acceptodds
Desk rejected at ICLR 2027

PhysioSplit: Delay-Aware Context-Selective Fusion under Cohort Shift

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

ECG and PPG provide complementary views of cardiovascular dynamics, but their timing, beat-level events, and slowly varying subject context can shift differently across patients, sensors, and acquisition protocols. Treating these factors as one undifferentiated representation makes it difficult to preserve transferable dynamics while adapting to a new subject. We introduce PhysioSplit, a delay-aware fusion framework that estimates a bounded fractional ECG–PPG arrival-delay proxy with a continuous-delay state-space encoder and factorizes the resulting representation into context and event codes. Context encoder adaptation updates only the context pathway during target calibration, while the event pathway and shared timing representation remain fixed. Across five cohorts, PhysioSplit reduces the external waveform degradation ratio from 2.23 for SiamQuality to 1.55 and lowers RMSE on the most challenging waveform setting from 13.21 to 6.74 mmHg. It achieves Macro-F1 values of 0.847 on the source cohort and 0.724 on the zero-shot ICD-derived phenotype task. Progressive ablations, delay-agreement diagnostics, temporal reliability measures, and frozen probes connect these gains to sub-sample alignment and structured information routing. The resulting model provides a compact interface for frozen transfer and small-support context calibration without treating learned codes as direct vascular measurements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multimodal learning, domain generalization, transfer learning, representation learning, state space models, physiological time series

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yxpkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PhysioSplit, which studies the cross-cohort transfer of a fusion model for electrocardiogram (ECG) and photoplethysmographic pulse wave (PPG) signals. The tasks include blood pressure waveform reconstruction and phenotype identification. The method models the fractional delays of the two signals using a continuous delay state-space encoder. Subsequently, the representation is decomposed into slowly changing individual context and rapidly changing cardiac events. At the target end, only the context encoder is updated, while the other modules remain frozen. The main contribution is the combination of delay modeling, representation decomposition, and selective adaptation. It not only designs the fusion representation but also clearly limits the range of parameters that can be updated during deployment. The experiments cover five cohorts. VitalDB is the only training source. MIMIC-III and UCI are used for zero-sample transfer. RWW is used for individual calibration with few samples. MC-MED is used for weakly labeled phenotype identification in ICD. The authors also conducted stepwise ablation, delay consistency, representation stability, and freeze probe analyses. These experiments provide support for performance improvement and expected information division. The verification range is still the representation learning and prediction effects under the evaluated conditions, not equivalent to clinical effectiveness verification.

Critical strengths
  1. The performance improvement achieved through freezing on the external test has a substantial impact. The paper only uses VitalDB as the training source and conducts frozen tests on MIMIC-III and UCI without using target support samples. It shows that the benefits reported by the paper are not solely dependent on the additional reference blood pressure obtained during testing, but also include clear zero-shot transfer improvements. For work aiming at cross-cohort generalization, this is more convincing than simply improving the accuracy of the source domain.
  2. Binding the representation decomposition with the update range during deployment has clear methodological research value. The method not only divides the context and event code, but also modulates through an asymmetric decoder, making the context encoder a dedicated interface for target-end calibration while keeping the parameters of the event encoder and the shared temporal module unchanged. This advantage enables the paper to go beyond the scope of simply stacking fusion modules. It transforms "which information should be retained and which parameters can be adapted" into a testable structural design, which is the main basis for my recognition of the method's contribution, but it cannot be asserted based on this that context update is superior to other calibration strategies.
Critical weaknesses

The benefits of selective calibration have not been reliably isolated, and the contribution of the core method lacks fair comparison. The main text claims that w/o CEA is used to isolate the context update at the target end, but on A and B without performing target calibration, the variant's RMSE is 7.48 and 8.97, while the full model is 5.43 and 5.89. This indicates that this comparison cannot be directly understood as "calibration when the checkpoint is turned off during testing", and the differences in the training stage and the update benefits during deployment have not been clearly distinguished. Moreover, in Setting C, the full model obtains target support set calibration and the external baseline remains frozen, and there is no report of global update or event branch update comparisons under the same support budget. Therefore, the existing results cannot fully determine whether the benefits come from selective context update or from meta-training and additional target supervision. This does not negate the zero-shot transfer results, but directly weakens one of the most important method contributions of the paper, which is the primary reason for my inclination to reject the manuscript. The temporal direction of the core delay operator and the "alignment" explanation have an unresolved contradiction. Unless there is an unexplained timestamp redefinition or re-indexing operation, the operator is inconsistent with the claimed delay compensation direction. This may be a problem with the formula expression and cannot be asserted to be an actual code error based on this; however, before clarification, the decrease in waveform error and time delay correlation are not sufficient to prove that "fractional delay alignment" is indeed the source of benefits. Since this mechanism is the main technical contribution of the paper, this uncertainty directly affects my judgment of the reliability of the method.

