate rather than waveforms.Prior work often uses PPG or other biosensors; uses multiple datasets for respiratory-flow prediction, while reviews respiratory-rate methods and presents prediction architectures.Source: §2 · pp. 3–4
3 Neural Model for RSA Analysis
The method uses paired ECG and nasal-cannula recordings to train a respiratory-waveform regression model. It prepares ECG-derived R-wave features and respiratory annotations as aligned inputs and targets, while restricting training data to respiratory rates where RSA is expected to be informative.
3 Neural Model for RSA Analysis
- Three R-peak detectors—a recurrent model, U-Net, and Transformer encoder—were trained, and the recurrent model was selected for further work using AAMI-aligned metrics.ECG annotations were prepared and windowed for training, validation, and testing; the recurrent model was used to generate additional information for respiratory regression.Source: §3 · pp. 4–5
- The respiratory-regression input adds RRI intervals and RPA peak amplitudes to the ECG, interpolating both vectors to ECG length and min-max normalizing RRI to <0, 1>.The paired dataset contains single-channel ECG and nasal-cannula respiratory signals recorded with the Aidmed One device.Source: §3 · pp. 4–5, Figure 2 · p. 6
- Respiratory signals were annotated semi-automatically, manually verified, synchronized with ECG, smoothed, windowed, and transformed for the model’s loss criterion.The annotations include respiratory peaks and inhalation and exhalation phases; windows were band-pass filtered from 0.05 to 3 Hz.Source: §3 · p. 6
- The dataset removed signal windows with maximum respiratory frequency below 9 breaths per minute, a step associated with limiting RSA; typically, RSA relates strongly to the HF band at approximately 9–24 breaths per minute.RSA amplitude decreases above a threshold and starts vanishing above around 0.12 Hz (approximately 7.2 breaths per minute).Source: §3 · pp. 6–7
- The respiratory regressor uses two bidirectional LSTM layers, two linear layers, normalization, and interpolation to account for different ECG and respiratory sampling frequencies.The best reported training results used a loss combining absolute error and Kullback-Leibler divergence, with α = 0.9.Source: §3 · p. 7
4 Experiments and results discussion
The experiments evaluated ECG-based respiratory waveform regression against a classical approach across test subsets spanning different breathing rates. The neural model performed better on most subsets, with high-frequency waveforms as the exception.
4 Experiments and results discussion
- The R-wave detector achieved acc = 0.999 and F1 = 0.989 on Fantasia, and acc = 0.998 and F1 = 0.968 on the authors’ dataset.The detector was then used to increase the respiratory regression model’s training-data channels.Source: §4 · p. 7
- Final-epoch NCC was 0.938 on training data and 0.9436 on validation data.The regression model was trained for 30 epochs with batch size 32.Source: §4 · p. 8
- Tests used constant-rate subsets of 12 low-, 21 medium-, and 8 high-rate examples, plus two mixed-rate samples.Table 1 summarizes test metrics across BPM ranges; average NCC is mean normalized cross-correlation and average RPE is mean relative percentage error.Source: §4 · pp. 8–9, Table 1 · p. 9
- For 3 of 4 test subsets, the neural model achieved better average L1 and NCC metrics than the classical algorithm; high-frequency waveforms were the exception.Table 2 presents the comparison between the neural model and NeuroKit.Source: §4 · p. 9, Table 2 · p. 10
- Above around 0.12 Hz, or approximately 7.2 breaths per minute, RSA starts vanishing, and very fast breathing can make it inconsistent.The paper limited training examples in relation to this RSA behavior.Source: §4 · p. 10
5 Conclusions and future works
The neural method reconstructs respiratory waveforms from ECG using RSA and outperforms the classical algorithm in regression accuracy. Its performance is strongest in the medium breathing-rate range and deteriorates at extremes outside the training range.
5 Conclusions and future works
- The proposed neural method was superior to the classical algorithm in regression accuracy.The analysis used the obtained metrics to confirm this comparison.Source: §5 · p. 10
- Training on 6–9 breaths per minute yielded effective performance in the medium range, but performance deteriorated at extreme rates outside that training range.Testing covered the full BPM range <4, ∞>.Source: §5 · p. 10
- Lower respiratory frequencies showed a significant accuracy drop, possibly due to increased RRI-channel noise and few training examples.Source: §5 · p. 10
- Future work targets broader respiratory-frequency coverage, wearable-device data, personalization, and factors that may shift individual sensitivity to breathing.Planned analyses include gender, age, and ANS activity measures such as eye activity or EEG signals.Source: §5 · p. 11