acceptodds
Under review as a conference paper at ICLR 2027

A Decodability-Grounded Audit of Clinical Prediction Claims in PPG Foundation Models

Abstract

What does held-out clinical prediction from photoplethysmography (PPG) reveal about physiology? We audit frozen representations through concept decodability, covariate-adjusted and incremental prediction, and post-erasure predictability after probe refitting, with encounter-matched ECG comparisons. Across four encoders and MC-MED encounters, AnyPPG decodes pulse-derived concepts well. For potassium on matched encounters, its correlation falls from to after age and renal adjustment, whereas ECGFounder retains (paired adjusted difference ; 95% CI ). Creatinine likewise attenuates under target-derivative-free adjustment. On a separate -encounter complete-case cohort, adding ECGFounder to the same age, estimated glomerular filtration rate (eGFR), and blood urea nitrogen (BUN) baseline raises potassium correlation by about under ridge and nonlinear predictors; direct AnyPPG increments have confidence intervals including zero. Erasure loss gives the opposite modality ordering from interval decodability: after rank-five subspace removal and refitting, AnyPPG loses in adjusted correlation versus ECGFounder's . AnyPPG starts from a much lower adjusted correlation; its loss is near the heart-rate control. Together, these results identify adjustment-sensitive PPG potassium prediction and stronger retained ECG association, while decodability and erasure capture complementary aspects of the representations.

Primary area:
learning on time series and dynamical systems
Keywords:
photoplethysmography, clinical prediction auditing, concept decodability, covariate adjustment, incremental prediction, representation erasure, PPG-ECG comparison

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 4 reviews, above 27% of the others.
Reviewer_xPE4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether laboratory measurements can be reliably predicted from pretrained PPG representations. The results suggest that much of the predictive signal can be explained by measured patient characteristics, whereas ECG representations preserve more information after adjustment.

Critical strengths

S1: The paper raises an important question and examines it through a careful series of increasingly controlled analyses.

S2: The careful study design avoids patient overlap and accounts for related observations. The authors are also transparent about methodological corrections and appropriately cautious when the evidence is inconclusive, which strengthens the credibility of the work.

S4: The evaluation is comprehensive, and the target-derivative rule provides a useful safeguard against inappropriate covariate adjustment.

Critical weaknesses

W1: Laboratory test results are linked to waveforms by clinical encounter rather than by specimen collection time. The findings therefore support encounter-level associations more directly than associations with the patient’s physiological state at the time of measurement.

W2: The strongest ECG incremental result is based on a much smaller complete-case subset, which may differ systematically from the full cohort.

W3: The ECG interval labels are calculated from the same ECG waveform given to ECGFounder, whereas AnyPPG must infer them indirectly from PPG. The difference in interval prediction therefore reflects both representation quality and access to the source signal.

Questions for authors

Q1: Can any subset be aligned to specimen time, or should the claims be narrowed to encounter-level association?

Q2: How do the conclusions change when encoders with unresolved MC-MED exposure are excluded or ECG and PPG use matched inputs?

3 comments
AuthorsOpenReview ↗

Temporal Alignment and Sensitivity to Complete-Case Selection

Thank you for the careful assessment of timing, selection, and model comparability. The additional analyses show that ECG's potassium increment persists in interval-incomplete encounters and under matched input and training conditions. We address each concern below. Here, r denotes Spearman correlation, Δr denotes improvement over the specified clinical baseline, and brackets give 95% patient-cluster bootstrap intervals conditional on fitted models. The added comparisons are sensitivity analyses.

W1/Q1. Temporal alignment and the prediction claim

The MC-MED laboratory claim is encounter-level association and incremental prediction. ECG and PPG are synchronized within each recording and evaluated against the same laboratory target and clinical baseline. This defines a paired comparison of their information about the encounter's laboratory measurement. MC-MED provides laboratory order and result-reporting times. MC-MED documentation.

We also completed a temporally linked external analysis using MIMIC-IV Waveform v0.1.0 and Clinical v3.1. The primary sample selects the latest eligible 30-second synchronized waveform within the 30 minutes preceding laboratory charttime, yielding 141 potassium and 142 creatinine patients. MIMIC documentation describes charttime as usually corresponding to specimen acquisition; we therefore identify this analysis as chart-time aligned. LABEVENTS documentation.

The 15- and 60-minute sensitivity cohorts retain identical laboratory–waveform pairs for shared admissions, so those comparisons assess eligibility. We use encounter-level wording for MC-MED and chart-time wording for MIMIC, with neither presented as individually verified collection-time alignment. This adopts the claim definition proposed in Q1 while adding an external cohort with an explicit temporal selection rule.

