acceptodds
Under review as a conference paper at ICLR 2027

OPLR: Alleviating Normality Bias in PPG-to-ABP Reconstruction via Ordered Prototype Latent Refinement

Abstract

Continuous arterial blood pressure (ABP) waveforms enable high-resolution hemodynamic monitoring but typically require invasive arterial lines. PPG-to-ABP reconstruction offers a noninvasive alternative, yet long-tailed BP distributions induce a systematic failure we term Normality Bias: predictions drift toward the normotensive mean, attenuating hypertensive peaks and elevating hypotensive troughs even when aggregate metrics appear strong. We formalize this bias through a stratified signed-error diagnostic and introduce two model-agnostic metrics: the Normality Bias Index (NBI), which captures monotonic regime-dependent drift, and TailBias, which quantifies bias at the clinically critical extremes. Across architecture families, these diagnostics reveal that normality bias is pervasive. We propose Ordered Prototype Latent Refinement (OPLR), a representation-level remedy that imposes an ordered regime geometry (Low < Normal < High) on the latent space and performs routing-guided prototype–token refinement before decoding. This design targets three coupled failure modes of normality bias: regime absorption (rare-regime segments encoded near the dominant Normal mode), routing monopoly (routing collapsing onto the Normal prototype), and attenuation of extreme waveform amplitudes. Under strict subject-wise, calibration-free evaluation on PulseDB and zero-shot transfer to two additional cohorts, OPLR achieves the lowest NBI and TailBias among all compared models and the lowest RMSE and MAE in every regime, including the aggregate, with gains significant under subject-level paired tests on PulseDB. The diagnostics are robust to binning choices, and the regime-level results are stable under shifts of the reporting boundaries. Matched-backbone ablations isolate the ordered prototype as the component that moves the bias metrics, show that the advantage is not explained by additional inputs alone, and find that data- and loss-level long-tailed remedies on the same backbone reduce it only marginally.

Primary area:
learning on time series and dynamical systems
Keywords:
long-tailed learning, physiological waveform reconstruction, prototype-based latent refinement, PPG-to-ABP, blood pressure estimation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_FZvHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors uncover a widespread failure mode in PPG-to-ABP models that they term "normality bias", which is driven by training data imbalances that skew predictions towards normotensive values. To evaluate this phenomenon, the authors develop NBI and TailBias as robust scalar diagnostics. Their proposed solution, OPLR, imposes an ordered geometric structure on the latent space (Low < Normal < High) to decouple rare regimes from the dominant normal mode before decoding. Experimental results confirm that OPLR surpasses standard fixes and delivers impressive performance.

Critical strengths
  • Mean tracking is a rampant issue in noninvasive blood pressure monitoring work, so it is good to see the authors addressing this issue head-on. I especially like the proposed metrics (Normality Bias Index and TailBias); I would consider using them in my own research.
  • The construction of the objective function is clear and intuitive
  • The appendix is almost entirely comprehensive, covering deeper ablations and the sensitivity of various hyperparameters in the method. The appendix also provides further evidence demonstrating the benefit of the proposed approach (e.g., Figures 4 and 5)
Critical weaknesses
  • As the authors already note, their external validation sets are modest in comparison to their training sets: 30 subjects from MIMIC-IV and 4 subjects from the Queensland dataset after preprocessing. However, this supplements the subject-wise calibration-free cross-validation they performed on PulseDB, which itself is a mixture of MIMIC-III and VitalDB.
  • Close inspection of Figure 4 reveals that OPLR slightly prioritizes the Low prototype when the true SBP is just above the hypotensive boundary (90-93 mmHg). Although the difference in routing probabilities is slight and the smoothness of the routing probabilities is ideal for regression, this could hint at a bias towards false alarms for hypotension. One could argue that smoothness actually complicates predictions at the boundaries, whereas stronger routing could pull samples towards the correct prototype for the sake of improved alarm delivery. Therefore, while I agree that regression is the right approach to noninvasive blood pressure monitoring, I would like to see some classification metrics as well (i.e., thresholding the continuous outputs, not retraining a classification head).
Questions for authors
  • In Table 11, the width of the Medium regime always remains 50 mmHg. Thus, it is always ~4x the size of the other two classes combined. Is there a reason not to try a configuration with closer boundaries (e.g., 40 mmHg apart)? As far as I understand, OPLR is not beholden to a specific degree of class imbalance, and the boundaries could be adjusted slightly while preserving clinical utility. I would just be curious to see how increasing the representation of Low and High cases relative to the Medium regime changes performance.
  • Section A.4 states, "The requirements for OPLR are an ordered target variable, a heavy-tailed distribution whose extremes are the cases that matter, and a sequence output whose local morphology must be preserved. Other physiological waveform tasks meet all three." I’m trying to think of when this is actually true. For blood chemistry quantities like glucose, detecting that they fall within a range is important, but the morphology of the signal is not. For physiological waveforms like ECG, morphology matters, but extreme values do not hold special meaning. Could the authors offer some examples in the Appendix?
3 comments

