Multi-Tier Validation of a Single-Lead Wearable ECG Signal-Processing Pipeline: Clinical Comparison and Algorithmic Benchmarking on Public Expert-Annotated Databases

Figures

• 

  Figure 1. Graphical overview of the study’s design and contributions. The locked SmartGuard signal-processing pipeline (left; Section 2.2) is evaluated by three complementary validation tiers (centre; Section 2.3), each mapped to its role within the V3 framework—clinical validation (Tier 1) and analytical validation under clean (Tier 2) and deliberately adverse (Tier 3) conditions. Each tier feeds one of the study’s three outputs (right): a reusable validation template with an operational checklist, per-tier accuracy evidence, and an explicit mapping from residual failure modes to concrete algorithmic remedies. Grey arrows indicate the flow of the locked pipeline output into the three validation tiers; orange arrows link each tier to its principal output. Navy outlines denote the device and pipeline stages; the beige, blue, and green tier backgrounds distinguish the clinical, clean-benchmark, and stress-test tiers, respectively; orange outlines mark the study outputs.

• 

  Figure 2. Block diagram of the SmartGuard single-lead ECG signal-processing pipeline. Raw ECG samples from the device’s internal acquisition channels are calibrated to millivolts, assembled into a unified waveform, and cleaned. R-peaks are detected with an artefact-corrected published detector, followed by discrete wavelet transform (DWT) delineation of P–Q–R–S–T fiducial points. Prespecified derivative-based QRS-boundary and tangent-based T-wave-end refinements are applied, followed by robust beat aggregation and extraction of HR, RR, QRS, QT, P/Q/R/S/T wave durations and amplitudes, and Bazett-corrected QTc. Production tuning constants are intentionally omitted. Black arrows indicate the order of processing; box colours group the stages: blue, raw input; green, calibration, assembly, and cleaning; yellow, detection and delineation; orange, prespecified refinements and robust aggregation; violet, feature extraction and output.

• 

  Figure 3. End-to-end illustration of the SmartGuard signal-processing pipeline on a representative 60-s recording from the Tier 1 study. (a) Raw single-lead signal after analogue-front-end conversion, showing the two consecutively acquired 30-s segments; (b) Cleaned signal with R-peak detection (n = 81; the transient at the segment junction is suppressed by artefact correction); (c) 3.6-s excerpt with DWT fiducials and custom refinements: P onsets/peaks/offsets, refined QRS onsets/offsets, and tangent-method T-wave ends; (d) Derivative-based QRS-width refinement for a representative beat—boundaries are placed where the smoothed absolute derivative falls below 4% of its local maximum, narrowing the DWT estimate from 125 ms to 90 ms (record medians 124 ms and 86 ms, respectively); (e) Tangent-method T-wave end—the tangent through the steepest post-peak slope is extrapolated to the baseline defined at the QRS-onset amplitude. Colour coding across panels: dark blue, ECG signal; red, R peaks, the 4% derivative threshold, and the tangent line; green, P-wave fiducials; orange, refined QRS boundaries; purple, T-wave fiducials and tangent-derived T-wave ends; dotted grey, DWT-derived reference…

• 

  Figure 4. Representative expert-annotated excerpts from the two public validation databases. (a) LUDB record 1 (lead I): clean 10-s sinus rhythm with cardiologist-verified P (green triangles), QRS (red circles) and T (purple squares) peak annotations and wave boundaries (orange lines); (b) QTDB record sel16265 (MIT–BIH Normal Sinus Rhythm Database): sinus rhythm with upright T waves; (c) QTDB record sele0107 (European ST-T Database): deeply inverted T waves and low-amplitude P waves; (d) QTDB record sel808 (MIT–BIH Supraventricular Arrhythmia Database): broad QRS complexes, tall asymmetric T waves, and distinct U waves (cyan diamonds). Panels (b–d) illustrate the morphological heterogeneity that makes QTDB a deliberately harder delineation target than LUDB (Section 2.3.3). In all panels the dark-blue line is the ECG signal (amplitude in millivolts); orange vertical lines mark expert-annotated wave boundaries.

• 

  Figure 5. Bland–Altman plots for Tier 2 (LUDB, n = 200). Each scatter point represents one record. Solid red line: mean bias (pipeline−reference). Dashed red lines: upper and lower 95% limits of agreement (bias ± 1.96 × SD). Dotted grey line: zero reference. Heart rate (top-left) and QRS duration (top-right) show tight limits of agreement around small positive biases. QT (bottom-left) and QTc (bottom-right) show low aggregate bias but substantially wider limits of agreement, driven by a subset of records with discrepant T-wave end placement. Point colours identify the metric panels only (blue, heart rate; green, QRS duration; orange, QT; purple, QTc) and carry no additional meaning.

• 

  Figure 6. Bland–Altman plots for Tier 3 (QTDB, n ≤ 103). Compared with Tier 2 (Figure 5), limits of agreement are wider for all four panels—particularly for QT and QTc (bottom row)—reflecting the increased morphological complexity and longer-duration recordings of the QTDB stress-test set. The QRS duration panel (top-right) shows a clear positive bias of approximately +17 ms, indicating that the derivative-based QRS-width refinement systematically extends QRS boundaries on the heterogeneous QTDB morphologies. Representative examples of the underlying QTDB morphologies are shown in Figure 4c,d.

• 

  Figure 7. Cross-tier comparison of pipeline mean absolute error (MAE) for the five metrics most directly comparable across the three validation settings. Tier 1: Paired clinical recordings of healthy adults against the Bionet Cardio 7 Lead I reference (n = 31). Tier 2: Per-record comparison against cardiologist annotations on LUDB (n = 200). Tier 3: Per-record comparison against hybrid expert annotations on QTDB (n ≤ 103). Numerical values are shown above each bar.