acceptodds
Under review as a conference paper at ICLR 2027

Long-Horizon Physiological Analysis via Native Calibration of a Multimodal Waveform Foundation Model

Abstract

Bedside monitors record PPG and ECG waveforms continuously, but laboratory values such as hemoglobin and potassium are measured only when blood is drawn: hours apart in intensive care, months apart between emergency-department visits. We study long-horizon physiological analysis: estimating from the waveforms how a lab value has changed since the previous draw. Our method, native calibration, uses the patient's own history: each earlier draw provides a measured value and a waveform recorded at the same time. It estimates the current value as an earlier measured value plus the change read from the current and past waveforms, both embedded by a frozen waveform foundation model (FM). We also introduce a new multimodal FM, a transformer pretrained without labels to relate ECG and PPG and to match a patient's PPG across time. We evaluate on two kinds of task: 1) tracking lab changes, namely hemoglobin between emergency visits and potassium within intensive-care stays; and 2) reading values from the current waveform alone, namely first-draw emergency labs, intra-operative haemodynamics and 24 standard PPG benchmark tasks. On held-out patients, native calibration improves the estimated change in both hemoglobin and potassium over strong non-waveform baselines, and part of this gain is specific to the patient: it shrinks when the patient's past waveforms are replaced by those of other patients with the same lab values. Our FM achieves state-of-the-art results on first-draw labs and the standard PPG tasks, ranking first among ten encoders, and is competitive with the best public encoder on intra-operative haemodynamics.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
PPG, Physiological Foundation Models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_zzirOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a framework to predict a patient’s lab value changes between blood draws using bedside PPG and ECG waveforms. The proposed method uses the patient’s previously measured value and reads the change from their current and past waveforms. The authors also introduce a new multimodal waveform foundation model. On held-out patients, the method modestly improves over strong non-waveform baselines, partly because it uses the patient’s own past waveforms. The new encoder ranks first on average among ten encoders, slightly ahead of the best public model.

Critical strengths
  1. Most waveform foundation-model work reads a value from a single window of seconds to minutes. This paper instead asks how waveforms can track a patient’s physiology across the gaps between blood draws, from hours in the ICU to months between ED visits. This long-horizon setting is clinically relevant and under-explored.

  2. The paper is transparent about pretraining overlap and reproducibility limitations.

Critical weaknesses
  1. The donor control overstates the patient-specific gain. From Table 4, the donor arm gains less over the floor than the network that sees only the current window. A donor’s past windows are probably acting as noise. Please add a donor − current column with CIs, report own − current as the primary contrast, cap donor reuse.

  2. The stage-1 checkpoint lacks details. The authors mentioned Stage 1 is “a checkpoint … that our group pretrained earlier”. Section K states it “has no identifiers that join to MIMIC-III or MC-MED”, so its overlap with the evaluation patients cannot be determined. This is a reproducibility gap and a risk of label-free transductive leakage that would directly favour the own-past contrast. Please properly cite the previous work in third person so that its details and pretraining corpus can be determined.

  3. The comparison with public encoders may not be fair. First of all, the public encoders were pre-trained on some of the evaluation datasets used. The pretraining exposure is uneven. Secondly, 518 of 528 ED test patients (98%) may appear in the stage-1 corpus of the proposed model. In the ED task, the four encoders that may have seen MC-MED have the four largest own − donor contrasts (+0.025 to +0.038). Every unexposed encoder is at most +0.012. Comparing against models without this access is not fair, and a Limitations note is not sufficient. Please retrain stage 1 on a split or dataset that excludes evaluation patients, or restrict the claims to the ICU task, which has no overlap.

  4. The encoder’s main contribution is joint ECG–PPG pretraining, yet every comparator is a single-modality or general time-series model. Please compare against, or discuss, a few ECG–PPG multimodal foundation models.

  5. The distribution of gap lengths and their units are not reported. Only the ED median is reported, but not for ICU tasks. Please add a histogram or table of gap lengths per task and split, with sample sizes per bin. This would support the “long-horizon” framing and the no-decay claim. In addition, it is unclear what the unit for delta t is and whether it's the same for ED and ICU. It matters for log(1 + delta t) since ICU gaps are hours and ED gaps are months.