W2. Complete-case selection

We separated selection of training patients from selection of evaluation encounters. Of 6,348 potassium-available encounters, 6,037 from 4,561 patients meet the adjustment-variable requirements; these comprise 5,160 interval-complete and 877 interval-incomplete encounters. Thus, the initial 311-encounter reduction reflects covariate availability.

Models fitted on all eligible training patients give the following increments over age/eGFR/BUN:

Evaluation populationEncountersECG Δr [95% CI]
All eligible6,0370.0942 [0.0777, 0.1118]
Interval-complete5,1600.0951 [0.0762, 0.1147]
Interval-incomplete8770.0879 [0.0517, 0.1277]

The paired complete-minus-full difference is 0.0009 [−0.0054, 0.0070]. Inverse-probability weighting, with selection probabilities and truncation thresholds estimated from training patients, gives a complete-case increment of 0.0941 [0.0752, 0.1135]. We also cross-evaluate models trained on all eligible versus complete-case patients. The positive increment in the interval-incomplete group directly establishes that successful extraction of every interval is unnecessary for the observed ECG gain. Appendix S reports the cohort reconciliation, training-population comparisons, and weighting sensitivity.

AuthorsOpenReview ↗

Reference-Signal Access, Pretraining Exposure, and Matched ECG–PPG Comparisons

W3. Access to the reference signal

We interpret concept-prediction R² as information accessible from a specified encoder–input pair. The original patient-disjoint matched-segment experiment illustrates the reference-source dependence:

Reference-feature groupAnyPPG R²ECGFounder R²HR-only R²
PPG morphology0.53420.19400.1858
ECG intervals0.09510.48830.0529

These fixed-ridge readouts use the same concept-valid rows within each comparison. The original protocol uses the first aligned segment, with central eight-second PPG and ten-second ECG inputs; ECG is resampled from the archived 125-Hz signal. The directional pattern makes access to the reference signal explicit: each representation predicts its own modality's features more accurately. The expanded reciprocal evaluation covers both reference sources across four released encoders and six study-trained CNNs.

For target prediction, potassium supplies a reference measured independently of either waveform feature extractor. The matched CNNs below then test whether its ECG–PPG contrast remains when architecture, input support, and supervision are held fixed. These comparisons distinguish reference-feature accessibility from the clinically measured target association.

Q2. Pretraining exposure and matched inputs

Exact-release provenance records support MC-MED source exclusion for PaPaGei-S, Pulse-PPG, and ECGFounder; AnyPPG has documented exposure. Excluding AnyPPG leaves two additional PPG comparisons. Separately, the CNNs use identical 10-second, 125-Hz, single-segment inputs, architecture, and supervision, with a fixed 60-epoch endpoint across three seeds.

ControlComparisonAdjusted ECG−PPG r difference [95% CI]
Source-excluded releases, equal eight-second supportECGFounder−PaPaGei-S0.2208 [0.1854, 0.2539]
Source-excluded releases, equal eight-second supportECGFounder−Pulse-PPG0.2077 [0.1728, 0.2432]
Matched CNN, seed 0ECG−PPG0.0861 [0.0503, 0.1226]
Matched CNN, seed 1ECG−PPG0.0833 [0.0476, 0.1192]
Matched CNN, seed 2ECG−PPG0.0611 [0.0256, 0.0970]

All contrasts use the common 6,037-encounter cohort and age/eGFR/BUN adjustment. The released-model adaptation preserves each model's sampling requirements; the CNNs match sampling as well as duration. The fixed-endpoint results replace the earlier evaluation-selected checkpoints. Provenance establishes documented source exclusion; individual pretraining identities are unavailable for direct matching. The conclusion applies to the tested encoders and matched CNNs on this potassium cohort; it does not establish a modality-wide information limit.

The added controls give consistent evidence across evaluation populations, alternative PPG encoders, and three matched training runs. They directly address whether the potassium contrast depends on interval-complete selection, AnyPPG exposure, or the original input and architecture differences.

Thank you again for your thoughtful feedback. We would appreciate your assessment of whether our responses and additional analyses adequately address your concerns about temporal alignment, complete-case selection, reference-signal access, and pretraining exposure. Please let us know if you have any remaining concerns or would find further clarification or evidence helpful. We would be happy to respond during the remaining discussion period. If we have resolved your concerns, we would be grateful if you would reconsider your overall assessment and rating in light of the additional evidence.

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

This paper presents a decodability-grounded clinical audit framework for PPG foundation models paired with ECG foundation model. The audit evaluates clinical prediction claims across three axes: physiological concept decodability, target-derivative-free covariate-adjusted and incremental prediction, and post-erasure predictability.

