acceptodds
Withdrawn from ICLR 2027

Structured Trajectory Supervision for Reliable Physiological Time-Series Regression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Many regression targets are summary statistics of an unobserved structured trajectory: the scalar label is a reduction of a higher-resolution signal the model never sees. We ask whether estimating the trajectory as the model's representation—and reducing it to the scalar afterwards—is a beneficial form of dense structured supervision, and what can be learned about a model's behaviour from that representation. Using wrist photoplethysmography (PPG) heart-rate (HR) estimation as the testbed, our estimator, PPI-Net, predicts a 128-point pulse-to-pulse-interval trajectory with per-point uncertainty and derives HR by inverse-variance-weighted aggregation. On an identical backbone, replacing a scalar head by the trajectory head lowers participant-level MAE from 5.11 to 4.04 BPM under leave-one-subject-out evaluation (paired  BPM, better on 13/15 subjects, )—three times a re-run noise floor of 0.34 BPM that we measure and treat as the threshold of interpretability. The representation also supports mechanistic analysis that a scalar cannot: the signed error decomposes into shrinkage toward the cohort mean under subject-level shift (unbiased point 92 BPM, slope ), and the delta-method uncertainty inherits a factor that inverts its ranking in the high-HR stratum, which we derive and then verify. Under one corrected protocol, PPI-Net outperforms a standard 1D-CNN and a classical adaptive filter and is on par with the strongest estimator we could retrain from public code, BeliefPPG (3.64 vs. 3.830.20; ), with a complementary error profile; the shared failure reproduces in three held-out cohorts, and a calibration–test-separated confidence score turns it into selective regression (MAE at 50% coverage). We state explicitly what the evidence does not support, including a sensor ablation that reverses our own earlier conclusion.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
structured supervision, intermediate representations, time-series regression, selective prediction, uncertainty quantification, distribution shift, evaluation methodology, reproducibility, physiological signals, wearable sensing, heart rate estimation, photoplethysmography

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8zvHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposed to apply regression for a trajectory instead of a scalar. They propose PPI-Net and perform experiments on heart rate estimation. The authors compared with a strong estimator, Belief-PPG, and provide evidence that the proposed method performs similarly with the state-of-the-art technique with small improvements.

Critical strengths

The authors spent a good time to find the state-of-the-art techniques and compared with them. I appreciate for this effort. The idea is motivated well enough while the choosing application fits the motivation. The figures are easy to read and generally clear.

Critical weaknesses

The experiments are extremely shallow. The authors only considered one application, heart rate tracking, with a single dataset Dalia. There are multiple datasets out there including WildPPG (cited by authors). There are also several application where the authors can and should try their methods to provide evidence. Some examples are heart rate variability (HRV), respiration rate (RR), blood pressure and more. All of these tasks are regression based and well-defined for the technique. The current experiments are not enough for ICLR bar.

The presentation is weak and multiple references to Appendix are ?. This also shows the rigor of the work.

I believe the work needs a major revision.

Questions for authors

See above for the experiment point especially.

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

This paper targets wrist PPG heart rate estimation to study regression where scalar labels come from structured underlying trajectories. Instead of direct scalar regression, the model learns an ECG-derived PPI trajectory and aggregates these outputs for heart rate inference. It combines PPG, spectrogram, and accelerometer inputs and predicts per point heteroscedastic scales. On PPG‑DaLiA leave‑one‑subject‑out evaluation, trajectory supervision improves performance over a matching scalar head baseline.