Questions for authors
  1. The hemoglobin-only floor term needs justification. Only the ED hemoglobin floor includes the time from arrival to the start of waveform monitoring. Why only hemoglobin? Was the term added after seeing, in validation patients, that monitoring timing correlates with hemoglobin?

  2. The preprocessing pipeline for public encoders needs clarification. Were all signals first downsampled to 40Hz for PPG and 120 Hz for ECG, then later resampled to each public encoder's native required sampling rate?

2 comments
AuthorsOpenReview ↗

Response to Reviewer zzlr Part 1/2

Thank you for a detailed and specific review. The reviewer's points on pretraining exposure led us to retrain the encoder on training patients only and to recompute every result that uses it. We also reorganised the paper. Native calibration is now its single contribution, and the encoder is one of 9 encoders it is applied to. Its two checkpoints are now called the stage-1 and stage-2 encoders (Ours (initial) and Ours in the submission). This comment answers weaknesses 1 to 3; the next answers weaknesses 4 and 5 and the two questions. Table, figure and section numbers refer to the revised PDF.

W1. Donor − current, own − current first, capped donors

We added donor − current with its 95% interval for every encoder to Table 3. Own − current now leads: the introduction states it in words, and Figure 2 and Table 3 report it. Own − donor for every encoder is in Figure 3 and Table 24. The reviewer was right that the donor control can overstate the patient-specific gain. Donor − current is below zero for 3 of the 8 encoders in the ED (Stage 1, AnyPPG, Chronos-Bolt-small) and for 3 of the 5 in the ICU (Stage 2, Stage 1, MOMENT-base). For those rows own − donor slightly overstates what the own past adds, and own − current is the conservative figure. With the stage-2 encoder, donor − current is −0.0004 [−0.0031, +0.0020] in the ED, which is what a valid control should show, and −0.0023 [−0.0045, −0.0004] in the ICU. Own − current is above zero on both tasks: +0.039 [+0.023, +0.056] for ED hemoglobin and +0.010 [+0.0050, +0.015] for ICU potassium. Own − donor was the contrast fixed in advance, so it keeps the Bonferroni-adjusted interval. Table 25 gives all contrasts for the stage-1 and stage-2 encoders. We also capped donor reuse, so that each row donates at most once per draw (Appendix H.6). On validation patients, capped minus standard donor is −0.0006 [−0.0029, +0.0016] for the stage-2 encoder with NC-attn, and its interval includes zero for both checkpoints and both estimators (Table 29).

W2. The stage-1 checkpoint

The reviewer was right about the risk. Mapped to patients, 37.6% of the windows of the submitted stage-1 training set came from validation or test patients of the evaluation tasks. We retrained both stages from scratch on training patients only: a pretraining window is kept only if its patient is identified, belongs to an evaluation cohort of its data source, and is a training patient in every such cohort. The retrained stage-1 training set contains 0 such windows. Stage 1 is now trained in this work, without labels, on 3M windows, 89.2% of them from MC-MED and the rest from the MIMIC-III and VitalDB parts of PulseDB. Its checkpoint is the stage-1 encoder. Appendix D.1 and Table 11 give how each window was linked to its patient, the windows kept and dropped per source, and the checks. Appendix D, Appendix E and Appendix K give the corpus, objective, hyperparameters and compute of both stages.