Questions for authors
  1. Does the variant without CEA truly isolate the role of context calibration during the test? Does this variant use the same meta-training process and the same checkpoint as the complete model, only disabling the update during the test? If so, why are the RMSE values of A and B on the target calibration columns in Table 1 still 7.48/8.97 and 5.43/5.89 respectively, even though the update is disabled during the test? If not, please clearly state which training factors the ablation has changed. Additionally, given that the external baseline in Setting C remains frozen, can the existing evidence distinguish between "the value of context selective update" and "the ordinary calibration gain after obtaining additional target supervision"? This answer will directly affect my judgment of the independent contribution of CEA.
  2. Is the actual implementation of the time index of CDSE consistent with the claimed delay compensation direction in the paper? Are there any unexplained timestamp redefinitions, re-indexing, or output delay conventions in the actual implementation? Please illustrate how events are paired using a set of known delays and their actual read indices. This will determine whether this is a correctable expression error or an implementation defect affecting the effectiveness of the core mechanism.
4 comments
AuthorsOpenReview ↗

Response to Reviewer yxpk (Part 1/2: W1–W2, Q1)

Thank you for identifying the distinction between deployment-time calibration and source training, and for checking the temporal direction of the CDSE operator.


W1: The benefit of selective calibration is not fully isolated by the current table.

A: Table 1 gives the relevant Setting-C comparison: 6.74 ± 0.24 mmHg with context-only CEA versus 11.64 ± 0.52 mmHg with no update, using four support windows, four query windows, and three inner steps (Main Text Section 5.1). The A/B values (5.43 ± 0.14, 5.89 ± 0.17 mmHg) are frozen-transfer results and are not used for this deployment attribution.

The representation evidence in Main Text Table 4 is consistent with this routing: context probes explain age/BMI with (R^2=0.531/0.489), event probes with (R^2=0.094/0.078), while the RWW SBP probe gives AUROC (0.712) for context and (0.597) for event. The deployment update therefore acts on the code that carries the measured persistent variation, while the event path remains fixed for short-time waveform structure.


W2: The temporal direction of the CDSE operator and its alignment explanation require clarification.

A: The CDSE read-index written in Eq. 2 is

pdel[k]=(1−γ)p[k−d]+γp[k−d−1].p_{\mathrm{del}}[k]=(1-\gamma)p[k-d]+\gamma p[k-d-1].

For the known offset ECG[50]→PPG[70] at 125 Hz, d=20d=20. Substituting these indices into the two expressions gives:

R3-Table1: Known-offset CDSE read-index comparison.

ConventionOutput indexRead/pairingInterpretation
Printed PPG operator90reads PPG[70]shifts PPG[70] to output 90
Causal ECG delay70pairs ECG[50] with PPG[70]aligns the earlier ECG feature to PPG time

The causal fractional-delay form is

edel[k]=(1−γ)e[k−d]+γe[k−d−1].e_{\mathrm{del}}[k]=(1-\gamma)e[k-d]+\gamma e[k-d-1].

For a fractional offset of 20.520.5 samples, d=20d=20 and γ=0.5\gamma=0.5; at output 70, the delayed ECG feature is 0.5e[50]+0.5e[49]0.5e[50]+0.5e[49], paired with PPG[70]. Output kk therefore pairs the interpolated ECG feature with p[k]p[k]. Reading p[k+d]p[k+d] instead uses a future PPG sample. The table gives the algebraic consequence of each written read-index convention.