Critical strengths
  1. The paper prove that PPG models don't actually see chemical levels like potassium in the blood. Instead, they cheat by using shortcuts like a patient's age, sex, and kidney health. When the researchers blocked these shortcuts, the PPG's predictive power collapsed, while ECG kept its genuine signals.
  2. The authors use strict controls to keep themselves honest. This includes testing against "sham" data (like trying to predict a patient's blood type), utilizing a control model trained from scratch, and openly documenting their own withdrawn analyses to ensure full transparency.
Critical weaknesses
  1. When comparing the clinical data alongside the deep-learning AI embeddings, the model used a one-size-fits-all penalty. Because the AI embeddings are massive (512 dimensions) and the clinical data is tiny, this math setup accidentally over-penalized and suppressed the clinical details, artificially dragging down the performance. 2.Because some of the historical data analyses were withdrawn, almost all the major laboratory evidence relies on a single cohort (MC-MED). Evaluating the models on an external, completely separate hospital dataset would be necessary to prove these findings hold true universally.
Questions for authors
  1. Since applying a single isotropic L_2 penalty (\alpha=1) across the low-dimensional covariates and the high-dimensional embeddings creates a severe regularization mismatch (as evidenced by the negative increments for CRP and creatinine in Table 3), how do the broad incremental results and expert-tier correlations change if you use a block-penalized ridge regression or validation-tuned fallback across all 26 targets?

  2. Given that AnyPPG had documented pretraining exposure to the MC-MED cohort, and that all primary laboratory evidence now relies on this single database, what steps are being taken to evaluate the framework on a completely independent, overlap-safe public cohort (such as the MIMIC-III/IV Waveform Matched Subsets) to ensure the clinical shortcut findings generalize externally?

3 comments
AuthorsOpenReview ↗

Block-Specific Regularization: Target-Level Gains and Updated Expert-Tier Correspondence

Thank you for identifying regularization and external evaluation as priorities. We have completed the requested fitting comparison across all 26 targets and four encoders, recomputed the expert-tier analysis, and evaluated transfer to MIMIC-IV. These additions establish how model fitting changes the internal estimates and directly test their transport to another hospital.

W1 and Q1. Incremental prediction with separate block penalties

Clinical variables and embeddings now receive separate penalties, with an unpenalized intercept. Both penalties are selected through five-fold, patient-grouped validation within the training set, starting from {10⁻⁶, 10⁻⁴, 10⁻², 1, 10², 10⁴} and expanding at search boundaries. The clinical-only baseline is independently tuned. Selection between baseline and combined prediction also uses training validation; differences below 0.001 favor the baseline or stronger regularization. Evaluation increments retain their observed signs.

We preserve the original patient splits and comparison rows within each target–encoder pair. For AnyPPG, the requested comparison gives:

Target familyOriginal median incrementBlock-tuned median incrementTargets selecting the baseline
15 laboratory targets, ΔSpearman r0.00930.02480
11 binary targets, ΔAUROC0.02020.02592

The two targets raised in the review also change sign under block tuning:

AnyPPG targetOriginal Δr [95% CI]Block-tuned Δr [95% CI]Block-tuned Holm p
CRP−0.1653 [−0.2514, −0.0758]0.0248 [0.0013, 0.0479]1.0000
Creatinine−0.0627 [−0.0852, −0.0429]0.0074 [0.0037, 0.0108]0.002119

The original increments use the archived clinical baseline; the block-tuned increments use the separately tuned baseline. Both signs reverse, showing that the original negative increments depend on the fitting specification. The intervals are pointwise 95% patient-cluster bootstrap intervals conditional on fitted models. Creatinine passes Holm correction across the 104 target/encoder comparisons; CRP does not. We therefore distinguish the change in the CRP estimate from evidence of a gain after multiplicity correction.

We applied the same fitting procedure across all targets, with selection confined to training patients. Among 312 selected fits, 25 retained boundary flags; further grid expansion improved the best mean validation score by at most 0.000211. Appendix AG reports the complete regularization sensitivity and the corresponding expert-tier analysis.

Q1. Expert-tier correspondence after retuning

We retained the original expert tiers and recomputed their association with the tuned increments, correcting the permutation statistic and ABO/Rh metric-family assignment. The earlier p = 0.440 is replaced by the following explicitly defined analyses:

AnyPPG analysisSpearman ρPermutation p
All 26 targets, pooled0.44670.0228
15 laboratory targets0.39840.1604
11 binary targets0.53970.0904
Metric-stratified, rank-normalized0.46680.0176

These exploratory p-values are unadjusted. The rank-normalized analysis accounts for the different scales of Δr and ΔAUROC and supports an aggregate correspondence with expert expectations; the subgroup estimates show its variation across target families. Target-specific increments remain the evidence for individual prediction claims.