W3. Fairness against the public encoders

No validation or test patient of the lab and haemodynamics tasks is now in the pretraining data, and native calibration holds. Own − donor is +0.040 [+0.023, +0.058] in the ED and +0.012 [+0.0070, +0.018] in the ICU. In the ED the stage-2 encoder has the largest own − donor contrast, above AnyPPG's by +0.015 [+0.0045, +0.026]. The three encoders with MC-MED in their pretraining, the stage-1 and stage-2 encoders and AnyPPG, still have the three largest ED contrasts. Excluding the evaluation patients rules out leakage, but not an advantage of pretraining on the task's data source. We state this in Section 4.4 and Section 5. In the ICU, the contrasts of the stage-1 and stage-2 encoders do not differ, and the contrast is above zero for all 5 encoders used on ECG, including ECGFounder. The public encoders were not retrained. AnyPPG's pretraining lists MC-MED and PulseDB, and PaPaGei's lists VitalDB and MIMIC-III, so any overlap with our evaluation patients would favour them (Appendix J). On the standard PPG tasks the ranking changed. Recomputed among the eight encoders that remain, the submitted mean ranks were 2.54 [1.88, 3.25] for ours and 2.71 [2.08, 3.42] for AnyPPG. After retraining, AnyPPG ranks first (2.62 [1.96, 3.33]) and the stage-2 encoder second (3.08 [2.33, 3.92]). AnyPPG's scores did not change; only the two retrained rows did. The two are statistically on par. We withdraw the claim of ranking first there, and the encoder is no longer presented as a contribution. Three of these tasks lack subject identifiers, so the exclusion cannot be verified for them; Table 22 reports the other 21.

AuthorsOpenReview ↗

Response to Reviewer zzlr Part 2/2

This comment answers weaknesses 4 and 5 and the two questions. Table, figure and section numbers refer to the revised PDF.

W4. ECG–PPG multimodal foundation models

The reviewer asked for this comparison because the encoder was presented as a contribution. It no longer is. Native calibration is the paper's single contribution, and the encoder we pretrain is one of 9 frozen encoders it is applied to (Section 4.1). We now discuss ECG–PPG multimodal foundation models in Section 2, but we did not run more of them. CSFM, xMAE and the wearable encoders of Abbaspourazad et al. (2024) have no openly available weights. NormWear's weights are public, and we did not evaluate it. Two of the encoders we use are pretrained on both modalities. AnyPPG aligns its PPG encoder to a frozen ECG encoder on paired segments. The encoder we pretrain trains first on PPG with ECG as an auxiliary view and then on paired ECG–PPG segments. In the ICU task we also apply native calibration to ECGFounder. Section 5 lists the missing comparison as a limitation.

W5. Gap lengths and their units

Table 8 now gives, for each task and split, the number of rows, the median and the 10th to 90th percentiles of the gap, and the number of rows in each of five gap bins. Its second part counts ICU draws by their number of usable earlier draws. The unit differs between the two tasks, which are fitted separately. Between ED visits the gap is in days (test median 64 d). Within an ICU stay it is in hours (test median 10.9 h to the nearest usable earlier draw). The ED cross-visit floor, NC-lin, NC-attn and the estimator baselines all take the gap as log(1 + Δt), in the unit of their task (Section 3.3, Appendix C). Figure 7b gives own − donor by tertile of the ED gap.

Q1. Monitoring timing in the hemoglobin floor

On validation visits, the timing terms changed the hemoglobin floor's R² by +0.0002 [−0.0028, +0.0031]. Monitoring timing therefore does not predict hemoglobin beyond the case-mix floor. The hemoglobin floor went through extra checks on validation patients, and this was one of them. To check that the waveform's gain does not reflect when monitoring began, we added to this floor two offsets of the start of waveform monitoring, the arrival hour and the weekday. We kept them for hemoglobin. The floors of the other labs do not have them. Arrival hour and weekday carry no clock information, because timestamps are shifted per patient (Appendix C).

Q2. Resampling for the public encoders

Yes. Every encoder receives the same stored windows, PPG at 40 Hz and ECG lead II at 120 Hz. Each public encoder's preprocessing then resamples them to the rate or length that encoder expects, for example 125 Hz for AnyPPG and PaPaGei-S and 500 Hz for ECGFounder (Appendix E). The windows of the standard PPG tasks are likewise resampled to 40 Hz first (Appendix B).

Reviewer_k55VOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes native calibration, which uses a patient's own past lab results and waveforms to better estimate how their lab values have changed over time. It also introduces a new multimodal FM trained on PPG and ECG signals without labels. The proposed method seems to outperform other competing methods.