The two equations therefore define distinct read operations: Eq. 2 reads PPG history, while the causal alternative reads ECG history to pair an earlier ECG event with a later PPG event. Appendix D.1 gives the buffer convention and zero-padding rule for the stated recurrence.


Q1: Does w/o CEA isolate the role of context calibration during testing?

A1: w/o CEA is the Setting-C no-update control: it skips the deployment-time context update, while full CEA updates only the context branch from four support windows and evaluates four disjoint query windows after three inner steps (Main Text Sections 3.2–3.3 and 5.1; Appendix E.5). The measured query RMSE changes from 11.64 ± 0.52 to 6.74 ± 0.24 mmHg, an absolute reduction of 4.90 mmHg (42.1%). For patient-level analysis, the paired effect is

Δi=RMSEi,w/o CEA−RMSEi,CEA.\Delta_i=\mathrm{RMSE}_{i,\text{w/o CEA}}-\mathrm{RMSE}_{i,\mathrm{CEA}}.

Both Setting-C conditions start from the same selected checkpoint and use exactly the same query windows. CEA performs three context-update steps on four disjoint support windows; the no-update control performs zero adaptation steps. CDSE, the event branch, decoder dynamics, and task heads are frozen in both conditions. Thus checkpoint selection and query composition are held fixed, and the deployment intervention is the context-branch update. The source/A/B entries in the ablation table are separate transfer rows, not outputs of this deployment toggle.


AuthorsOpenReview ↗

Response to Reviewer yxpk (Part 2/2: Q2)

Q2: Is the actual CDSE time index consistent with the claimed delay direction?

A2: Eq. 2 and Appendix D.1 specify a PPG-history read: at output kk, the interpolator reads p[k−dk]p[k-d_k] and p[k−dk−1]p[k-d_k-1], with weights 1−γk1-\gamma_k and γk\gamma_k. Appendix D.1 also describes a PPG feature buffer and zero-padding when history is unavailable. This specifies the manuscript's operator. Under a causal ECG-to-PPG alignment convention, the earlier ECG feature would instead be delayed to the PPG timebase, using e[k−dk]e[k-d_k] and e[k−dk−1]e[k-d_k-1]. The integer and fractional components are

dk=⌊τkδt⌋,γk=τk−dkδtδt.d_k=\left\lfloor\frac{\tau_k}{\delta_t}\right\rfloor, \qquad \gamma_k=\frac{\tau_k-d_k\delta_t}{\delta_t}.

The printed expression delays PPG, whereas the causal-pairing expression delays ECG. The formula above and the table below distinguish these operations by their read indices and output times. Appendix D.1 discusses the causal pairing; the table below gives algebraic consequences of each expression.

At 125 Hz, a 20-sample delay is 160 ms, so ECG[50] at 0.400 s pairs with PPG[70] at 0.560 s. A 20.5-sample delay is 164 ms and uses the half-sample interpolation above. These conversions make the timestamp convention explicit in both samples and milliseconds.

The fractional case can also be checked from the input-event perspective. For a unit ECG impulse at index 50, with all other ECG samples zero, the causal expression gives:

R3-Table2: Integer and fractional causal-delay examples.

Delay (samples)edel[70]e_{\mathrm{del}}[70]edel[71]e_{\mathrm{del}}[71]Output centroid
201070
20.50.50.570.5

For the second row, the centroid is (70×0.5+71×0.5)/(0.5+0.5)=70.5(70\times0.5+71\times0.5)/(0.5+0.5)=70.5. Thus a fractional delay does not place the entire impulse at integer output 70: it distributes its weight between the adjacent outputs. Equivalently, output 70 interpolates the ECG at input position 49.5. The integer example pairs ECG[50] with PPG[70]; a corresponding 20.5-sample arrival lies at PPG time 70.5. This distinction separates the output read operation from the location of a delayed input event.

Manuscript locations.

  • Main text §§3.2–3.3 and 5.1, pp. 4–6; Appendix E.5, p. 13 — define the Set-C no-update comparison.
  • Main text Eq. 2–3, p. 3; Appendix D.1, p. 8; Algorithm S1 — locations of the delay operator and its read indices.
  • Appendix D.1, p. 8; Figure S7, p. 8 — causal pairing clarification; the table above provides the integer and fractional algebraic examples.

