Abstract
Left Ventricular Hypertrophy (LVH) is a major cardiovascular risk factor, linked to heart failure, arrhythmia, and sudden cardiac death, often resulting from chronic stress like hypertension. Electrocardiography (ECG), while varying in sensitivity, is widely accessible and cost-effective for detecting LVH-related morphological changes. This work introduces a bilateral signal warping (BSW) approach to improve ECG-based LVH diagnosis. Our method creates a library of heartbeat prototypes from patients with consistent ECG patterns. After preprocessing to eliminate baseline wander and detect R peaks, we apply BSW to cluster heartbeats, generating prototypes for both normal and LVH classes. We compare each new record to these references to support diagnosis. Experimental results show promising potential for practical application in clinical settings.
Introduction
Left Ventricular Hypertrophy (LVH)—the thickening of the heart's left ventricle—is a major precursor to heart failure and sudden cardiac death. While Electrocardiography (ECG) is the standard screening tool, traditional diagnostic rules, such as the Sokolow-Lyon criteria, often suffer from low sensitivity, missing many patients who require intervention.
Modern deep learning models have achieved high accuracy in detecting LVH, but their "black-box" nature limits clinical adoption. Physicians require interpretable evidence to trust automated diagnoses. This research proposes a system that mimics clinical reasoning by comparing a patient's heartbeat to a library of validated prototypes using a novel Bilateral Signal Warping (BSW) technique.
Methodology
The proposed diagnostic pipeline is designed to ensure both signal quality and interpretability through several key stages, from initial preprocessing to the generation of dynamic prototypes.
1. Eligible Candidate Selection
To ensure the prototype library contains only high-quality, representative signals, the system filters out noisy or irregular heartbeats. A variability metric ($v_h$) is employed to select patients with consistent cardiac cycles, ensuring the resulting prototypes are clean and reliable.
2. Bilateral Signal Warping (BSW)
Standard Dynamic Time Warping (DTW) aligns signals by shifting them in time but fails to account for amplitude variations. In LVH, the amplitude of the QRS complex is a critical diagnostic marker. BSW addresses this by jointly optimizing for both amplitude scaling ($r(t)$) and time shifting ($s(t)$).
3. Dynamic Prototype Generation and Matching
The library is built using an iterative matching process. The Blossom algorithm, a graph theory method for non-bipartite matching, identifies similar heartbeat pairs across different patients. These are then warped together to form a single prototype. For diagnosis, a new patient's ECG is compared against these normal and LVH prototypes, and a distance metric $d$ is calculated based on the warping required for a match.
Datasets
The study utilized the PTB-XL 12-lead ECG dataset. The initial pool consisted of 9,069 normal records and 456 LVH records. After filtering for signal regularity and balancing classes, 215 LVH and 256 normal records were used to generate the prototype library. Performance was validated on a separate set of 200 patients (100 normal, 100 LVH).
Results and Discussion
The BSW prototype matching method was compared against gold-standard clinical criteria, including Modified Cornell and Sokolow-Lyon.
- Superior Sensitivity: The BSW method significantly outperformed traditional criteria. While traditional methods are highly specific, they often fail to identify actual LVH cases. In contrast, the prototype matching achieved a balanced performance, correctly identifying 80% of LVH cases in the test sample, compared to just 17% for the Modified Cornell criteria.
- Interpretability: Unlike deep learning models, this system provides visual evidence. Clinicians can see a patient's heartbeat overlaid on the closest LVH prototype, highlighting specific morphological features like "strain" patterns or inverted T-waves.
Conclusion
Bilateral Signal Warping offers a promising, interpretable alternative for LVH diagnosis. By capturing both timing and amplitude variations, it provides a "glass-box" approach that aligns with clinical workflows. Future work will focus on expanding the prototype library to include other cardiac conditions and improving the system's robustness to noisy real-world data.
Summary updated