Aims: Signal Quality Indices (SQIs) are essential for identifying usable segments in ECG data, particularly when collected via wearable devices in uncontrolled, non-clinical settings. However, most SQIs are general-purpose and do not account for the specific requirements of feature extraction or downstream analysis. For example, heart rate (HR) estimation depends primarily on accurate QRS detection, while other tasks, such as arrhythmia detection, may require clean P or T wave morphology. In this work, we propose a task-specific SQI framework, using accurate HR estimation as a case-study for implementation.
Methods: We implement a labelling strategy that classifies 10-second ECG segments as "Clean" or "Noisy" based on whether a defined beat detector produces HR estimates within 10% of the HR calculated from GT beat annotations. These task-specific labels are then used to train a 1D ResNet classifier to classify ECG segments according to their suitability for HR estimation. Due to the limited availability of labelled, noisy real-world ECG data, the model is pre-trained using synthetic data and finetuned and evaluated using real-world ECG signals. Datasets include synthetic ECG with randomly generated noise parameters, semi-synthetic signals from the MIT-BIH Noise Stress Test, and real-world data from the PhysioNet 2014 Challenge. Model generalisability is evaluated using the TELE ECG dataset as an external dataset, composed of telehealth recordings with dry electrodes.
Results: The model achieved F1 scores of 0.82 and 0.85 on the PhysioNet and MIT-BIH test sets, respectively, and 0.79 on the external TELE ECG dataset. These results demonstrate that combining synthetic data with limited real-world data can yield reliable performance across sensor types and noise conditions.
Conclusion: This work introduces a task-specific framework to ECG signal quality assessment. While focused on HR estimation, the framework is generalisable to other features or tasks, provided an appropriate labelling strategy can be defined.