Low-latency HRV analysis from ultra-short ECG windows using a modular deep-learning framework

Figures

• 

  Fig. 1. Training and data preparation workflow, including dataset inputs, preprocessing, and supervision target construction.

• 

  Fig. 2. Inference/deployment workflow: preprocessing, frozen encoder, discriminator gating, and conditional RMSSD estimation.

• 

  Fig. 3. Graph showcasing compiled model architecture with a single metric head used as an example. Numbers denote exemplary hyperparameters.

• 

  Fig. 4. Representative ECG windows from the datasets used in this study. Top: LUDB (clinical 12-lead). Middle: PTB-XL (clinical 12-lead). Bottom: Apple Watch (single-lead consumer device). Each panel shows a single window after per-window min-max normalization to [0, 1] (as defined in Eq. (3)).

• 

  Fig. 5. Bland-Altman agreement plots for RMSSD on the combined test set. Left column shows the full (unclipped) scale; right column shows a fixed clipped view for easier visual comparison (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x\in [0,500]$$\end{document} ms and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y\in [-300,600]$$\end{document} ms). Solid lines indicate bias, and dashed lines indicate 95% limits of agreement (bias \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm 1.96\,\textrm{SD}$$\end{document}). Baseline methods are evaluated on the subset of valid windows where they return a finite RMSSD estimate.

• 

  Fig. 6. Examples of incorrect rejections by the discriminator. Both ECG windows were flagged as low-quality (non-processable) by the discriminator, yet they remain visually interpretable and contain identifiable QRS complexes, indicating that the gate can be overly conservative in borderline cases.