Thank you for the careful reading. Page numbers refer to the revised PDF.

W1. External cohorts. Revision: Section 8, "External evidence" (p. 10).

Agreed. Limitations keeps the cohort-size and external-validation qualifications and now states that evaluation in additional independent cohorts remains necessary.

W2. Boundary alarms and classification metrics. Revision: new Appendix L and Table 14 (pp. 29–30); Section 8, "Residual hypotensive bias" (p. 10); Figure 4 caption (p. 21).

As you suggested, we thresholded the continuous outputs, with no classifier retrained and no threshold tuned: SBP from each reconstructed waveform and the reference SBP are both thresholded at 90 mmHg (91,834 PulseDB test segments with a valid SBP for all models).

ModelSensitivity (%)Specificity (%)F1AUROC
OPLR3.2799.88.060.793
U²-Net1.7599.72.031.757
LSCT0.9999.60.017.716
PPG2ABP2.2399.82.040.733
ArterialNet2.9498.83.038.664
T2T-GAN0.7699.88.014.592

For reference SBP in [90, 93) mmHg, the interval you raised, OPLR produces 10 false Low alarms among 1,915 segments (0.52%); the lowest count among the baselines is 11. The slight Low-prototype preference in Figure 4 therefore does not produce more false alarms, and the Figure 4 caption now notes that routing probabilities are latent preferences that do not themselves define alarms. OPLR has the highest sensitivity, F1, and AUROC, but sensitivity at the fixed threshold is at most 3.3% for all six models, which Limitations now states. For hypertension (>140 mmHg), AUROC is 0.782 for OPLR against 0.756 for the best baseline.

On stronger routing: routing is soft by design. The routing log-prior is clamped from below and the token update is residual, so a misrouted segment receives an additive correction instead of having its content replaced (Section 4, "Prototypes → Tokens", p. 6). Sharper routing would raise the cost of a wrong route where the regime is most uncertain, near 90 and 140 mmHg. We did not train a sharper-routing variant, so this is a design rationale, not an empirical comparison.

Q1. A 40-mmHg Normal interval. Revision: Table 11, blocks "40-mmHg normotensive interval" and "Retrained with 95/135 mmHg anchors" (p. 27); Appendix I, last two paragraphs (p. 26).

The default boundaries are clinical anchors for hypotension and hypertension. We added three 40-mmHg intervals and a model retrained with narrower anchors.