Critical strengths

The native calibration improved estimates of hemoglobin and potassium change. The donor-control experiment was implemented to attribute the gain to the patient’s own history. The proposed method ranked first among ten encoders on first-draw labs.

Critical weaknesses
  1. The paper is hard to follow, with symbols used before they are explained. There is little intuition given for the design choces. The authors should clearly define symbols, add more motivations and intuitions through out the paper.

  2. The core idea of "calibration" is not well motivated. As implemented in the paper, if the gap between y_j and g(e_j) is constant per patient, then the method reduces to a simple bias correction. You can just define a new function h(e_j) = g(e_j) + b where b is the constant difference/bias. No special calibration is needed in this case. The authors need to show that g(e_i)-g(e_j) add values beyond this simple bias correction.

  3. The performance difference between models are very tiny, with the gains only at the level of 0.0x. It is unclear how statistically signficant these gains are.

Questions for authors

see the weakness

2 comments

Thank you for the review. We reorganised the paper around the method, rewrote the method section and ran the comparison the reviewer asked for. Table, figure and section numbers refer to the revised PDF.

1. Symbols and intuition

The method is now Section 3, ahead of all evaluation details, and it is written in the order a reader needs. Section 3.1 states the inputs and the challenge and defines the notation, one sentence per symbol: the current draw ii with unknown value yiy_i and window embedding eie_i, and an earlier draw jj of the same patient with measured value yjy_j, embedding eje_j and elapsed time Δtij\Delta t_{ij}. Section 3.2 gives the estimate y^i=yj+g(ei)−g(ej)\hat y_i = y_j + g(e_i) - g(e_j) and its intuition: the measured value sets the level, the waveforms supply the change, and a constant patient-specific offset cancels. It also states the relation to a bias correction. Section 3.3 then states what we run as a second equation (Equation 2): the final estimate is the floor's prediction plus a waveform correction, which is computed in two steps and fitted to the floor's residual. Equation 1 is its special case, and the two estimators are described inside this frame. Section 4.1 defines the metric by its formula and each control and contrast in one sentence. The introduction gives the same intuition before any symbol, and Figure 1 shows the principle and how it is combined with the floor in our experiments.

2. Calibration versus a simple bias correction

The reviewer is right about the form. Equation 1 equals a population prediction g(ei)g(e_i) plus the patient's past residual yj−g(ej)y_j - g(e_j), which is the bias correction the reviewer describes. We now say so in Section 3.2. What differs is how gg is fitted. A bias correction fits gg to levels across patients. Native calibration fits it on within-patient differences, so that g(ei)−g(ej)g(e_i) - g(e_j) is trained to predict the change. We tested whether this matters on the held-out ED hemoglobin pairs (Table 2). The bias correction uses the same features with the same penalty grid and cross-fitting as NC-lin, and it has its own donor control. With the stage-2 encoder (Ours in the submission), NC-lin beats it by +0.039 [+0.014, +0.066] and NC-attn by +0.032 [+0.0070, +0.057]. The bias correction's own − donor contrast is indistinguishable from zero (+0.012 [−0.0044, +0.027]). That of native calibration exceeds it by +0.024 [+0.0013, +0.046] for NC-lin and +0.028 [+0.0074, +0.048] for NC-attn. With the stage-1 encoder (Ours (initial) in the submission), NC-lin still beats the bias correction (+0.030 [+0.0072, +0.056]), while NC-attn is borderline (+0.019 [−0.0000, +0.040]), and neither difference in own − donor excludes zero (Table 27). We fixed these contrasts before running the analysis on held-out patients, and the intervals are unadjusted. On validation patients, NC-lin's margin over the bias correction was not significant (Table 23).