We thank Reviewer yxpk for identifying the CDSE indexing issue and for prompting the explicit causal pairing convention.

Dear Reviewer yxpk,

As the discussion period is nearing its end, we would appreciate your thoughts on our responses and the revised manuscript. Could you let us know which concerns remain most important to your assessment? We would welcome the opportunity to address them in the remaining time. If the responses and revisions have helped address your concerns, we would be grateful if you would consider revisiting your rating.

Thank you again for your time and constructive feedback.

The author's response clarified the experimental protocol for Setting C. The Q1 reply clearly stated that the complete model and the non-updated control used the same checkpoint and exactly the same query window, with the only difference being whether the support set was utilized for context update. According to this explanation, the RMSE decreased from 11.64 to 6.74 mmHg, which can support the target-end calibration gain on this model. I accept this clarification and no longer consider the possible checkpoint or meta-training differences in Setting C as a major concern.

However, the independent contribution of context selective update has not been fully established. The existing control indicates the effectiveness of context update compared to no update at all, but the reply does not provide the results of other calibration strategies under the same target supervision conditions. Therefore, it is currently difficult to determine how much additional value is brought by restricting the update to the context branch itself. Probe results such as age and BMI support the information division of the branches, but they cannot replace a direct comparison of the update strategy. This issue limits the attribution of the method's contribution and does not negate the reported calibration gains.

Regarding CDSE, the reply and R3-Table1 clearly distinguished the causal pairing scheme of PPG historical reading and delayed ECG in the original formula, and provided algebraic examples for integer and fractional delays. However, the reply still did not explicitly state which operator was actually used in the reported main experiment and ablation experiment, nor did it clarify whether this difference exists only in the paper's formula. Therefore, the core question still remains: It is currently impossible to determine whether this is a correctable expression error or a problem related to the actual implementation and mechanism explanation. The algebraic example of the correct pairing method is not sufficient to confirm that the existing experiments actually used this operation.

Based on the current provided responses, I accept the author's clarification of the calibration protocol and also continue to accept the zero-shot transfer improvement reported in the paper. However, the independent value of selective update still lacks sufficient evidence, and the correspondence between the delay mechanism and the actual experimental implementation has not been clarified. I maintain the weak rejection recommendation, where the unresolved delay implementation issue is currently the more important basis for judgment.

Reviewer_RezMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PhysioSplit for cross-cohort ECG--PPG transfer. It combines fractional delay-aware fusion, context/event factorization, and context-only adaptation. The model is trained on VitalDB and evaluated on MIMIC-III, UCI, a wearable cohort, and MC-MED.

Critical strengths

S1: The experiments provide strong evidence connecting the architecture to its proposed mechanisms. The evaluation is comprehensive across settings.

S2: The experimental design reduces leakage through patient-level separation and disjoint support and query windows.

Critical weaknesses

W1: The deployment value of context encoder adaptation is limited because its support loss requires reference ABP—the measurement the model is intended to replace. Without reference ABP, only frozen inference can be used.

W2: The physiological meaning of the learned delay remains unclear because it is validated against an R-peak-to-PPG-foot proxy derived from the same input signals. This verifies timing consistency but not an independent physiological mechanism.

W3: It is unclear how consistently the method performs across patients given the reported variation is based mainly on three repeated runs rather than patient-level uncertainty.

W4: The learned delay is restricted to a fixed range of 100–400 ms, but the paper does not justify this range using the observed delay distributions or test sensitivity to alternative bounds. It is therefore unclear whether the reported gains depend on this hand-chosen constraint.

Questions for authors

Q1: Can adaptation use realistically available targets such as intermittent cuff measurements or unlabeled ECG--PPG consistency?

Q2: How does context-only adaptation compare with standard fine-tuning and other update choices?

3 comments
AuthorsOpenReview ↗

Response to Reviewer RezM (Part 1/2: W1–W4)

Thank you for separating the deployment target requirement, the meaning of the learned delay, and patient-level uncertainty. We address these three points directly below.


W1: The deployment value of CEA is limited when reference ABP is unavailable.

A: CEA requires reference ABP on the four support windows used in Setting C. With that support, query RMSE is 6.74 ± 0.24 mmHg, versus 11.64 ± 0.52 mmHg with no update (Table 1, Main Text Section 5.1). Without a reference target, deployment uses frozen inference; cuff-only and unlabeled consistency calibration are extensions beyond the measured protocol (Appendix E.3).

