Intelligent sparse sampling using min-max and adaptive methods down to 30 Hz achieved QRS detection F1 scores of 99.43–99.44%, outperforming the native 360 Hz baseline (99.19%).
Does intelligent sparse sampling (extrema-preserving downsampling) maintain QRS detection and HRV accuracy compared to native high-sampling-rate ECG?
Intelligent extrema-preserving downsampling of ECG signals to 30 Hz maintains or improves QRS detection accuracy and preserves HRV parameters while significantly reducing data volume and computational demands.
Absolute Event Rate: 99.44% vs 99.19%
Highlights What are the main findings? · Extrema-preserving downsampling maintains excellent QRS detection accuracy (F1) down to 12 Hz, often outperforming the native high-sampling-rate (360 Hz, 250 Hz) baseline. · Peak-preserving strategies preserve clinically relevant HRV parameters (mean RR, SDNN, RMSSD) with negligible error down to 12 Hz. What are the implications of the main findings? · Existing Pan-Tompkins-based system can safely downsample to 12–30 Hz, achieving 2–3× faster processing with no loss in quality. · Wearable and resource-constrained devices can reduce data volume and computational demands by 12–30× while retaining clinical-grade performance using classic algorithms. Abstract Background: Reliable QRS detection is fundamental to ECG analysis, yet the long‑held assumption that high sampling rates (≥250 Hz) are necessary has never been systematically challenged. Methods: We evaluated the Pan‑Tompkins algorithm on the MIT‑BIH Arrhythmia and QT Databases after applying four downsampling strategies (naïve decimation, max‑abs‑hold, min‑max, and adaptive min‑max) across target rates from 180 Hz down to 16 Hz. For rates below 45 Hz, sparse representations were upsampled to 45 Hz via Makima interpolation. Performance was assessed using F1 score, timing jitter, heart rate variability (HRV) parameters, compression ratio, and processing speed. Results: At 30 Hz, the min‑max and adaptive methods achieved F1 scores of 99.43–99.44%, outperforming the native 360 Hz baseline (99.19%). Mean timing error remained below 1.2 ms, and HRV parameters (mean RR, SDNN, RMSSD) were preserved within native baseline variability. Data compression reached up to 17× (59.5 MB → 3.47 MB for MITDB using gzip) and 12.3× using lightweight LZ4, and processing speed increased 2.7–3.1× without any modification to the detector. Naïve decimation failed below 45 Hz (sensitivity 94.4% at 30 Hz), confirming that extrema preservation is critical. Conclusions: The widely held assumption that high sampling rates are essential for QRS detection is false when downsampling intelligently preserves local extrema. Min‑max and adaptive methods enable clinically equivalent or superior performance at 30 Hz, offering dramatic reductions in data volume and computation for wearable and resource‑constrained ECG applications.
Jayarathna et al. (Thu,) conducted a other in ECG analysis. Intelligent sparse sampling (min-max and adaptive methods) vs. Native 360 Hz baseline was evaluated on F1 score for QRS detection. Intelligent sparse sampling using min-max and adaptive methods down to 30 Hz achieved QRS detection F1 scores of 99.43–99.44%, outperforming the native 360 Hz baseline (99.19%).