A Framework for Task-Specific Signal Quality Assessment: A Case Study in Heart Rate Estimation

Signal Quality Indices (SQIs) are essential for identifying usable ECG segments in long-term recordings. Most SQIs are designed for general-purpose use and do not account for the specific demands of feature extraction or downstream analyses. In this work, we introduce a task specific SQI tailored to heart rate (HR) estimation. We develop a labelling strategy that uses a beat detector to classify 10-second ECG segments as Clean or Noisy based on whether the derived HR is within 10% of ground truth.

Using a combination of synthetic, semi-synthetic, and real-world ECG data, we trained and fine-tuned a 1D ResNet to classify segments accordingly. The model achieved F1 scores of 0.92 and 0.85 on internal test sets (PhysioNet 2014 and MIT-BIH Noise Stress Test), and generalised well to an external test set (TELE ECG), with an F1 score of 0.80. This framework presents an adaptable method for building SQIs that are aligned to specific clinical or analytical tasks, offering a more reproducible and targeted alternative to existing approaches.