The deployment interface is therefore explicit: reference ABP supervises only the context-branch update, while CDSE, the event branch, decoder dynamics, and task heads stay fixed. The measured comparison is a 4.90 mmHg (42.1%) reduction under four support windows, four query windows, and three context-update steps.


W2: The physiological meaning of the learned delay remains limited because the reference proxy uses the same input signals.

A: The reference interval is the R-peak-to-PPG-foot interval derived from the same ECG and PPG inputs used by the model. In notation,

τproxy=tPPG foot−tR peak.\tau_{\mathrm{proxy}}=t_{\mathrm{PPG\,foot}}-t_{\mathrm{R\,peak}}.

With the prespecified 100–400 ms gate in Main Text Section 3.1 and Appendix E.7, this quantity measures timing consistency with a signal landmark. We therefore refer to it as an ECG-to-PPG arrival-delay proxy and interpret the agreement analysis as signal-landmark consistency; it is not used as an independent vascular-physiology measurement. The definition and interpretation are stated in Main Text Section 5.3 and Appendix E.4.

At 125 Hz, this range corresponds to 12.5–50 samples. The learned delay is decomposed as

dk=⌊τkδt⌋,γk=τk−dkδtδt,d_k=\left\lfloor\frac{\tau_k}{\delta_t}\right\rfloor,\qquad \gamma_k=\frac{\tau_k-d_k\delta_t}{\delta_t},

so the model represents both the integer offset and the fractional remainder.

The delay diagnostics in Main Text Table 4 report (r=0.934), MAE (4.7,\mathrm{ms}), and ICC (0.918) on VitalDB, and (r=0.908), MAE (6.1,\mathrm{ms}), and ICC (0.893) on MIMIC-III for the fractional-delay variant. These values connect the proxy definition to both agreement and reliability measurements.

Revision locations.

  • Main text §§3.1 and 5.3, pp. 3 and 9; Table 3, p. 8; Appendix E.4, p. 12 — use the term arrival-delay proxy and describe the agreement as signal-landmark consistency.

W3: The reported seed variation does not quantify consistency across patients.

A: The reported ± values summarize three independently initialized training runs; for example, Table 1 reports 6.74 ± 0.24 mmHg for CEA and 11.64 ± 0.52 mmHg for no update. Patient-level consistency is computed separately from paired predictions for each patient using a patient-level paired analysis or bootstrap interval (Appendix E.4).

For patient-level comparison, the corresponding effect is

Δi=RMSEi,no update−RMSEi,CEA,Δˉ=1N∑i=1NΔi,\Delta_i=\mathrm{RMSE}_{i,\text{no update}}-\mathrm{RMSE}_{i,\mathrm{CEA}}, \qquad \bar\Delta=\frac{1}{N}\sum_{i=1}^{N}\Delta_i,

with patients, rather than windows, as the resampling unit.

The seed-level standard deviation and the patient-level paired interval answer different questions: the former measures run-to-run variation, while the latter measures whether the per-patient difference is consistently positive. The RMSE table therefore reports seed variation, and the paired estimand is defined by (Delta_i) above.


W4: The fixed 100–400 ms delay range requires justification and sensitivity analysis.

A: The manuscript prespecifies 100–400 ms as the physiological ECG-to-PPG arrival-delay range and uses it to define the learned delay gate. The sensitivity table reports:

R2-Table2: Delay-bound sensitivity values.

Delay boundsSet ASet BSet C
100–400 ms5.43 ± 0.145.89 ± 0.176.74 ± 0.24
80–420 ms5.38 ± 0.175.66 ± 0.206.55 ± 0.30
150–350 ms5.80 ± 0.236.05 ± 0.337.00 ± 0.45

The three rows show the sensitivity of the reported endpoint to the delay range. Relative to 100–400 ms, widening the range to 80–420 ms changes Set-C from 6.74 to 6.55 mmHg (−2.8%), while narrowing it to 150–350 ms changes Set-C to 7.00 mmHg (+3.9%). This is a range-sensitivity analysis around the prespecified 100–400 ms operating range; the associated histograms and occupancy statistics describe how the learned delays use each range.