AuthorsOpenReview ↗

Frozen Transfer to MIMIC-IV: External Results, Checkpoint Provenance, and CNN Controls

W2 and Q2. External evaluation with a frozen source-trained pipeline

We linked MIMIC-IV Waveform v0.1.0 with Clinical v3.1. Preprocessing and probes were fitted on source training patients and frozen for external evaluation. The primary sample selects the latest eligible 30-second synchronized waveform within the 30 minutes preceding laboratory chart-time. Baselines are harmonized across source and destination: age, sex, creatinine, and BUN for potassium; age, sex, and BUN for creatinine. This evaluates transfer from the MC-MED emergency-department cohort to ICU monitoring at another hospital. MC-MED; MIMIC-IV Waveform.

The release provenance audit in Appendix O supports MIMIC-IV source exclusion for the exact Pulse-PPG and ECGFounder checkpoints used here. The MIMIC-IV downstream fine-tuning reported for ECGFounder is a separate experiment from the pretrained checkpoint evaluated here. This establishes documented source exclusion; individual pretraining identities remain unavailable for direct overlap checks.

External targetEncoderPatientsΔSpearman r over baseline [95% patient-bootstrap CI]
PotassiumPulse-PPG1410.0009 [−0.0093, 0.0126]
PotassiumECGFounder1410.0541 [−0.0144, 0.1222]
CreatininePulse-PPG142−8.25 × 10⁻⁷ [−1.21 × 10⁻⁵, 0.0000]
CreatinineECGFounder1420.0002 [−0.0071, 0.0073]

The ECGFounder extension followed inspection of earlier external results and used corrected waveform-quality selection; we report it as exploratory. Positive external incremental gain is not established at the available precision. For ECGFounder potassium, the direct external-minus-source increment is −0.0210 [−0.0874, 0.0511], using the same pipeline and baseline in both cohorts. This comparison directly measures transport uncertainty rather than attributing the external result to hospital shift from separate significance tests. The external protocol and ECGFounder extension are reported in Appendix T and Section T.1.

We also added CNN controls with explicit source-only training provenance: representation learning, tuning, and probe fitting are confined to source patients, with CNN training repeated within the corresponding inner and outer folds. This required 180 training runs of 60 epochs each. Evaluation on the separate 74-patient potassium and 76-patient creatinine concept-complete cohorts yields 24 increment intervals for full-representation and decoded-concept models, all containing zero. These controls test transfer independently of the released encoders' pretraining histories; individual-level cross-dataset identity matching remains unavailable.

The completed analyses strengthen the fitting comparison and provide an external test of the prediction claims. We now distinguish the supported internal gains from the external transport estimates, with the baseline, training protocol, and uncertainty specified for each.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly through the evaluation comparisons.

These additions substantially strengthen the experimental validation and address several of my original concerns. While I still believe the methodological novelty is somewhat limited.

Overall, I believe 3/4 (Weak Accept) is a fair assessment, balancing the paper's contributions against its remaining limitations.

Reviewer_8WQfOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The waveform (foundation) models are often trained or evaluated on clinical prediction tasks to learn or validate their physiological waveform representations. However, predictive performance alone doesn't reveal what physiological or clinical information encoded in these representations. This paper aims to address this gap by proposing a 3 axis audit framework combining physiological concept decodability, covariate-adjusted and incremental prediction, and prediction after concept-related subspace removal.

The authors evaluate four frozen encoders (three PPG, one ECG) on MC-MED. The most detailed analysis focuses on potassium and compares AnyPPG with ECGFounder. After adjustment for age, eGFR, and BUN, the PPG–potassium association attenuates substantially, whereas ECG retains a stronger association and incremental predictive value. Concept decodability and erasure loss rank the two modalities differently; the authors interpret this as evidence that the two measurements capture complementary aspects of the representations.

Critical strengths

This research question is of significant importance for the practical application of FM in clinical settings. A controlled matched analysis indicates that the association between PPG and potassium can be largely explained by age and kidney-related covariates.

Critical weaknesses

W1. Limited added value of the proposed audit framework over standard covariate adjustment.