Critical strengths
  1. Supervision on ECG‑derived PPI offers a natural intermediate representation, since these intervals directly define heart rate.
  2. When compared against the scalar‑head baseline with an identical model backbone, trajectory supervision reduces mean MAE from 5.11 to 4.04 BPM. Improvements are seen on 13 out of 15 participants.
  3. The work studies how the non‑linear mapping from PPI to heart rate reshapes uncertainty. It also highlights real‑world failure cases when relying on confidence‑based ranking.
Critical weaknesses
  1. The comparison against the scalar‑head baseline contains confounding factors beyond supervision. Though the encoder backbone is shared, the scalar baseline drops the trajectory‑refining BiGRU and pools across the full 64‑second window. In contrast, heart rate for the trajectory model is computed only from the central 8‑second predictions. Given these mismatches in how outputs are processed, it remains uncertain how much of the 1.07 BPM gain truly comes from structured trajectory supervision.
  2. Evaluation focuses almost entirely on final heart‑rate error, with no direct assessment of the predicted PPI trajectories. There are no reported metrics for point‑wise interval accuracy, nor any check of whether the model can track interval variations inside each window. Since heart rate comes from aggregating only the central segment of outputs, good heart‑rate performance does not guarantee that the underlying trajectory predictions are correct. It could be that the model only learns a window‑average interval or fits toward a smoothness prior, rather than recovering meaningful time‑varying PPI sequences.
  3. Primary quantitative results, ablations, and mechanistic analyses are all derived from PPG‑DaLiA, which contains only 15 participants. While additional datasets (WESAD, BAMI‑1) are mentioned, they only feed into a preliminary cohort‑shift check performed before the protocol corrections described in the main text. Only effect directions are interpreted here, without full LOSO evaluation. It is still unknown whether the proposed approach has good generalisation.
Questions for authors
  1. Does the trajectory approach keep its advantage when the baseline uses the same BiGRU and features corresponding to the central 8 seconds? Please keep the prediction head capacity and training settings as closely as possible.
  2. What are the point‑wise errors of predicted PPI trajectories relative to ECG‑derived references? How well do these trajectories track changes inside each window?
  3. Can the authors design experiments to disentangle performance gains coming purely from using PPI‑interval targets versus those arising from predicting a full time‑varying trajectory?
  4. Does the trajectory approach deliver better reliability than well‑matched scalar alternatives on challenging segments where motion artefact frequencies overlap with true heart‑rate frequencies?
  5. How can results obtained on a small dataset with only 15 participants support claims about the generalisability of the proposed method?
Reviewer_qA7fOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces PPI-Net, which reframes heart rate estimation from wrist-worn PPG sensors, traditionally a scalar regression task, as a PPI trajectory prediction problem. The model predicts a sequence of 128 PPI values ​​along with their point-wise uncertainties, subsequently deriving the heart rate via inverse-variance weighting. The authors argue that trajectory-based supervision provides richer physiological structural information and facilitates the analysis of model uncertainty and failure modes. In a leave-one-subject-out evaluation on the PPG-DaLiA dataset, PPI-Net achieves a MAE of 4.04 BPM, outperforming a scalar regression model built on the same backbone (5.11 BPM) and performing comparably to BeliefPPG. Further analysis reveals that primary errors stem from a regression toward the mean, where predictions for subjects with high heart rates tend to shift toward the training set average, and that the propagated uncertainty estimates may become unreliable at high heart rates. By applying a confidence-based filtering mechanism, grounded in signal quality and the degree of motion, the model can significantly reduce its MAE, albeit at the cost of lower coverage. In summary, the core contribution of this work lies in leveraging structured physiological trajectories to enhance the density of supervision signals, thereby enabling the interpretation of model errors and the implementation of selective prediction.

Critical strengths
  • Instead of directly regressing heart rate as a scalar value, the paper predicts the trajectory of beat-to-beat intervals, a representation with greater physiological significance, and derives the heart rate through weighted aggregation. This offers a novel modeling perspective for the regression of physiological time series.
  • Both the trajectory-based and scalar-based models utilize the same backbone network, with only the prediction head being swapped. This effectively isolates the benefits specifically attributable to trajectory-based supervision.
  • This paper is written in a very fluid, accessible manner, with a clear motivation.
Critical weaknesses
  • The core experiments rely primarily on a single, small-scale dataset (PPG-DaLiA). With a sample size of only 15 subjects, the findings are insufficient to support broad, generalizable conclusions.
  • The "regression to the mean" observed at high heart rates is primarily a descriptive finding. The paper notes that error increases as heart rate rises, attributing this to regression toward the population mean, but it has not experimentally verified the underlying causal mechanism.
  • The overall innovation of the paper lies largely in the combination of existing methods. The core architecture consists of standard components. The primary novelty stems from using PPI trajectories as a supervisory target and conducting analysis centered on this representation. Thus, the methodological contribution is oriented more toward the modeling framework and experimental analysis than toward the development of novel network architectures.
Questions for authors
  1. How do the authors justify the generalizability of their conclusions based primarily on PPG-DaLiA, which contains only 15 subjects? Can they provide results on additional independent datasets with larger and more diverse subject populations?
  2. What experiments support the claimed causal explanation that high-heart-rate errors are caused by regression toward the population mean?

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

Related papers

Discussion (0)

Sign in to comment.