ECG Denoising and R-Peak Detection under Indian Power-Line Noise: A PhysioNet Benchmark

ECG-Bench is a declared-protocol benchmark of ECG denoising filters, scored not only on signal-quality metrics but on the thing that actually matters downstream — whether R-peak detection survives the filtering. All experiments run on the MIT-BIH Arrhythmia Database from PhysioNet, the most heavily cited public ECG archive in existence.

Every ECG recorded in an Indian hospital or on a wearable fights the same three enemies: baseline wander from respiration and electrode drift, 50 Hz power-line interference (India's mains frequency — most textbook experiments default to the American 60 Hz), and muscle-artefact noise that resembles real cardiac content. Textbook answers exist for each — high-pass FIR filtering, IIR notch filtering, wavelet thresholding — but published comparisons usually report signal-to-noise ratio (SNR) alone, which hides a classic trap: an aggressively denoised signal can look clean while having blunted the very QRS features a detector needs.

The study implements four filter families in one framework, degrades each MIT-BIH record with declared, parameterised noise (three types at three SNR levels), and evaluates every combination twice: SNR improvement and distortion against the clean signal, then R-peak detection sensitivity, positive predictive value and F1 against the database's beat annotations with a ±150 ms tolerance.

The deliverable is a paper-style study with a reproducible Python pipeline: per-record and aggregate tables, the noise-vs-accuracy trade-off made explicit, and honest guidance on which filter to ship in a low-cost Indian ECG front end.

ECG systems deployed in India must handle 50 Hz power-line interference and respiration-band baseline wander, but filter choices in published comparisons and student projects are rarely justified by a like-for-like evaluation: different records, different noise models, different SNR definitions, and almost never a measurement of whether denoising actually improves or degrades R-peak detection — the step every clinical and consumer ECG application depends on. There is no reproducible benchmark that (a) injects declared, seeded noise of the three dominant Indian ambient types at controlled SNR levels into a standard public ECG database, (b) compares the classical filter families under identical conditions, and (c) scores them on both signal fidelity and downstream detection accuracy against ground-truth annotations. This dissertation fills that gap on the MIT-BIH Arrhythmia Database with a frozen protocol and a fully scripted pipeline, quantifying the SNR-versus-detection trade-off that single-metric evaluations hide, and giving Indian biomedical engineers a defensible basis for front-end filter selection.

Abstract

This dissertation benchmarks four ECG denoising approaches — FIR high-pass, IIR notch, wavelet shrinkage and cascaded combinations — on the MIT-BIH Arrhythmia Database under declared synthetic degradation. Three noise types (baseline wander, 50 Hz interference, pseudo-EMG) are injected at 6, 12 and 24 dB SNR with fixed seeds, and every filter × noise × level combination is evaluated on two pre-declared metric families: signal fidelity (SNR improvement, distortion) and clinical utility (R-peak detection sensitivity, PPV and F1 of a Pan–Tompkins detector against PhysioNet beat annotations, at ±150 ms tolerance). The study quantifies the gap between the two metric families — the conditions under which the highest-SNR filter is not the best filter for detection — and reports per-record breakdowns and cost. All numbers regenerate from the public database and a frozen protocol.

Introduction

India's ECG landscape spans home monitors to ICU telemetry, all fighting the same interference: 50 Hz mains coupling, respiration-driven baseline wander and EMG contamination in ambulatory use. Filtering is the most consequential stage of an ECG chain — a bad filter destroys morphology or detection — yet filter selection is often inherited from American 60 Hz examples or from papers that report only SNR.

Existing system and literature gap

The denoising literature is rich but its evaluation practice is inconsistent: papers compare filters on different records, noise realisations and SNR definitions, and rarely connect SNR gains to detection accuracy. A filter can raise SNR while smearing QRS slope or shifting detected R-peaks. What is missing is a declared-protocol benchmark on one public database, where identical degraded signals are scored both by signal metrics and by a fixed detector against ground-truth annotations — so the SNR-versus-detection trade-off is measured, not assumed — on MIT-BIH, the field's common reference.

Proposed work

• Download all 48 MIT-BIH records and annotations via wfdb-python into a versioned, integrity-checked store.
• Implement four filter families behind one interface: FIR high-pass (0.5 Hz, linear phase), IIR Butterworth notch (50 Hz, declared order and Q), wavelet shrinkage (symlets, soft thresholding) and their cascade.
• Generate declared degradations — synthetic baseline wander, 50 Hz sine with amplitude drift and band-limited pseudo-EMG — each injected at 6, 12 and 24 dB SNR with fixed seeds, so every filter sees identical inputs.
• Evaluate: SNR improvement and distortion against clean signals; then score an unmodified Pan–Tompkins detector's sensitivity, PPV and F1 against annotations at ±150 ms, with per-record tables and compute time.
• Analyse the divergence: identify where SNR gain and detection accuracy disagree, and state the operational recommendation.