High / Low boundariesHigh (% of test set)High RMSE/MAELow (% of test set)Low RMSE/MAE
>140 / <90 (default)18.214.47/12.122.312.89/10.78
>135 / <9524.913.69/11.476.012.07/10.29
>130 / <9032.613.03/10.922.312.89/10.78
>140 / <10018.214.47/12.1211.911.72/10.08
>135 / <95, retrained24.913.23/10.936.08.89/7.14

The first four rows re-partition fixed predictions: moving a boundary inward lowers that regime's error, as the group admits more near-normal segments. The last row is OPLR retrained with 95 and 135 mmHg as the training anchors (training proportions 6.4% / 70.4% / 23.2% for Low / Normal / High) and evaluated at those boundaries. At the same >135 / <95 boundaries, retraining lowers Low RMSE from 12.07 to 8.89 and High RMSE from 13.69 to 13.23, so OPLR benefits when the tails are better represented in training.

Q2. Other tasks. Revision: Appendix A.4 (p. 16).

We agree that the original statement was too broad and have narrowed it. As usually formulated, glucose and ECG tasks do not meet all three requirements, for the reasons you give. A task that plausibly does is intracranial-pressure waveform reconstruction from paired physiological signals (Megjhani et al., 2023): pressure is ordered, elevated pressure is the clinically decisive range, and pulse morphology is informative. Whether its extremes are underrepresented depends on the cohort; we validate only PPG-to-ABP.

Thank you for the responses to my queries. I remain positive in my assessment of this paper

Dear Reviewer FZvH,

Thank you for your time and your constructive suggestions; they made the paper stronger. We appreciate your support.

Best regards, The authors

Reviewer_9wvwOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This is a strong manuscript that highlights the problem of normality bias and proposes a practical solution based on imposing an ordered geometry. Although the primary use case is PPG-to-ABP reconstruction, the method has potentially broader applications. The manuscript is also clearly written and easy to follow, and the experiments effectively demonstrate the utility of the proposed approach. My only substantive suggestion is that several of the most interesting methodological details are relegated to the appendices. The paper would therefore benefit from moving a selection of these methodological innovations into the main text, space permitting.

Critical strengths

Normality bias is an important failure mode, which the authors tackle using an interesting approach.

The methods are well described and the approach is novel.

The application to ABP reconstruction from PPG is well taken as normality bias is an important failure mode for this particular task - espcially given that most training data sets are (PulseDB) are heavily biased to normal BP measurements.

Critical weaknesses

The manuscript does not include a dedicated related-work section; instead, prior work is addressed only briefly in a single paragraph. Because much of the relevant literature concerns machine-learning robustness, this work would be better positioned if the authors discussed and cited related work in fairness and robustness, including group distributionally robust optimization (DRO). Methods that explicitly impose order or structure on embedding spaces, such as deep metric learning, are also directly relevant and should be discussed.

Many of the technical details are relegated to the appendices, making portions of the manuscript difficult to follow. Most notably, the loss function in Equation 9 does not fully convey the complexity of the final optimization problem. Equation 20 provides a clearer and more informative formulation because it explicitly displays the three cost-function hyperparameters. The authors should consider presenting this formulation, or an equivalent level of detail, in the main text.

The proposed approach includes a substantial number of hyperparameters—nine are listed in Table 5 of Appendix C. In light of this, it is unclear whether a comprehensive hyperparameter search was feasible. The authors should therefore describe the search strategy, parameter ranges, selection criterion, and procedure used to determine the final set of hyperparameters.

A central challenge in noninvasive ABP estimation is accurately tracking changes in an individual person’s blood pressure over time. This task is distinct from reproducing population-level blood pressure statistics and is generally more demanding. Although addressing longitudinal, patient-specific performance may be beyond the scope of the present study, the distinction is clinically important and should be acknowledged in the Discussion or Limitations section.

The top panel of Figure 2 is not particularly informative in its current form, and the accompanying discussion does not resolve the ambiguity. Although the token-to-prototype formalism is clear, the role of the soft-routing distribution in the method remains insufficiently explained. In particular, because this distribution may preserve biases in the training data—for example, the predominance of normal blood-pressure waveforms in PulseDB—it is unclear why it provides a useful structural prior. The authors should explain how the soft-routing distribution is constructed and applied, and why it is expected to mitigate rather than perpetuate the underlying normality bias.