Revision locations.

  • Main text §3.1, p. 3; Appendix E.7, p. 14; Table 3, p. 8 — state 100–400 ms as the prespecified operating range.
  • Appendix E.7, p. 14; Table S4, p. 13 — document the alternative-range protocol together with delay histograms and occupancy statistics.

AuthorsOpenReview ↗

Response to Reviewer RezM (Part 2/2: Q1–Q2)

Q1: Can CEA use realistically available targets such as intermittent cuff measurements or unlabeled ECG–PPG consistency?

A1: The evaluated CEA procedure is supervised target calibration: reference ABP from four support windows updates only the context branch, while CDSE, the event branch, decoder dynamics, and task heads remain frozen (Main Text Sections 3.2–3.3; Appendix E.3). The measured Setting-C endpoint is 6.74 ± 0.24 mmHg with CEA versus 11.64 ± 0.52 mmHg without the update (Table 1). Without a reference target, the evaluated deployment mode is frozen inference.

Here, reference ABP is available on the support windows. Intermittent cuff-only calibration and an unlabeled ECG–PPG-consistency objective are extensions beyond this experiment.

For a target task, the evaluated inner-loop update is

ϕctx(m)=ϕctx(m−1)−αin∇ϕctxLcalib(m),Lcalib=Lwave+λreg∥ϕctx−ϕctx∗∥22.\phi_{\mathrm{ctx}}^{(m)}=\phi_{\mathrm{ctx}}^{(m-1)}-\alpha_{\mathrm{in}}\nabla_{\phi_{\mathrm{ctx}}}\mathcal{L}_{\mathrm{calib}}^{(m)}, \qquad \mathcal{L}_{\mathrm{calib}}=\mathcal{L}_{\mathrm{wave}}+\lambda_{\mathrm{reg}}\lVert\phi_{\mathrm{ctx}}-\phi_{\mathrm{ctx}}^{\ast}\rVert_2^2.

Only the context branch is updated in this equation; CDSE, the event branch, decoder dynamics, and task heads remain fixed.

The same four-support/four-query protocol gives 6.74 ± 0.24 mmHg with CEA and 11.64 ± 0.52 mmHg without the update. Thus the equation specifies which parameters are adapted, while Table 1 quantifies the resulting deployment change.

Revision locations.

  • Main text §§3.2–3.3, pp. 4–5; Appendix E.3, p. 11 — describe CEA as supervised reference-ABP calibration and frozen inference as the no-target mode.
  • Main text Discussion, p. 9 — identify cuff-only and unlabeled consistency calibration as extensions beyond the current evaluation.

Q2: How does context-only adaptation compare with standard fine-tuning and other update choices?

A2: Setting C is designed to measure context-only adaptation against its no-update control (w/o CEA). Table 1 gives 6.74 ± 0.24 versus 11.64 ± 0.52 mmHg, a 4.90 mmHg (42.1%) reduction under four support windows, four disjoint query windows, and three inner steps (Main Text Section 5.1; Appendix E.5). The claim is a quantified CEA-over-no-update improvement under this support budget, rather than a ranking of every possible adaptation policy.

R2-Table1: Setting-C comparison.

ConditionQuery RMSE (mmHg)
No update (w/o CEA)11.64 ± 0.52
Context-only CEA6.74 ± 0.24

Both conditions use four support windows, four disjoint query windows, and three inner steps; the no-update condition skips the deployment-time context update. Under this support budget, the measured difference is 4.90 mmHg.

This comparison is intentionally a context-only intervention: it tests whether adapting the designated branch improves the same frozen model on the same query windows. It does not assign a universal ordering to full-model fine-tuning, event-only updates, or other update rules.

Revision locations.

  • Main text §5.1, p. 6; Appendix E.5, p. 13; Table 1 — describe w/o CEA as the Setting-C no-update control.

We thank Reviewer RezM for the questions on deployment targets, delay interpretation, and uncertainty.

Dear Reviewer RezM,

As the discussion period is nearing its end, we would appreciate hearing whether our responses and revisions address your concerns, and which issues you feel still need attention. We would welcome the opportunity to address any remaining concerns within the time available.

