Introduction: The Challenge of ECG Delineation
Automated cardiac diagnosis fundamentally relies on "delineation"—the precise identification of the start, peak, and end of various waves within an Electrocardiogram (ECG). While detecting the prominent R-peak (representing ventricular contraction) is a well-solved problem, accurately locating the subtle boundaries of the P-wave (atrial activity) and T-wave (ventricular recovery) remains a significant challenge in clinical practice.
Traditional methods for delineation typically fall into two categories: derivative-based approaches and wavelet-based transforms. However, both have inherent limitations. Derivative methods often require manual threshold tuning that fails under noisy conditions or baseline wander. Wavelet methods, while more robust, require the selection of a "mother wavelet" that may not suit every patient's unique cardiac morphology. This research introduces a threshold-free geometric approach designed to work reliably across standard twelve-lead clinical setups.
Methodology: A Multi-Stage Geometric Pipeline
The proposed method shifts the focus from traditional signal processing thresholds to a robust geometric analysis. The pipeline consists of five distinct stages designed to ensure high accuracy and reliability.
- Preprocessing: Signals are down-sampled to 200Hz and processed through a Butterworth bandpass filter to eliminate high-frequency noise and low-frequency baseline drift.
- QRS Location: Using a modified Pan-Tompkins approach, the algorithm employs a "Reliable QRS" check. A beat is only confirmed if it appears in at least 8 of the 12 leads simultaneously, effectively filtering out false positives like tall T-waves.
- P and T Wave Search: Using the QRS complex as a landmark, adaptive search windows are established based on the heart rate (R-R interval).
- Local Distance Transform (LDT): This core innovation identifies boundaries by drawing a straight line between two points on the signal and calculating the perpendicular distance to every point on the curve. The point of maximum distance is identified as the boundary, mimicking human visual perception of "sharp turns" in a curve.
- Multi-Lead Fusion: The algorithm collects boundary estimates from all 12 leads and applies a moving-window rule to find a centralized consensus, ignoring outliers caused by lead-specific noise.
Experimental Results
The method was rigorously evaluated using the Lobachevsky University Electrocardiograph Database (LUDB), which contains 200 records from patients with various cardiovascular diseases. Each record includes manual annotations by expert cardiologists.
When tested against a strict 40ms tolerance window—the standard for clinical relevance—the algorithm achieved over 98% accuracy for peak detection and over 96% accuracy for all wave onsets and offsets. In broader performance metrics, the system demonstrated exceptional sensitivity (Se) and positive predictive value (PPV):
- P-wave: 98.43% Se, 96.44% PPV
- QRS complex: 99.89% Se, 99.86% PPV
- T-wave: 99.21% Se, 99.85% PPV
Discussion and Conclusion
By prioritizing geometry over sensitive thresholds, this method offers a "plug-and-play" solution that maintains high precision without the need for patient-specific recalibration. The use of multi-lead fusion acts as a "jury," ensuring that noise in a single lead does not compromise the overall diagnosis.
While the current study demonstrates robustness against common clinical challenges like baseline wandering and high T-waves, future work will focus on performance during complex arrhythmias (such as atrial fibrillation) and the computational optimization required for real-time deployment on low-power wearable devices. Ultimately, this morphological delineation provides a high-quality data foundation for advanced deep-learning-based disease classification.
Summary updated