Questions for authors

I am having trouble seeing specifically how the soft routing distribution is used - as I mentioned above. Can the authors clarify?

Can the authors clarify the procedure for hyper-parameter tuning?

Please include a prior work section, making sure to mention (and cite) prior work in the robustness space.

Can the authors clarify Figure 2? I do not think it helps to help the reader understand what the authors are doing.

3 comments

Thank you for the careful reading. Page numbers refer to the revised PDF.

W1/Q3. Related work. Revision: new Section 2, "Related Work" (pp. 2–3); comparison in Appendix K, Table 13 (p. 29).

The new section covers fairness and distributionally robust learning, including group DRO (Hashimoto et al., 2018; Sagawa et al., 2020), deep metric learning (Schroff et al., 2015; Kim et al., 2020), and continuous-target methods. Group DRO changes how much each group's risk counts, whereas OPLR restructures the representation that the decoder consumes. Proxy-based metric learning treats class proxies as unordered, whereas OPLR's prototypes are ordered and trained with a margin that grows with ordinal distance. Five continuous-target methods are now compared on our backbone: their NBI is 0.355–0.378, against 0.148 for OPLR.

W2. Objective. Revision: Eq. 9 in Section 4, "Failure-Mode-Aligned Objective" (p. 6); term definitions in Appendix C (pp. 18–19).

Eq. 9 is now the weighted objective that was previously Eq. 20:

L = L_time + L_freq + λ_ord L_ord + λ_dist L_dist + λ_peak L_peak, with λ_ord = 0.2, λ_dist = 0.1, and λ_peak = 0.5.

The target of L_dist, q = (0.025, 0.806, 0.169) for Low, Normal, and High, is now also stated in the main text ("Ordered Supervision and Anti-Monopoly Regularization", p. 5).

W3/Q2. Hyperparameter tuning. Revision: Appendix C.4, "Tuning procedure" (p. 19); Table 5, new "Range tried" column (p. 20).

Tuning was manual and iterative, with no automated or joint search. Settings were compared by the loss on the PulseDB validation split, and within each run we kept the checkpoint with the lowest validation loss. The final values are fixed across experiments and cohorts, except where an ablation varies one of them. Table 5 now lists the values tried (e.g., λ_ord and λ_dist 0.05–0.5, m 0.2–0.38), and Table 3 (p. 9) gives the two extremes m = 0 (NBI 0.369) and λ_dist = 0 (0.317), against 0.148 for the full model.

W4. Within-person tracking. Revision: Section 8, "Scope" (p. 10).

Limitations now distinguishes segment reconstruction from tracking an individual's blood pressure over time, which requires temporally linked recordings and is not evaluated here.

W5/Q1/Q4. Routing and Figure 2. Revision: Figure 2 and its caption (p. 5); Section 4, "Ordered Prototype Routing" (pp. 4–5), "Ordered Supervision and Anti-Monopoly Regularization" (p. 5), and "Prototypes → Tokens" (p. 6); Appendix D.1 and Figure 4 (pp. 20–21).

Construction and use. Each segment's pooled tokens give p = softmax(cos(g, r_k)/τ). Its softened form p_cond is used in two places: its log, clamped from below, biases prototype-to-token attention, and it weights the prototype summary Z_sum that conditions the decoder. During training a scheduled mixture with the regime label precedes softening; validation and inference use the predicted routing only. Figure 2 now draws this as one colored path and shows q only in the training-only box.

Why it does not perpetuate the bias. The routing target is each segment's own regime label, not the class prior, so a hypotensive segment is pushed toward the Low prototype however rare Low is. The prior q acts only on the batch mean of the routing, and because q equals the training regime frequencies, routing every segment to its own regime gives a batch mean of q in expectation, so the two terms are compatible. Empirically, removing the ordinal margin returns NBI to the host level (0.369; Table 3), a uniform q does not help (0.182 against 0.148; Appendix D.1), and realized routing is far more even than q: in Figure 4 the mean Normal probability stays below 0.5 in every SBP bin (maximum 0.450), although q_N = 0.806.