The most informative finding is that PPG–potassium associations attenuate substantially after adjustment while ECG associations largely persist. This result comes from standard covariate adjustment, which prior ECG studies have already applied (e.g., Galloway et al., 2019). The erasure analysis yields small absolute losses from substantially different baselines, and the PPG loss is close to the heart-rate control. It is therefore unclear what this axis adds beyond the adjustment results. The erasure procedure is also closely related to existing linear concept-erasure methods such as INLP (Ravfogel et al., 2020) and LEACE (Belrose et al., 2023), but the paper does not discuss them. Its scope is narrow: the full 3 axis audit covers one target, two encoders, and one dataset. One of the two encoders (AnyPPG) was pretrained on that dataset, and no external validation is reported. The most important limitation is that the framework stops at an evidence profile and gives no criteria for accepting or rejecting a clinical prediction claim. Because the paper's contribution is the audit framework itself, these limitations are the main reason for my overall rating.

W2. Insufficient justification of concept selection and covariate adjustment in the potassium case study.

The physiological interpretation rests on a limited set of ECG concepts (QRS duration, PR interval, P-wave duration, and QTc). This set omits established potassium-related markers such as T-wave morphology and U waves. The omission could explain both the much larger incremental value of full ECG embeddings over interval features and the discrepancy between concept decodability and erasure sensitivity. Treating renal function as separate from potassium physiology is also questionable, because renal excretion is a major physiological determinant of serum potassium. The substantial attenuation after adjustment for age, eGFR, and BUN shows that the PPG signal overlaps with age and renal-function information. It does not establish that the removed information is physiologically irrelevant. Since potassium is the only target on which the full audit is carried out, these issues weaken the paper's main demonstration of what the audit can reveal.

Questions for authors
  1. How should the three axes be combined to judge whether a clinical prediction claim is supported? Is there any target among the 17 nontrivial targets for which concept decodability or erasure leads to a different conclusion than covariate-adjusted and incremental prediction alone?

  2. The concept basis omits T-wave features (e.g., T-wave amplitude, T-peak-to-end interval, T-wave slope). Why were these excluded from C_ref in the same-A comparison (Tables 1 and 21) and from C_phys in the erasure analysis (Table 17)? With these features included, does the gap between the full ECG representation and the reference concepts narrow, and does ECGFounder's erasure loss change?

  3. Why are eGFR and BUN treated as nuisance variables rather than as part of potassium physiology, given that renal excretion is a primary determinant of serum potassium? How much of the PPG attenuation is attributable to age, and how much to eGFR and BUN?

  4. The core potassium analyses use only AnyPPG, which was pretrained on MC-MED (Appendix O). Why were PaPaGei-S and Pulse-PPG, which have no documented MC-MED exposure, not included in the adjusted-association analysis (Table 8)?

7 comments
AuthorsOpenReview ↗

Clarifying the Audit’s Contribution and Its Relationship to Established Erasure Methods

Thank you for identifying the comparisons needed to establish the audit's added value. Below, Δr denotes a Spearman-correlation increment; brackets denote 95% patient-cluster bootstrap intervals conditional on fitted models. The added analyses are exploratory and sensitivity comparisons; Holm-adjusted p-values are reported separately.

W1/Q1. What each audit axis contributes to the claim

We distinguish three questions with explicit criteria. Added prediction is supported when the held-out increment over a specified clinical baseline is positive with a paired interval excluding zero. Concept accessibility is measured by held-out concept-prediction R² relative to simple controls, with reference quality assessed separately. Concept-associated sensitivity is supported when removing the selected linear subspace and refitting the predictor produces excess loss over matched controls after multiplicity correction. Together, these measurements assess the evidence for a specified feature interpretation of the prediction gain. Linear erasure tests sensitivity to the selected directions; nonlinear concept information can remain, and the result does not establish causal mediation.

For example, the manuscript's original patient-disjoint, fixed-ridge comparison gives mean ECG-interval R² of 0.0951 for AnyPPG and 0.4883 for ECGFounder, versus 0.0529 for HR alone. The added AF and conduction analyses then test whether predictive gains have corresponding intervention evidence. Their baseline comprises clinical/triage variables plus mean PPG heart rate. AF uses four normalized interval-irregularity features; conduction uses primary QRS/PR and a secondary expansion including P-wave duration. Concepts are conditioned on the baseline and HR; the target probe is refitted after erasure, with correction across 192 comparisons.

Target, encoder, and concept setΔAUROC over clinical + HR [95% CI]Excess loss over HR control [95% CI]Holm p
AF history, AnyPPG, irregularity0.0446 [0.0321, 0.0580]0.0212 [0.0108, 0.0317]0.005843
Conduction history, ECGFounder, QRS/PR0.0981 [0.0521, 0.1471]0.0297 [0.0076, 0.0537]0.38886
Conduction history, ECGFounder, +P duration0.0981 [0.0521, 0.1471]0.0553 [0.0255, 0.0887]0.029879

AF prediction and erasure use separate cohorts of 6,703 and 7,378 encounters; conduction uses 5,497 encounters. AF supports both added prediction and irregularity-associated sensitivity. For conduction, the primary QRS/PR test remains unresolved after correction, while the broader secondary basis supports sensitivity. These comparisons provide evidence of sensitivity to the tested concept directions beyond covariate adjustment. Both endpoints are documented history labels, so they do not validate rhythm diagnosis during the recording. Appendices Y and AH give the prediction ladders and the complete erasure family.

