Fixed-Point ECG Denoising for Wearable AFEs with Morphology Preservation

Wearable electrocardiography (ECG) systems increasingly combine an analog front end (AFE) for acquisition with an on-device digital back-end that denoises the signal before feature extraction and inference. In this setting, baseline wander (BW) and power-line interference (PLI) are dominant contaminants, yet their suppression can inadvertently deform diagnostically relevant morphology, particularly the ST segment. At the same time, many learning-based denoisers remain difficult to verify and to deploy under the fixed-point arithmetic and low-latency constraints typical of ultra-low-power microcontrollers and near-sensor accelerators. This work presents a fixed-point-oriented denoising back-end for wearable analog front ends (AFEs) that targets BW and PLI while explicitly managing distortion and quantization effects. The pipeline combines: (i) a multi-rate BW estimator with beat-aware masking and interpolation inspired by ST-preserving correction concepts, (ii) a fixed-point biquadratic-section (biquad) notch filter for 50/60 Hz PLI designed to limit transient artifacts, and (iii) a low-order low-pass stage to reduce electromyographic (EMG)-like noise within a wearable bandwidth. This paper also proposes a delay-tolerant morphology metric, QRS-aligned cross-correlation (QRS-CC), computed via QRS-aligned maximum cross-correlation (CC), and provides a Q1.15 fixed-point format simulation model with explicit rounding and saturation (applied to Stages 2–3). Finally, this paper specifies a reproducible evaluation protocol on PTB-XL (a large publicly available 12-lead ECG dataset) to support consistent corpus-level benchmarking.