acceptodds
Under review as a conference paper at ICLR 2027

BeatFlow-ECG: Rectified Flow for ECG Reconstruction from Indirect Wearable Signals

112 views

Abstract

Continuous cardiac monitoring outside clinical settings requires signals that are both informative and practical to collect during daily life. Electrocardiography (ECG) provides rich information about cardiac rhythm and waveform morphology, while wearable photoplethysmography (PPG) is easier to acquire continuously but is only an indirect cardiovascular measurement and is highly sensitive to motion. We present BeatFlow-ECG, a conditional rectified-flow model for reconstructing single-channel ECG from synchronized PPG and inertial measurements. BeatFlow-ECG models reconstruction as conditional transport from noise to ECG using a convolutional encoder-decoder with a transformer bottleneck and explicit flow-time conditioning. Motion information is incorporated through IMU-derived conditioning features, motion-dependent loss weighting, and an easy-to-hard training curriculum. We evaluate the model under leave-one-subject-out protocols on PPG-DaLiA and WESAD. BeatFlow-ECG achieves the best results among the evaluated deterministic, adversarial, and diffusion-based baselines across all reported waveform and beat-timing metrics, with Pearson correlations of 0.983 and 0.986 and R-peak F1 scores of 0.946 and 0.955, respectively. Compared with Conditional DDPM-1D, L1 error decreases from 0.085 to 0.062 on PPG-DaLiA and from 0.074 to 0.055 on WESAD. Additional analyses on PPG-DaLiA show higher correlation in fixed R-peak-relative waveform regions and lower reconstruction error across low-, medium-, and high-motion subsets.

Primary area:
learning on time series and dynamical systems
Keywords:
ECG reconstruction, photoplethysmography, wearable sensing, rectified flow, flow matching, physiological signal processing, multimodal learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_mNMTOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BeatFlow-ECG, a conditional rectified-flow model for reconstructing single-channel ECG from synchronized PPG and inertial signals. The framework combines a convolutional encoder-decoder, Transformer bottleneck, flow-time conditioning, motion-aware training, and an STFT loss. Evaluation on PPG-DaLiA and WESAD under LOSO protocols shows improved waveform and beat-timing performance over the considered baselines.

Critical strengths

The paper addresses a relevant wearable cardiac-monitoring problem and presents a coherent conditional flow-based framework. The evaluation is reasonably comprehensive, including waveform metrics, R-peak/heart-rate metrics, regional P/QRS/T correlations, motion-stratified analysis, and ablations.

Critical weaknesses
  1. The scope of “ECG reconstruction” is unclear. Each 10-s ECG target is independently z-score normalized, which removes absolute baseline and amplitude information. Therefore, the model mainly reconstructs normalized morphology and timing rather than the physical ECG waveform. Pearson correlation and R-peak F1 also do not directly assess peak-amplitude fidelity. The manuscript should either restrict its claims accordingly or provide amplitude-sensitive evaluation.
  2. The contribution of motion-aware training is not fully isolated. The ablation removes motion weighting and STFT loss together, so the specific effect of motion-dependent loss weighting cannot be determined. In addition, LOSO results are reported without clear subject-level variability or confidence intervals, making it difficult to assess consistency across subjects.
Questions for authors
  1. Is absolute ECG amplitude reconstruction outside the intended scope of the method, given the per-window z-score normalization? Are any target ECG statistics used during inference or inverse normalization?

  2. Are the P/QRS/T regional correlations computed using ground-truth or predicted R peaks?

  3. Can the authors provide an ablation removing only motion-dependent loss weighting while retaining the STFT loss?

  4. Can subject-level variability or confidence intervals across LOSO folds be reported?

  5. There are some abbreviations without any explanation

  6. In the results section, the authors only present the comparison results between baseline models and the proposed model using PPG-DaLiA. What about the WESAD dataset?

  7. There are some details in the network structure not clearly mentioned, like AdaRMSNorm, how is it implemented in the current model?

Reviewer_6n9eOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work adapts the rectified flow generative framework to wearable biosignal reconstruction, presenting BeatFlow-ECG, a model that reconstructs single-lead ECG from wrist PPG paired with IMU motion data. The whole design is centered on fighting motion artifacts: it adds IMU-derived motion conditioning features, uses motion-weighted loss, and adopts an easy-to-hard training curriculum across motion levels. Evaluated under leave-one-subject-out (LOSO) protocol on two public datasets, it achieves better waveform and beat-timing accuracy than diffusion models and classic deterministic baselines, and runs notably faster than diffusion-based methods.

Critical strengths

(1) Really nails the core pain point of wearable PPG. Motion artifacts wreck signal quality, and they didn’t just slap IMU data on as an extra feature. They built motion adaptation into every step of the pipeline, i.e., input conditioning, loss weighting, even the training curriculum. The whole logic is end-to-end consistent. (2) Evaluation is thorough and done properly. LOSO testing instead of random train/test splits actually measures generalization to new users, which is what matters for wearable models. They test against a solid range of baselines and add on morphology-sensitive analysis, motion-stratified tests, component ablations, and inference speed benchmarks. Pretty much covers the checks you’d want to see.

Critical weaknesses

(1) The rectified flow backbone and conv+transformer architecture are all borrowed from prior work; all new additions are engineering tweaks for motion adaptation, with no theoretical advance to the generative model itself. (2) CardioGAN uses PPG-only input while this model uses PPG+IMU, giving it an inherent information advantage. No IMU-augmented deterministic baseline is provided for a controlled comparison. (3) Tested only on healthy subjects, with zero evaluation on pathological ECG patterns. Single-lead ECG reconstruction also has very limited clinical utility. (4) 120 iterative steps are still dozens of times slower than single-pass deterministic models, which is problematic for low-power real-time wearables. (5) Only whole-component removals are tested, with no performance breakdown across motion intensities or deep dive into failure modes.

Questions for authors

(1) Do you plan to validate on pathological ECG data? Right now results are only on healthy subjects, which limits clinical relevance. (2) Could you add a deterministic baseline that also includes IMU features? Right now CardioGAN uses PPG only, so the comparison is not controlled. It’s hard to tell if the gain comes from multi-modal input or from the rectified flow formulation itself. (3) Have you tried reducing the number of inference steps? For example, what happens to accuracy if you drop to 60 steps? Real-time performance and power consumption are very critical for wearable devices.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.