Dear Reviewer 9wvw,

Thank you again for your review. As the discussion period is closing soon, we would like to check whether our response addresses your points. In brief:

  • Main-text detail: the routing mechanism, the weighted objective (Eq. 9), and the prior q are now in the main text, following your main suggestion.
  • Related work: a dedicated section now covers fairness, group DRO, and deep metric learning.
  • Hyperparameters: Appendix C.4 and Table 5 now give the tuning procedure, the ranges tried, and the selection criterion (validation loss).
  • Within-person tracking: Limitations now distinguishes segment reconstruction from tracking an individual over time.
  • Figure 2 and routing: the figure is redrawn, and Section 4 explains how routing is constructed, where it is used, and why it does not perpetuate the bias.

If anything remains unclear, we would be glad to answer before the deadline.

Best regards, The authors

I thank the authors for clarification. While I support the work, the lack of any analysis for predicting patient-specific BP tracking remains a deficiency, hence I will retain my score.

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

The authors identify Normality Bias as a major problem in ABP prediction using PPG signals. They propose a smart architectural innovation to overcome this issue. They extract latent tokens from the PPG waveform. They are routed through ordered prototypes and refined for their bias. The refined outputs are then passed to the decoder, along with additional conditioning from the prototype summary. The authors show that this yields significantly greater RMSE gains, especially in the lower and higher ABP regimes.

Critical strengths
  1. The authors identify a very important problem in physiological signals: Normality Bias.
  2. The authors are also very transparent about their shortcomings regarding potential data leakage.
Critical weaknesses
  1. While the paper is positioned for ABP prediction using PPG, DIR baselines are also widely applicable here. Thus, the authors must compare against the works cited in Appendix A.3, specifically those that target feature/representation DIR, such as FDS, RankSim, Rank-N-Contrast, ConR, or Balanced MSE.
  2. The ablation studies are incomplete. We only add the four components to the backbone and also remove two components from the full. What is required is an ablation study where one component is sequentially added to the next. For e.g., you should have a table comparing “Bare backbone”, “Bare backbone+L_Peak”, “Bare backbone+L_Peak+Decoder conditioning”, “Bare backbone+L_Peak+Decoder conditioning+Freq Enc.”, and Full. This is more informative than the current Table 3.
  3. The paper targets a very specific use case. I wonder whether the authors have looked at related problems in DIR and whether their proposed ideas work in those domains as well?
Questions for authors

Relative improvement in Low BP regimes still underperforms compared to High BP regimes. Is there any reason why that happens?

4 comments

Thank you for the constructive suggestions. Page numbers refer to the revised PDF.

W1. DIR baselines. Revision: Appendix K, "Imbalanced-regression methods for continuous targets" and Table 13 (pp. 28–29); summary in Section 7, "Imbalanced-regression methods" (p. 10).

We added FDS, RankSim, Rank-N-Contrast (RnC), ConR, and Balanced MSE (BMC) on our backbone with OPLR removed (same split and seed). FDS, RankSim, and ConR follow their official procedures. RnC and BMC use our training recipe: RnC is trained jointly on a single view, and BMC uses a smaller noise-scale initialization because the official one did not reach a competitive solution (Overall RMSE 35.98). The host is the same backbone under our recipe, trained with a schedule shortened to reduce training time. Because the official procedures differ from our recipe, the FDS, RankSim, and ConR rows differ from the host in more than the added method; all settings are in Appendix K.

