Improving ECG diagnostic performance under ambulatory noise with scalable long-range state-space denoising modelling

Background Long-duration ECG monitoring is increasingly used to detect intermittent and subtle cardiac abnormalities, yet real-world recordings are frequently contaminated by non-stationary artifacts (baseline wander, motion, muscle activity) that impair interpretability and can degrade automated diagnosis—particularly as recording length increases. Purpose To develop and clinically validate an efficient long-range ECG denoising approach that improves waveform fidelity and preserves diagnostically relevant information for downstream disease prediction in noisy ambulatory conditions. Methods We introduced a convolutional encoder-decoder denoiser augmented with selective state-space (Mamba) blocks placed at the network bottleneck to model long-range temporal context with linear-time scaling. The approach was evaluated on (i) clean synthetic ECGs (McSharry simulator), and two clinical benchmarks: (ii) European ST-T (90 ambulatory 2-hour recordings) and (iii) PTB-XL (21,837 10-s 12-lead ECGs)—see fig 1. Noise segments from the MIT-BIH Noise Stress Test Database were added to generate realistic artifact regimes (baseline wander, electrode motion, muscle artifact, plus white noise) at recommended SNRs. Reconstruction was quantified using RMSE (fig 2.A) and SNR. Clinical utility was assessed on PTB-XL by classifying superclasses (CD, HYP, MI, NORM, STTC) from denoised 12-lead ECGs using a classifier trained on clean data, reporting macro-AUROC (fig 2.B). Results Across synthetic and clinical cohorts, Mamba augmentation consistently improved reconstruction versus convolutional (UNet, IMUnet, DRNet, DAE) and recurrent (DRNN) architectures, and Mamba-equipped filter (MECGE) baselines with SNR gains of up to 3 dB. On PTB-XL, our model achieved RMSE 0.31 ± 0.01 (fig 2.A) and SNR 16.16 ± 0.10 dB. Importantly, denoising improved diagnosis: disease prediction on denoised PTB-XL ECGs reached macro-AUROC >0.913 (fig 2.B), with strong performance across clinically relevant categories (e.g., CD and STTC). Benefits increased with longer context: SNR gains rose from ~1 dB (5-s segments) to ~2 dB (40-s segments), supporting utility for extended monitoring. Conclusion Selective state-space modeling at the bottleneck of encoder-decoder models provides a scalable, clinically effective backbone for long-range ECG denoising, improving robustness to ambulatory artifacts while preserving diagnostic information for downstream classification. While further analysis and validation is required, this supports deployment-oriented ECG cleaning for long-duration monitoring and large-scale clinical pipelines.Fig. 1.Pipeline and EvaluationFig. 2.