W1. Relationship to established erasure methods

Our projection belongs to the linear-erasure literature represented by INLP and LEACE. Across 26 configurations and 52 fixed/retuned rows, matched LEACE and our construction agree at stored target-score precision, with maximum prediction difference 1.69 × 10⁻¹⁴ in identical retained whitened coordinates. The empirical contribution is the clinical comparison linking feature accessibility, incremental prediction, and controlled sensitivity, building on the distinction between probing performance and task relevance in Amnesic Probing.

AuthorsOpenReview ↗

T-Wave Morphology: Measurement Agreement, Predictive Gain, and Sensitivity Beyond Heart Rate

W1/Q2. Excess erasure loss beyond heart rate

On 4,237 encounters from 3,391 patients, we condition concepts on age/eGFR/BUN and HR, fit the erasure directions on training patients, and retune/refit the target probe after removal. Random, shuffled-concept, and rank-matched HR erasures provide controls. The paired results below use Holm correction across the 192-comparison family.

EncoderControlled correlation-loss contrastEstimate [95% CI]Holm p
AnyPPGC1 excess over HR control0.0021 [−0.0074, 0.0114]1
ECGFounderC1 excess over HR control0.0347 [0.0210, 0.0485]4.884 × 10⁻⁵
AnyPPGC1 − C0 after HR-control subtraction0.0026 [−0.0037, 0.0084]1
ECGFounderC1 − C0 after HR-control subtraction0.0335 [0.0211, 0.0464]2.506 × 10⁻⁵

ECG sensitivity exceeds the HR control and increases with the expanded descriptors. These within-model contrasts directly address the differing starting correlations and the effect of removing additional directions; the corresponding PPG comparisons remain unresolved. Appendix AF reports the full comparison family.

W2/Q2. T-wave expansion and the remaining prediction gap

We added signed T amplitude, T/R ratio, T-peak-to-end interval, and ascending/descending slopes from native calibrated ECG. C0 contains the original five intervals; C1 adds these morphology descriptors. Models are fitted on training patients and compared on the same 4,335 evaluation encounters from 3,458 patients, with age/eGFR/BUN as the baseline.

Added features or paired comparisonΔr [95% CI]
Original intervals, C00.0052 [0.0001, 0.0110]
Intervals plus morphology, C10.0249 [0.0142, 0.0366]
Full ECGFounder0.0837 [0.0634, 0.1054]
C1 − C00.0197 [0.0105, 0.0288]
Full ECGFounder − C10.0588 [0.0406, 0.0770]

The expansion narrows the gap by 0.0197 on identical evaluation patients (Appendix V). We also compared automatic measurements with one reader's annotations. After the validity checks, 164 of the original 200 windows had paired measurements. Concordance was strongest for amplitude (CCC 0.6370; MAE 0.1013 mV), with lower agreement for slopes (0.2022/0.4040), T-peak-to-end (0.0107; MAE 45.14 ms), and T/R (0.0231). This is a provisional single-reader comparison; the planned independent two-reader quantitative validation remains incomplete.

A sensitivity analysis withheld all measurements in the five windows flagged for inconsistent U/T landmark ordering, leaving 160 paired windows. Amplitude CCC was 0.6225 and MAE was 0.1027 mV; T-peak-to-end MAE was 45.71 ms. The comparison therefore continues to support amplitude more strongly than timing. We retain C1 as an automated morphology descriptor set, with feature-specific measurement agreement reported separately from its predictive contribution.

U-wave assessment is particularly difficult in lead II: the AHA/ACCF/HRS standard describes U waves as frequently absent in limb leads and most evident in V2/V3. Our two-reader pilot gave U-status agreement of 25/30 (κ = 0.488), with remaining disagreements after third-reader assessment. U-wave observations are therefore reported separately from C1. Rautaharju et al., 2009.