MethodOverall RMSEHigh RMSELow RMSENBITailBias
Host12.8617.1817.000.36715.84
FDS12.5318.0215.350.35515.14
RankSim12.2519.4413.290.37816.11
ConR12.3918.7814.050.37015.83
BMC12.9418.9914.740.36615.75
RnC12.6618.3014.710.36215.60
OPLR11.1414.4712.890.1486.63

All five leave the bias metrics near the host level (NBI 0.355–0.378, TailBias 15.14–16.11, against 0.148 and 6.63 for OPLR), and their best Low and High RMSE (13.29 and 18.02) remain above OPLR's.

W2. Sequential ablation. Revision: Table 3 and Section 7, "Ablation and component analysis" (p. 9); single additions moved to Table 10 (p. 26); Appendix H, "Seeds" and "Training settings" (p. 25).

Table 3 now follows your proposed order; the fourth row is the host of W1.

ConfigurationOverall RMSEHigh RMSELow RMSENBITailBias
Bare12.1717.0316.370.36615.86
+ L_peak12.5118.1815.600.36915.88
+ Conditioning12.6117.8015.230.36115.66
+ Frequency encoder12.8617.1817.000.36715.84
Full OPLR11.1414.4712.890.1486.63

The first three additions leave NBI at 0.361–0.369 and TailBias at 15.66–15.88; both move only when the ordered prototype is added. The prototype alone gives NBI 0.339 (Table 10), so the components are complementary. With seeds 43 and 44, NBI is 0.365 and 0.372 for the host and 0.093 and 0.154 for the full model. "Conditioning" denotes the metadata and PPG-statistic inputs with their FiLM pathways.

W3. Broader applicability. Revision: Section 8, "Scope" (p. 10); Appendix A.4 (p. 16).

We checked a standard DIR benchmark for this response (to be added to the appendix). With the definitions of Section 3 unchanged, we computed NBI and TailBias on IMDB-WIKI-DIR (Yang et al., 2021) for the three ResNet-50 checkpoints released by the benchmark authors (inference only; released MAEs reproduced).

ModelMAENBI [95% CI]TailBias (years)
LDS7.870.282 [0.259, 0.305]14.06
FDS7.660.287 [0.266, 0.309]14.18
LDS + FDS7.680.282 [0.261, 0.305]13.97

The same drift is present there: signed error falls monotonically from about +14 years at ages 0–10 to about −14 at 90–100 and crosses zero at 30–40, the most populated training decade, even though these models were trained with established DIR remedies. OPLR is built for PPG-to-ABP: it refines a token sequence, conditions a waveform decoder, and uses an extrema loss, none of which carries over directly to a scalar output. A proper transfer requires adapting these parts and training the encoder with them, which is a natural next experiment; we make no claim of transfer here and state this as scope. Among waveform tasks, intracranial-pressure waveform reconstruction is a candidate (Megjhani et al., 2023).

Q1. Low versus High. Revision: Section 8, "Residual hypotensive bias" (p. 10).

Against the same backbone without OPLR, the two gains are comparable: over the two runs of that backbone (host in Table 13; Table 12, p. 29), OPLR lowers Low RMSE by 16–24% and High RMSE by 16–20%. Against the strongest baseline, the margin is smaller in Low (about 3%; U²-Net) than in High (about 18%; PPG2ABP) because U²-Net is unusually strong in Low; against the other four baselines, OPLR's Low RMSE is 21–55% lower. Low also has 6.8 times fewer training segments than High, and a residual hypotensive bias remains; Limitations now states both.

Dear Reviewer PqCh,

Thank you again for your review. As the discussion period is closing soon, we would like to check whether our response addresses your concerns. In brief:

  • W1: all five DIR methods you named are now compared on our backbone (Table 13); their NBI stays at 0.355–0.378, against 0.148 for OPLR.
  • W2: Table 3 now follows the sequential order you proposed, with two additional seeds.
  • W3: we measured the same bias on a standard DIR benchmark (IMDB-WIKI-DIR, NBI 0.28).
  • Question: against the same backbone, the Low and High gains are comparable (16–24% and 16–20%); the smaller margin over the strongest baseline in Low arises because U²-Net is unusually strong there, and Low has 6.8 times fewer training segments.