Thank you for your time and consideration.

Reviewer_uZZmOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
2
Presentation
1(−1)
Soundness
1(−1)
Read review
Paper summary

PhysioSplit studies ECG-PPG transfer through a bounded fractional-delay state-space encoder, a context/event representation, and context encoder adaptation (CEA). It reconstructs arterial blood pressure (ABP) waveforms and predicts blood-pressure-related phenotypes. VitalDB is the sole training source. Settings A and B are frozen transfers to MIMIC-III and UCI; Setting C uses supervised context calibration on a 32-subject wearable cohort, with four support windows, four disjoint query windows, and three inner updates. Setting D evaluates a fixed phenotype head against visit-level ICD-derived weak labels on MC-MED, separately from ABP-derived labels. The paper supports its performance claims with encoder ablations, delay diagnostics, representation reliability measures, and frozen probes (Section 4, p. 5; Appendices E-F, pp. 21-26).

Critical strengths
  1. An important transfer problem and a coherent architectural response. The paper addresses how a physiological model can retain useful temporal information while adapting to changes in subjects and acquisition conditions. This is relevant to ICLR readers interested in multimodal representations, distribution shift, and constrained personalization. The delay-aware state-space encoder, temporally pooled context branch, sequence-based event branch, and asymmetric decoder conditioning form a coherent set of inductive biases for that problem (Sections 3.1-3.2, pp. 3-5). The resulting adaptation interface is a meaningful contribution candidate. Appendix B.3 appropriately treats the factorization as a computational design rather than an identified physiological decomposition.

  2. Encouraging frozen external results, supported by several diagnostics. Table 1 (p. 6) reports RMSEs of 5.43 and 5.89 mmHg on A and B, compared with 8.03 and 9.67 mmHg for SiamQuality. These tests use no target calibration, so they provide evidence beyond the calibrated wearable setting. Appendix E.5 and Table S4 (p. 23) also include DANN, CORAL, and GroupDRO comparisons. The progressive encoder ablation, delay agreement, and context/event probes provide useful additional evidence about the proposed design (Tables 2-3, p. 8; Table S7, p. 26), although they do not settle the attribution questions below.

Critical weaknesses
  1. The distinctive value of the proposed information partition is not yet sufficiently established. The architecture is well motivated, and the encoder ablations and probes provide relevant supporting evidence. However, the main adaptation comparison does not clearly separate the benefit of learning the representation from the benefit of restricting deployment updates to context. Section 5.1 and Appendix E.5 describe w/o CEA as the no-update control, yet its frozen Set-A RMSE is 7.48 rather than the complete model's 5.43 (Table 1, p. 6). This could reflect different training conditions, so it is an attribution ambiguity rather than proof of an erroneous score. The reliability and probe results (Table 4, p. 9; Table S7, p. 26) show persistence and predictive information, but do not establish the comparative benefit of the chosen adaptation subspace. The authors acknowledge that global and event-only updates were not evaluated (Appendix E.5, p. 23). The resulting uncertainty concerns the paper's main architectural lesson: how much of the adaptation benefit depends on the proposed context/event partition and selective update? It does not negate the observed performance of the complete system.

  2. The external robustness evidence needs clearer population and evaluation boundaries. A/B assess frozen waveform transfer, C assesses supervised wearable calibration, and D uses ICD-derived phenotype labels. The manuscript appropriately distinguishes these regimes, but its headline external mean still combines frozen A/B with calibrated C (Table 1, p. 6; Appendix E.3, p. 22). The reported waveform SDs reflect three training runs and do not establish the consistency of paired gains across patients; the window-to-patient weighting also needs clarification (Appendix E.4, p. 23). The population underlying D is additionally unclear. Table S2 (p. 22) lists 70,545 MC-MED visits, whereas the original MC-MED data descriptor reports 70,545 unique patients and 118,385 visits. The manuscript may use a selected visit subset, so this is not evidence of an incorrect sample count. However, the relationship between the reported cohort and the source population needs to be explained. Clear patient/visit counts, separate deployment-regime summaries, and patient-level comparisons would make the scope and reliability of the transfer claim more convincing.

