Background: Wearable ECG data are increasingly collected in ambulatory settings, where noise is prevalent. However, many ECG signal quality assessment methods have been developed using curated datasets that may not reflect real-world conditions. This study systematically evaluates ECG quality assessment methods on prolonged real-world wearable data with clinician-annotated labels.
Methods: Four publicly available approaches: template matching, fuzzy logic, rule-based, and CNN-LSTM, were evaluated using two datasets: (1) a synthetic noise benchmark derived from the MIT-BIH Arrhythmia Database (460 segments; with electrode motion [EM], baseline wander [BW], and muscle artifact [MA] introduced at 6 SNR levels), and (2) real-world wearable ECG data from a 14-day patch monitor (3,403 segments). Segments were independently annotated by two cardiology clinicians using an ordinal scale reflecting clinical interpretability (e.g., interpretable vs. non-interpretable; κ = 0.88). The task was formulated as a classification problem, and performance was evaluated on balanced and imbalanced test sets. Computational efficiency was assessed using mean processing time per segment and throughput.
Results: The CNN–LSTM method achieved the highest performance (accuracy 88.5%, recall 79.9%) on the synthetic benchmark, and across noise types: EM (80.8%, 66.2%), BW (accuracy 80.8%, recall 88.7%), and MA (94.2%, 91.5%). However, its advantage over the rule-based approach with balanced real-world data was modest (accuracy/recall 75.0%) with only a 4% relative improvement to the rule-based method. In contrast, computational cost differed substantially: CNN–LSTM required 5509.0 milliseconds per segment (0.18 segments/seconds), whereas the rule-based method required 5.9 ms (170.4 segments/seconds), corresponding to ~950× higher throughput.
Discussion: While deep learning methods achieved the highest performance, its gains were limited in real-world wearable ECG data and came at higher computational cost. These findings highlight a key trade-off between performance and efficiency, suggesting that lightweight methods may be more suitable for continuous ambulatory ECG monitoring.