If anything remains unclear, we would be glad to answer before the deadline.

Best regards, The authors

Thank you for the detailed responses!

My concern about scope remains. Your response (and Appendix A.4) confirms that OLPR is validated on a single task and is not expected to transfer without redesign. I will weigh this in my final assessment.

From my understanding of the sequential ablation, the performance improves only when the ordered prototype is added. Otherwise, the other components lead to a higher overall RMSE (and a similar NBI to the bare model) when added. Could you please clarify why this occurs?

Thank you for the follow-up.

Sequential ablation. The pattern becomes clear when the two tails are read together. Without the ordered prototype, each addition trades one tail for the other: added alone to the bare backbone, every component lowers Low RMSE and raises High RMSE, while NBI and TailBias do not move (Table 10; TailBias 15.76–15.93). With the original schedule, the three together behave the same way. All rows below use that schedule:

Added to bare backboneOverall RMSELow RMSEHigh RMSENBI
nothing12.1716.3717.030.366
L_peak12.5115.6018.180.369
conditioning12.4514.7818.580.363
frequency encoder12.3115.0118.600.370
all three, no prototype (Table 12)12.7515.3518.170.36
ordered prototype alone12.1813.8717.830.339
all together (Full OPLR)11.1412.8914.470.148

This is a transfer of error between the tails, not a reduction of the regime-dependent drift, which is why NBI stays at 0.36–0.37. For the single additions it also accounts for more than half of the rise in Overall RMSE: High segments are 18.2% of the test set and Low segments 2.3%, so the loss in High outweighs the gain in Low. The fourth row of Table 3 is the same three-component configuration trained with a shortened schedule; there Low RMSE is 17.00 and High 17.18, so the tail values of this configuration vary between runs, whereas NBI does not (0.367 and 0.36; 0.365 and 0.372 with two further seeds).

The prototype alone is the only single addition that moves the bias metrics (NBI 0.339, TailBias 13.65), and it gives the largest Low gain, but it does not improve High. Among the rows above, only the full model improves both tails at once (Low 16.37 to 12.89, High 17.03 to 14.47), and that is where NBI falls to 0.148 and TailBias to 6.63. The other components are therefore not redundant: of the total NBI reduction (0.218), only 0.027 appears with the prototype alone, and the remaining 0.191 appears only when they are present. Our reading is that L_peak and the conditioning path give the model the means to change amplitude, and the ordered prototype gives it the regime information that determines the direction; without that information the change goes in the same direction regardless of regime, which is the trade above. The ordering itself is necessary: without the ordinal margin, the full model's NBI returns to 0.369 (Table 3, bottom).

Scope. We agree that OPLR is validated on a single task, and we would like to explain why it is scoped this way. Medical data carry well-defined patient characteristics and modality-specific signal properties, and models for them commonly exploit these properties. Such task-specific methods are typically evaluated on the target task across cohorts, whereas evaluation across domains is the standard for general-purpose methods; for example, Yang et al. (2021) evaluated LDS and FDS on five benchmarks spanning images, text, and polysomnography.

OPLR belongs to the first group. It is built around properties that PPG-to-ABP has: both signals are pulsatile waveforms driven by the same cardiac cycle, and the ordered variable (SBP) is the peak of the output waveform itself, so regime, pulse amplitude, and morphology are tied together. The regime anchors (the clinical thresholds of 90 and 140 mmHg), the waveform decoder, and the extrema loss all rely on this link. A task without such a waveform output would need these parts redesigned, and a waveform task with the same structure (Appendix A.4) would still need its own validation; both are separate studies that were not feasible within the rebuttal period.

Our claims are therefore limited to PPG-to-ABP, validated on three cohorts. The part that is not task-specific, the NBI and TailBias diagnostics, is also shown on IMDB-WIKI-DIR. We would be grateful if the paper could be assessed on the basis of these claims, and we appreciate your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.