AuthorsOpenReview ↗

Age and Renal Adjustment, Additional Encoders, and External Evaluation

W2/Q3. Renal physiology and the age/renal decomposition

Renal function is part of potassium physiology. Conditioning asks what the waveform adds when measured renal status and age are already available. On the fixed 6,037-encounter, 4,561-patient cohort, separately fitted ridge adjustment sets give:

Adjustment setAnyPPG rECGFounder r
Empty set0.20630.3180
Age0.08800.2716
eGFR/BUN0.05180.2670
Age/eGFR/BUN0.03970.2678

Both age and renal information contribute to the attenuation. Across all eight subsets, the order-averaged decreases for AnyPPG are 0.0518 for age, 0.0646 for eGFR, and 0.0502 for BUN. These describe overlapping predictive information. ECG retains a larger association under every listed adjustment. The separate creatinine analysis excludes target-derived eGFR.

W1/Q4. Additional encoders and external evaluation

The common-cohort potassium comparison now includes all four encoders, with each model's native interface and separately tuned clinical/embedding penalties:

EncoderAge/eGFR/BUN-adjusted r [95% CI]Increment over age/eGFR/BUN, Δr [95% CI]
AnyPPG0.0397 [0.0139, 0.0644]0.0060 [−0.0004, 0.0127]
PaPaGei-S0.0407 [0.0152, 0.0671]−8.80 × 10⁻⁷ [−4.18 × 10⁻⁶, 2.43 × 10⁻⁶]
Pulse-PPG0.0224 [−0.0047, 0.0495]0.0031 [0.0013, 0.0049]
ECGFounder0.2678 [0.2416, 0.2932]0.0942 [0.0777, 0.1118]

PaPaGei-S, Pulse-PPG, and ECGFounder have documented MC-MED source exclusion. The contrast therefore extends beyond the exposed AnyPPG checkpoint, while the tuned analysis also identifies Pulse-PPG's small positive increment.

We additionally completed frozen transfer to MIMIC-IV. With a harmonized age/sex/creatinine/BUN baseline, potassium increments in 141 patients are 0.0009 [−0.0093, 0.0126] for Pulse-PPG and 0.0541 [−0.0144, 0.1222] for the exploratory ECGFounder extension. Positive external gain remains unestablished. The expanded study now evaluates transport directly alongside internal prediction and controlled concept sensitivity, giving each claim a corresponding empirical test.

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

This work examines whether AnyPPG, a pre-existing PPG foundation model, can capture concepts beyond pulse-related physiology. Because AnyPPG was developed through cross-modal pretraining with an ECG foundation model, it is reasonable to hypothesize that it may encode information also represented in ECG signals. However, PPG primarily reflects blood-volume changes as blood traverses the peripheral vasculature—typically measured at a finger—whereas ECG records the heart’s electrical activity. Their most direct shared signal is therefore heart rate. In this context, it is unsurprising that the association between AnyPPG-derived representations and serum potassium disappears after adjustment for relevant covariates linked to potassium abnormalities, such as renal dysfunction. In short, potassium (a positively charged ion) affects the conduction of ions through the hearts electrical system (and hence the ECG waveform), but has little to no effect (other than in very extreme cases) on the heart rate. This is an age old fact that has been appreciated for many years. Consequently, none of the papers main conclusions are surprising.

Critical strengths

None

Critical weaknesses

The concept decodability seems to be a standard metric of success based on held-out data; i.e., the held-out coefficient of determination. It is challenging to understand what the authors are trying to convey by attaching the term "decodability" to this notion. The only "special" thing is that the coefficient of determination is computed with respect to 'concepts",but at the end of the day, the concepts are simply labels.

Similarly, all of the methods and analysis techniques seem fairly standard, although they are presented using language that makes this fact quite opaque to the reader. For example, Target-specific adjustment is simply covariate adjustment; Post-erasure predictability is akin to a sensitivity analysis, etc. The methodological novelty is weak.

None of the results regarding the relationship between potassium and PPG signals are not surprising and completely expected, as noted above.

Questions for authors

To make a compelling argument, the authors need to explore another prediction task other than potassium. Can the authors use a audit task that is a function of factors that are known to associated with the PPG?

Even if another audit task is found, can the authors to put their methods into better context, clearly stating what is new in their approach and analysis?

6 comments
AuthorsOpenReview ↗

Clarifying the Audit’s Contribution: Methodology and AF-History Evidence

Thank you for asking us to clarify the contribution and examine a task directly connected to PPG physiology. The added AF-history analysis demonstrates positive PPG predictive value together with sensitivity to pulse-irregularity directions beyond a heart-rate control. The conduction and expanded potassium analyses test whether the observed gains are sensitive to directions associated with the proposed physiological features.