3. Small gains and their significance

Every difference we claim carries a 95% bootstrap interval, paired and clustered by patient on the clinical tasks, and where an interval includes zero we say so. For example, the patient's own past windows add to the current window for 6 of the 8 ED encoders, not for all, and Section 4.4 says so. Own − donor on both tasks, the contrast fixed in advance, also has Bonferroni-adjusted intervals: [+0.020, +0.060] for ED hemoglobin and [+0.0064, +0.019] for ICU potassium. The increments look small because the floors are strong: they already contain the patient's previous lab values. In absolute terms, native calibration raises the R² of the hemoglobin change from the floor's 0.281 to 0.348, a relative increase of 24%. For the potassium change it rises from 0.345 to 0.380, or 10% in relative terms. The mean absolute error changes by −4.8% for ED hemoglobin and by −3.6% for ICU potassium (Appendix I). We present the result as a better prior for the next draw, not a substitute for measuring it (Section 4.2, Section 5).

Thanks for the response. While I appreciate the authors’ efforts, I am still not fully convinced of the significance of the contribution and performance improvement. I will therefore keep my original score.

Reviewer_d1KROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work examines how photoplethysmography (PPG) and electrocardiography (ECG) signals can be used to predict changes in laboratory measurements over periods ranging from hours to months. The authors propose native calibration, which uses previous laboratory measurements and changes in waveform embeddings to estimate current laboratory values. They also develop a multimodal waveform foundation model pretrained using self-supervised learning objectives. Experiments on emergency department, ICU, and intraoperative datasets show that the proposed method improves hemoglobin and potassium change prediction compared to clinical baselines. The foundation model also achieves competitive performance on laboratory prediction, hemodynamic estimation, and 24 standard PPG benchmark tasks.

Critical strengths
  • interesting problem though have some clinical validity concerns
Critical weaknesses
  • what works have shown that waveforms can reliably estimate continuous blood lab values? from an ml perspective its great but i worry about the clinincal motivation and whether such a relationship has been established
  • method is a simple residual correction like calibration method
  • the mechanism behind patient specfic improvements are hard to decouple from your results and may be overclaiming
  • improvements are tiny
  • confounding from overlapping pretraining populations
Questions for authors

address weaknesses as further improvements and clarification in rebuttal

5 comments
AuthorsOpenReview ↗

Response to Reviewer d1KR Part 1/2

Thank you for the review. We reorganised the paper around native calibration, which is now its single contribution; the encoder is no longer presented as one. This comment answers points 1 to 3; the next answers points 4 and 5. Table, figure and section numbers refer to the revised PDF.

1. Clinical motivation

We are not aware of work showing reliable estimation of continuous lab values from a standard bedside PPG, and we do not claim it. Our claim is that the waveform adds to a strong non-waveform floor that already holds the patient's previous lab values. What earlier work shows is narrower, and we now cite it in Section 1 and Section 2. Potassium changes the ECG's repolarisation, and T-wave measures track potassium changes within haemodialysis patients (Dillon et al., 2015). In a large ED cohort, regression from the 12-lead ECG estimated potassium with moderate accuracy but had little practical utility for calcium, sodium and creatinine (von Bachmann et al., 2024). Deep learning on the ECG screens for hyperkalemia (Galloway et al., 2019) and anaemia (Kwon et al., 2020). Continuous noninvasive hemoglobin monitors show small mean bias but wide limits of agreement with laboratory hemoglobin (Kim et al., 2014). Recent work predicts lab values from waveform foundation-model embeddings (Wang et al., 2025; Kataria et al., 2025; Gustafsson et al., 2026). Most of this work estimates the value itself from a single recording, yet values differ between patients for reasons the waveform does not explain. This is why we estimate the change within a patient, anchored on a measured value. Per-patient calibration has precedent: in dialysis patients, a potassium regression from the ECG calibrated per patient reported lower error than a population model (Attia et al., 2016).

2. A simple residual correction

In form, yes. Equation 1 is a population prediction plus the patient's residual at the earlier draw, that is, a per-patient bias correction, and we now say so in Section 3.2. What differs is how the map is fitted: on within-patient differences, with one map applied to both windows, instead of on levels across patients. We tested whether this matters against the bias correction itself on held-out ED hemoglobin pairs, with the same features, penalty grid and cross-fitting as NC-lin (Table 2). With the stage-2 encoder (Ours in the submission), NC-lin beats it by +0.039 [+0.014, +0.066] and NC-attn by +0.032 [+0.0070, +0.057]. The bias correction's own − donor contrast is indistinguishable from zero (+0.012 [−0.0044, +0.027]), while that of native calibration is not. With the stage-1 encoder (Ours (initial) in the submission), NC-lin still beats it, NC-attn is borderline, and neither difference in own − donor excludes zero (Table 27). We fixed these contrasts before the run, and the intervals are unadjusted. On validation patients, NC-lin's margin over the bias correction was not significant (Table 23).

