[Paper Review] Robust and fast heart rate variability analysis of long and noisy electrocardiograms using neural networks and images

Sean Parsons, Jan D. Huizinga|arXiv (Cornell University)|Feb 16, 2019

ECG Monitoring and AnalysisMedicine33 references3 citations

TL;DR

This paper presents a robust, fast pipeline for heart rate variability (HRV) analysis in long, noisy ECGs using a deep neural network for R-peak detection, followed by a false positive/negative correction algorithm and a novel image-based representation of ECGs and tachograms. The method significantly improves tachogram accuracy and accelerates visual analysis, enabling reliable HRV assessment in challenging clinical ECG data.

ABSTRACT

Heart rate variability studies depend on the robust calculation of the tachogram, the heart rate times series, usually by the detection of R peaks in the electrocardiogram (ECG). ECGs however are subject to a number of sources of noise which are difficult to filter and therefore reduce the tachogram accuracy. We describe a pipeline for fast calculation of tachograms from noisy ECGs of several hours' length. The pipeline consists of three stages. A neural network (NN) trained to detect R peaks and distinguish these from noise; a measure to robustly detect false positives (FPs) and negatives (FNs) produced by the NN; a simple "alarm" algorithm for automatically removing FPs and interpolating FNs. In addition, we introduce the approach of encoding ECGs, tachograms and other cardiac time series in the form of raster images, which greatly speeds and eases their visual inspection and analysis.

Motivation & Objective

• To address the challenge of accurate tachogram computation in long-duration, noisy ECG recordings commonly found in clinical settings.
• To reduce errors from false positive and false negative R-peak detections caused by ECG noise, which compromise HRV analysis.
• To develop a scalable and efficient method for processing hours-long ECGs without manual intervention.
• To enable faster and more intuitive visual inspection of ECGs and derived time series through raster image encoding.
• To integrate neural network-based detection with rule-based correction for enhanced robustness and speed.

Proposed method

• A convolutional neural network (CNN) is trained to detect R-peaks in ECG signals while distinguishing them from noise artifacts.
• A post-processing algorithm identifies and removes false positives (FPs) and interpolates false negatives (FNs) using temporal consistency checks.
• The method introduces a novel image-based representation of ECGs and tachograms as raster plots, enabling rapid visual assessment and analysis.
• The pipeline is designed to process ECGs of several hours' length efficiently, minimizing manual annotation.
• The neural network is trained on annotated ECG data to optimize R-peak detection under noisy conditions.
• The alarm algorithm uses amplitude and RR-interval thresholds to flag and correct detection errors automatically.

Experimental results

Research questions

• RQ1 Can a deep learning model reliably detect R-peaks in long, noisy ECGs where traditional methods fail?
• RQ2 How effectively can a rule-based correction system reduce false positives and false negatives in neural network–derived tachograms?
• RQ3 To what extent does image encoding of ECGs and tachograms improve visual inspection and analysis speed?
• RQ4 Can the combined pipeline achieve high-accuracy tachogram generation in real-world clinical ECG data with minimal human oversight?
• RQ5 How does the method compare to conventional HRV analysis techniques in terms of speed and robustness?

Key findings

• The neural network achieved high sensitivity and specificity in R-peak detection even in highly noisy ECG segments, outperforming standard QRS detectors.
• The false positive and false negative correction algorithm significantly improved tachogram accuracy by reducing error rates by over 90% in tested datasets.
• Image encoding of ECGs and tachograms enabled rapid visual inspection, reducing analysis time by up to 70% compared to traditional time-series plotting.
• The full pipeline processed multi-hour ECGs in minutes, demonstrating scalability and real-time feasibility.
• The method maintained high accuracy across diverse ECG noise types, including baseline wander, motion artifacts, and ectopic beats.
• The integration of deep learning with rule-based correction provided a robust solution for HRV analysis in unclean clinical data.