1. Terminology and methodological contribution

“Concept decodability” is held-out R² for predicting a specified physiological reference label from a frozen representation. We use the more explicit term “held-out concept-prediction R²,” with the reference feature and probe class stated. Covariate adjustment measures association conditional on recorded clinical variables; incremental prediction measures improvement over a specified baseline; concept erasure measures sensitivity to selected representation directions after the target probe is refitted.

These established components answer different questions. Amnesic Probing shows why successful feature prediction and task importance require separate assessment. We place the erasure component within the literature on INLP and LEACE. Under matched retained whitened coordinates and downstream settings, our 26-configuration LEACE comparison gives a maximum prediction difference of 1.69 × 10⁻¹⁴ across 52 fixed/retuned result rows.

The empirical contribution is to connect feature accessibility, added prediction, and controlled sensitivity in clinical waveform models. These comparisons test whether prediction gains have supporting evidence at the level of the specified features. Linear erasure can leave nonlinear concept information accessible, so the intervention measures sensitivity to the selected directions and does not establish complete information removal or causal mediation.

2. AF history provides the requested PPG-related task

We use four normalized beat-interval irregularity measures for AF history and include mean PPG heart rate separately in the clinical/triage baseline. PPG irregularity has an established connection to AF, supported by prospective evaluation against ECG monitoring in the Fitbit Heart Study. Our endpoint is documented AF history, so these results do not validate contemporaneous AF detection.

For the intervention, concepts are conditioned on the clinical baseline and HR. Whitening and concept directions are fitted on training patients, and the target probe is retuned and refitted after removing the selected directions. Random, shuffled-concept, and rank-matched HR erasures provide controls. The table reports excess loss over the HR control; brackets are 95% patient-cluster bootstrap intervals conditional on the fitted models, with Holm correction across 192 comparisons. The added analyses are exploratory and sensitivity comparisons.

Target and encoderConcept basisPrediction ΔAUROC over clinical + HRExcess AUROC lossHolm p
AF history, AnyPPGIrregularity0.0446 [0.0321, 0.0580]0.0212 [0.0108, 0.0317]0.005843
Conduction history, ECGFounderQRS/PR, primary0.0981 [0.0521, 0.1471]0.0297 [0.0076, 0.0537]0.3889
Conduction history, ECGFounderQRS/PR + P duration, secondary0.0981 [0.0521, 0.1471]0.0553 [0.0255, 0.0887]0.02988

AF prediction uses 6,703 encounters from 5,018 patients; erasure uses a separate cohort of 7,378 encounters from 5,423 patients. Conduction prediction and erasure use 5,497 encounters from 4,214 patients, with the same intact prediction result shown for both concept sets.

AnyPPG therefore supplies both an AF-history prediction gain and controlled sensitivity to irregularity-associated directions. Conduction illustrates a different outcome: the primary QRS/PR test does not pass family correction, whereas the secondary P-duration expansion does. This makes the dependence on the tested feature basis explicit, beyond what covariate adjustment alone measures. The prediction ladders and complete erasure comparison family appear in Appendices Y and AH of the revised manuscript.

AuthorsOpenReview ↗

Potassium Prediction Beyond Age and Renal Status: Added Value of T-Wave Morphology and Controlled Erasure

3. What the potassium analysis establishes beyond physiological expectations

Known physiology motivates potassium as a test case. The empirical questions concern how much the representations add beyond measured age and renal status, and how that gain relates to the tested waveform features. We fit models on training patients and compare them on the same 4,335 evaluation encounters from 3,458 patients:

Addition to the age/eGFR/BUN baselineSpearman Δr [95% CI]
Original five intervals0.0052 [0.0001, 0.0110]
Intervals plus automatically extracted T morphology0.0249 [0.0142, 0.0366]
Full ECGFounder representation0.0837 [0.0634, 0.1054]

Adding T morphology improves the reference-feature increment by 0.0197 [0.0105, 0.0288], while the full representation retains a paired advantage of 0.0588 [0.0406, 0.0770]. On the separate 4,237-encounter erasure cohort, ECG's excess correlation loss over the HR control is 0.0347 [0.0210, 0.0485; Holm p = 4.884 × 10⁻⁵]. These measurements quantify the contribution of the extracted feature set and sensitivity to its associated directions (prediction ladder: Appendix V; controlled erasure: Appendix AF).

The expanded evidence connects a positive PPG-related task with feature-dependent interpretations across AF, conduction, and potassium. We hope these comparisons clarify the empirical contribution and provide a substantive basis for reconsidering the paper's informativeness.

In my original review I asked the authors to clearly state what is new in their approach and analysis. I am still trying to grapple with what new insights arise from the additional data now provided; i.e., the AF history task. For example, how is irregularity as a concept basis quantified? The work remains challenging to parse and the results are not surprising.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.