3. Mechanism

We agree that the mechanism is not established, and Section 5 now says so. The own past window is closer to the current window in embedding space than any donor's (distance 13.6 against 15.9–17.6). A persistent device, skin or recording signature therefore remains an alternative to calibration, and the evidence that a past window helps only together with its own value rests on one estimator. We also withdrew a claim. The dose-response in the patient's own history is no longer a contribution, because its slope interval now touches zero (+0.0019 [−0.0001, +0.0040]); we call it suggestive only (Figure 7). What the controls do show is narrower. The patient's own past windows add to the current window: own − current is +0.039 [+0.023, +0.056] for ED hemoglobin and +0.010 [+0.0050, +0.015] for ICU potassium. In the ED, with the stage-2 encoder, the past windows of another patient with the same lab value add nothing (donor − current −0.0004 [−0.0031, +0.0020]). In the ICU they slightly lower the gain (−0.0023 [−0.0045, −0.0004]), so there own − current is the conservative figure. On validation patients, the own past also beats demographically matched donors, look-alike donors nearest in embedding space and capped donors (Table 28, Table 29). Hand-crafted PPG features give no gain from the own past, and the negative-control analytes are null (Appendix H.8).

AuthorsOpenReview ↗

Response to Reviewer d1KR 2/2

This comment answers points 4 and 5. Table, figure and section numbers refer to the revised PDF.

4. Small improvements

Every difference we claim carries a 95% bootstrap interval, paired and clustered by patient on the clinical tasks, and own − donor on both tasks also has a Bonferroni-adjusted interval ([+0.020, +0.060] and [+0.0064, +0.019]). The increments look small because the floors are strong: they already contain the patient's previous lab values. Native calibration raises the R² of the hemoglobin change from the floor's 0.281 to 0.348 (+24% relative), and that of the potassium change from 0.345 to 0.380 (+10%). We present it as a better prior for the next draw, not a substitute measurement (Section 4.2, Appendix I).

5. Overlapping pretraining populations

The concern was justified, and we retrained the encoder. In the submitted stage-1 training set, 37.6% of the windows came from validation or test patients of the evaluation tasks. We retrained both pretraining stages on training patients only (Appendix D.1), and the retrained stage-1 set contains 0 such windows. No validation or test patient of the lab and haemodynamics tasks is in the pretraining data. All results of its two checkpoints, now called the stage-1 and stage-2 encoders, were recomputed, and native calibration's gains hold (Table 5). Two caveats remain. The three encoders with MC-MED in their pretraining, these two and AnyPPG, still have the three largest ED contrasts, so the exclusion rules out leakage but not an advantage of in-domain pretraining. Three of the 24 standard PPG tasks lack subject identifiers, so the exclusion cannot be verified for them; Table 22 reports the other 21. The public encoders were not retrained, and their pretraining may include our evaluation patients, which would favour them (Appendix J). On the standard PPG tasks, AnyPPG now ranks first (2.62 [1.96, 3.33]) and the stage-2 encoder second (3.08 [2.33, 3.92]), statistically on par. Recomputed among the same eight encoders, the submitted mean ranks were 2.54 [1.88, 3.25] for ours and 2.71 [2.08, 3.42] for AnyPPG. We withdraw the claim of ranking first there, and the encoder is no longer presented as a contribution.

Reviewer_d1KROpenReview ↗

Tracked Changes

Can the authors provide a revision with tracked changes (via a different text color). It is hard to follow what was a prior result and what was a new result. Thank you and I'll do my best to provide an assessment and whether it warrants a score change after reading the revision with tracked changes before the deadline tomorrow (AOE)

Comments on the paper

1 comment

We thank the reviewers for their careful reading. This comment summarises the revision for all reviewers and the area chair. The responses to each reviewer answer their points in turn. Table, figure and section numbers refer to the revised PDF.