Feasibility

• Technical: everything runs in Python (NumPy, SciPy.signal, PyWavelets, wfdb) on a 16 GB laptop; the full factorial benchmark completes in minutes.
• Data: MIT-BIH is public, citable and the field's standard reference; no ethics timeline is required.
• Economic: zero licence cost; no hardware.
• Operational (research): protocol, noise models, seeds and tolerances freeze before evaluation; every table regenerates with two commands — the audit trail an M.Tech viva demands.

In scope

• The 48 records of the MIT-BIH Arrhythmia Database (channel 1), with PhysioNet beat annotations as detection ground truth.
• Four filter families with declared parameters; one unmodified Pan–Tompkins detector fixed across all conditions (the detector is the measuring instrument, not a variable).
• Three synthetic noise types at three SNR levels with fixed seeds; SNR-improvement and distortion metrics; detection metrics at a declared ±150 ms tolerance.
• Room for per-record analysis of difficult records (e.g. paced-rhythm and heavy-arrhythmia segments) with honest reporting.

Out of scope

• Real patient recording campaigns or hardware front-end design (instrumentation amplifiers, ADC chain) — the study stops at the digital filter stage.
• New detector design and arrhythmia classification; the detector is held constant to isolate filter effects.
• Real-ambulatory noise corpora (listed as future scope with the PhysioNet Noise Stress Test Database as the extension path).
• Multichannel fusion and deep-learning denoisers; the benchmark's contribution is the protocol and the classical families' honest comparison.

The work follows a quantitative, comparative-experimental research design in six phases over 22 weeks, with the evaluation protocol (noise models, SNR levels, metrics, tolerance windows, seeds) frozen at the end of Phase 1 — before any filter is scored.

Validation inside the study: filter frequency-response tests against design specifications; a synthetic ECG with planted R-peaks recovered end-to-end; detector performance on clean MIT-BIH validated against published Pan–Tompkins-class sensitivity figures before any degraded run; and full re-runs reproducing every table from fixed seeds.

The pipeline separates signal generation from measurement: filters and the detector are instruments, the degradation machinery is the experiment, and the scoring layer is the only place metrics are computed.

flowchart TD
  A["MIT-BIH Arrhythmia Database\n48 records + annotations"] --> B["Ingestion via wfdb-python\nintegrity checks"]
  B --> C[("Versioned store\nper-record signals + beats")]
  C --> D["Noise degradation machinery\nBW, 50 Hz, pseudo-EMG\n6/12/24 dB, fixed seeds"]
  C --> E["Clean reference\nfor fidelity metrics"]
  D --> F["Filter families\nFIR HP, IIR notch, wavelet, cascade"]
  F --> G["Pan-Tompkins detector\nunmodified across conditions"]
  E --> H["Scoring layer\nSNR gain, distortion"]
  G --> I["Detection scoring\nsensitivity, PPV, F1 at ±150 ms"]
  C --> I
  H --> J["Results tables and figures\nper-record + aggregate"]
  I --> J

Three design decisions matter at the viva. First, the Pan–Tompkins detector is held unmodified across every condition: it is the measuring instrument, so any change to it mid-benchmark would invalidate cross-condition comparison — the protocol freezes it exactly as the truth definition is frozen in other benchmarks. Second, the degradation machinery writes seeded noise files once; filters consume the identical degraded signals, so filter differences cannot be attributed to different noise realisations. Third, the scoring layer computes both metric families from paired series — SNR gain and distortion against the clean reference, detection metrics against the database annotations — so the study's central contribution, the divergence between the two metric families, is a direct table comparison rather than a cross-paper argument.

• Extend the degradation machinery to the PhysioNet Noise Stress Test Database's real ambulatory noise (baseline wander and muscle artefact recordings) and re-run the benchmark on real contamination.
• Add adaptive interference cancellation (reference-channel LMS) and compare against the fixed 50 Hz notch.
• Score a second, structurally different detector to test whether the filter ranking transfers across detectors.
• Study filter behaviour on paced-rhythm and malignant-arrhythmia records specifically, where detection failure is costliest.
• Port the recommended cascade to an embedded target and measure real-time cost for a low-cost Indian single-lead ECG front end.