TL;DR
This paper systematically evaluates ten common QRS detection algorithms across six diverse ECG databases, highlighting their performance variations in real-world scenarios like noisy signals and dynamic telehealth applications. The study reveals that while most algorithms perform well on clean data, their accuracy significantly drops with signal degradation, with the OKB algorithm generally showing the best balance of accuracy and computational efficiency.
What Problem Does This Paper Address?
Accurate and real-time detection of the QRS complex in electrocardiogram (ECG) signals is crucial for monitoring cardiovascular diseases. While numerous QRS detection algorithms exist, many have been evaluated on limited or idealized datasets, often excluding real-world challenges like noise, arrhythmias, and data from portable monitoring devices. This lack of comprehensive evaluation across diverse and challenging scenarios makes it difficult for researchers and developers to select the most suitable algorithm for specific applications, particularly for resource-constrained mobile and wearable ECG systems. This paper aims to bridge this gap by providing a systematic performance analysis of ten widely used and efficient QRS detectors on a variety of challenging ECG databases.
Methods
The study employs a rigorous evaluation methodology involving four key experimental setups:
- High-quality vs. Low-quality ECG: Comparing algorithm performance on clean ECG signals versus those with significant noise.
- Normal vs. Arrhythmic ECG: Assessing performance on standard sinus rhythm ECGs versus those exhibiting various arrhythmias.
- Paced Rhythm ECG: Evaluating robustness to artificial pacing signals.
- Dynamic Telehealth ECG: Testing performance on ECGs collected in unsupervised, mobile environments.
A unified preprocessing pipeline was applied to all ECG signals before QRS detection. This pipeline included:
- Flat Line Detection: Identifying and handling segments with constant amplitude.
- Signal Detrending: Removing linear trends from the ECG signal using a least-squares fit.
- Band-Pass Filtering: Applying a third-order Butterworth band-pass filter (0.05–40 Hz) to isolate relevant frequency components.
Ten QRS detection algorithms were selected based on their widespread use, low computational complexity, and suitability for real-time execution on mobile devices. The evaluation metrics used were standard measures of detection performance: sensitivity (Se), positive predictivity (+P), and the F1 measure. Computational complexity was quantified by measuring the time cost to analyze a 10-second ECG segment.
Datasets
The study utilized six distinct ECG databases to cover a range of application scenarios:
- High-quality ECGs (Database A): 100 records from the 2014 PhysioNet/CinC Challenge training set (250 Hz sampling).
- Low-quality ECGs (Database A): 100 records from the 2014 PhysioNet/CinC Challenge augmented training set (360 Hz sampling), characterized by very poor signal quality.
- Normal Sinus Rhythm (NSR) ECGs (Database B): 18 long-term records from the MIT-BIH NSR database (128 Hz sampling), representing healthy subjects.
- Arrhythmia ECGs (Database B): 44 records from the MIT-BIH arrhythmia database (360 Hz sampling), excluding paced ECGs.
- Paced Rhythm ECGs (Database C): 4 records from the MIT-BIH arrhythmia database (records 102, 104, 107, 217) containing pacing signals.
- Telehealth ECGs (Database D): 250 lead-I ECGs from the TELE database (500 Hz sampling), collected in an unsupervised telehealth environment.
All ECG records had manually annotated QRS complex locations, serving as the ground truth for evaluation.
Key Insights and Ideas
- Signal Quality is Paramount: The most striking finding is the dramatic performance drop of nearly all algorithms when faced with low-quality ECG signals. This underscores the critical need for robust preprocessing or algorithms specifically designed for noisy environments.
- Arrhythmias are Manageable (Mostly): While arrhythmias introduce complexity, most tested algorithms maintained high performance, suggesting that existing algorithms are reasonably well-equipped to handle common arrhythmic patterns, especially when trained on databases like MIT-BIH.
- Paced Beats are a Specific Challenge: The RS slope algorithm showed a significant decline in performance on paced rhythm ECGs, highlighting that certain algorithmic approaches can be particularly vulnerable to artificial pacing artifacts.
- Telehealth Demands Robustness: Dynamic telehealth ECGs, collected in unsupervised settings, presented another significant challenge, with most algorithms struggling to maintain high accuracy. This points to the need for algorithms that can adapt to varying signal quality and artifacts common in such scenarios.
- OKB Algorithm Stands Out: The Optimized Knowledge-Based (OKB) algorithm consistently demonstrated superior performance across multiple challenging scenarios, particularly in low-quality and telehealth ECGs, while also maintaining high computational efficiency.
Key Results
- High-Quality ECGs: All ten algorithms achieved very high detection accuracy (F1 >99%).
- Low-Quality ECGs: F1 scores dropped significantly to below 80% for most algorithms. OKB achieved the highest F1 at 75.35%, while RS slope performed worst at 63.66%.
- Normal vs. Arrhythmic ECGs: Most algorithms achieved F1 scores above 95%. RS slope was slightly lower (94.24% on NSR, 94.44% on ARR). OKB again showed top performance (97.89% on NSR, 97.09% on ARR).
- Paced Rhythm ECGs: All algorithms performed well (>94%) except for the RS slope method, which yielded a low F1 of 78.99%.
- Dynamic Telehealth ECGs: Detection accuracies declined significantly (<80%) for most algorithms, with OKB being an exception at 80.43%. The Sixth Power algorithm performed worst at 74.08%.
- Computational Efficiency: Most algorithms were highly efficient (<4 ms per 10s segment). RS slope (94.07 ms) and Sixth Power (8.25 ms) were significantly slower. OKB was the fastest at 1.54 ms.
Limitations and Future Work
- Algorithm Selection: The study acknowledges that only ten algorithms were selected, and many other efficient QRS detectors exist. The selection was based on availability and reported efficiency.
- Implementation Variability: Some algorithms were implemented by the authors based on published descriptions, which might introduce minor differences compared to original implementations.
- Unified Preprocessing: A single preprocessing pipeline was used for fair comparison, but individual algorithms might benefit from tailored preprocessing. The effect of potential double filtering was not explored.
- Poor Signal Quality: The paper highlights that performance on poor signal quality ECGs was generally poor. Future work should focus on improving detection results in such noisy conditions, potentially through advanced preprocessing, multimodal data fusion, or algorithms specifically designed for robustness. The authors suggest that artifact masking technology, multilead ECG data fusion, and exploring different signal quality measures could be promising avenues.
Summary updated