The paper is now about native calibration alone

Reviewer k55v found the paper hard to follow, and reviewers zzlr and d1kr questioned the encoder comparison. We reorganised the paper. Native calibration is its single contribution: a readout for frozen waveform encoders that uses a patient's earlier blood draw as a calibration point and estimates only how the lab value has changed since then. The method is now Section 3. It opens with the inputs and the challenge and defines every symbol before use. The evaluation details moved to Section 4.1. The abstract and introduction now state the method and its main result, and all controls are in Section 4. The encoder we pretrain is no longer presented as a contribution, and the title no longer names it (it now ends "via Native Calibration of Waveform Foundation Models"). It is one of 9 frozen encoders that native calibration is applied to. Its two checkpoints are now called the stage-1 and stage-2 encoders (Ours (initial) and Ours in the submission).

What we retrained and why

Reviewers zzlr and d1kr pointed out that this encoder's first pretraining stage was undocumented and could include evaluation patients. This was right. Mapped to patients, 37.6% of the windows of the submitted stage-1 training set came from validation or test patients of the evaluation tasks. We retrained both stages from scratch on training patients only (Appendix D.1). The retrained stage-1 training set contains 0 windows of validation or test patients. Architecture, objectives, hyperparameters and evaluation protocol are those of the submission. We fixed the final checkpoint in advance and evaluated each task family once on held-out test patients. Every result of the two checkpoints was recomputed. The public encoders were not rerun, so their rows are those of the submission.

Results on held-out patients

With the stage-2 encoder, the analysis fixed in advance:

  • ED hemoglobin, from PPG: the R² of the change rises from 0.281 to 0.348 (+0.067 [+0.044, +0.090]). The patient's own past windows add +0.039 [+0.023, +0.056] to the current window. A donor's past windows add nothing (−0.0004 [−0.0031, +0.0020]).
  • ICU potassium, from ECG: it rises from 0.345 to 0.380 (+0.035 [+0.022, +0.051]), and the own past windows add +0.010 [+0.0050, +0.015].
  • Bias correction: on held-out ED hemoglobin pairs, native calibration beats a per-patient bias correction that uses the same features, by +0.039 [+0.014, +0.066] with NC-lin and +0.032 [+0.0070, +0.057] with NC-attn (Table 2).
  • Other encoders: the gain over the floor is above zero for all 13 encoder–task pairs, and the own past windows add to the current window for 11 of them (Table 3, Figure 3).

Table 5 sets the submitted and revised headline numbers side by side.

What we withdraw

  • The encoder as a contribution, and with it the claim of ranking first on the standard PPG tasks. Recomputed among the eight encoders that remain, the submitted mean ranks were 2.54 [1.88, 3.25] for ours and 2.71 [2.08, 3.42] for AnyPPG. After retraining, AnyPPG ranks first (2.62 [1.96, 3.33]) and the stage-2 encoder second (3.08 [2.33, 3.92]). AnyPPG's scores did not change. The two are statistically on par (Section 4.6).
  • The dose-response in the patient's own history as a contribution. Its slope interval now touches zero, and we call it suggestive only (Appendix H.4).
  • Models and analyses that were not retrained: the two single-stream variants, the second front-end row of Ours (initial), the ResNet comparison, the context-length analysis and the test results at intermediate checkpoints.

What we added

  • Donor − current for every encoder (Table 3), and all contrasts in Table 24 and Table 25.
  • A held-out comparison of native calibration with a per-patient bias correction (Table 2, Table 27).
  • Capped donors (Appendix H.6), gap lengths and their units (Table 8), and the standard-task results on the tasks with subject identifiers (Table 22).
  • The clinical background on lab values and waveforms, and a discussion of ECG–PPG multimodal foundation models (Section 2).

Changes are marked in blue in the revised PDF. Because the main text was reorganised, most of it is marked; Appendix A lists what changed in substance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.