Overall, this is a relevant representation-learning study with a plausible architectural rationale and encouraging external results. The recommendation is weak rejection at present because the evidence does not yet make the distinctive adaptation benefit and the consistency of the cross-patient gains sufficiently clear. An identifiable adaptation comparison and patient-level result summaries would clarify how strongly the empirical gains support the central design claim and could materially improve the

Questions for authors
  1. What evidence isolates the value of selective context adaptation? Please clarify the training and deployment changes in w/o CEA, its relationship to w/o MAML, and whether the reported Set-C results include a before/after comparison from the same complete-model checkpoint on identical support/query episodes. If available, which comparison under matched support supervision and model-selection rules best supports updating context specifically? These results would determine how strongly the proposed information partition is supported as an adaptation design.

  2. Which populations and aggregation rules underlie the external results? For MC-MED, does 70,545 in Table S2 denote unique patients or retained visits? Please explain how the analyzed visits were selected from the source cohort, including whether one visit per patient was chosen, and report the retained patient, visit, and class counts. For the waveform evaluations, please specify the window-to-patient aggregation and weighting, and identify or provide paired patient-level improvements with uncertainty for frozen A/B and calibrated C separately, using the existing test predictions. These answers would clarify both the evaluation denominator and how consistently the gains hold across the evaluated populations.

6 comments
AuthorsOpenReview ↗

Response to Reviewer uZZm (Part 1/3: W1–W2)

Thank you for pointing out the distinction between source training, deployment-time context calibration, and population-level aggregation. We address each issue with the reported numbers and the estimand used in each setting.


W1: The distinctive value of the proposed context/event partition is not yet sufficiently isolated.

A: The partition is supported by two complementary results. First, the deployment comparison in Setting C gives 6.74 ± 0.24 mmHg with context-only CEA versus 11.64 ± 0.52 mmHg for w/o CEA (Table 1, Main Text Section 5.1), a 4.90 mmHg reduction (42.1%). Both conditions use the same selected checkpoint, support/query episode partition, and source-only model-selection rule; CEA updates only the context branch on four support windows, while the control evaluates the same query windows with zero adaptation steps.

Second, the representation diagnostics test whether the two codes carry different information. The complete representation reaches source ICC$(2,1)=0.812$, within-patient ICC =0.834=0.834, and RWW LOSO ICC =0.634±0.041=0.634\pm0.041 (Main Text Section 5.3, Table 4). Frozen probes predict age and BMI from the context code with R2=0.531R^2=0.531 and 0.4890.489, versus 0.0940.094 and 0.0780.078 from the event code; the RWW SBP probe gives AUROC 0.7120.712 for context and 0.5970.597 for event. These numbers support the intended division: context carries persistent subject-level information used by CEA, while event features retain short-time waveform information. A/B (5.43 ± 0.14 and 5.89 ± 0.17 mmHg) are frozen-transfer results and answer the separate transfer question.


W2: The external robustness evidence needs clearer population and evaluation boundaries.

A: The external results use three different evaluation units (Main Text Section 4.1; Appendix E.1–E.4; Table S2): A/B are frozen waveform transfer (5.43 ± 0.14, 5.89 ± 0.17 mmHg); C is supervised RWW calibration on 32 subjects; D is zero-shot, visit-level phenotype recognition on MC-MED. The MC-MED database contains 70,545 unique patients and 118,385 ED visits; Setting D uses the retained visits with valid three-class ICD-10 labels after the stated filters.

R1-Table2: Evaluation cohorts, analysis units, and target information.

SettingCohort and unitTarget information
AMIMIC-III-Ext-PPG, 2,371 subjectsFrozen transfer
BUCI cuff-less BP, 942 subjectsFrozen transfer
CRWW, 32 subjectsSupport-window ABP
DMC-MED, ED visits with ICD-10-derived labelsZero-shot phenotype

For waveform results, the reported pooled quantity is

RMSEpool=∑i∑t(y^it−yit)2∑iTi,\mathrm{RMSE}_{\mathrm{pool}}=\sqrt{\frac{\sum_i\sum_t(\hat y_{it}-y_{it})^2}{\sum_iT_i}},

so the unit of aggregation is the retained waveform sample. The phenotype result instead averages window probabilities within each visit before computing Macro-F1 (Appendix E.1 and E.4).

AuthorsOpenReview ↗

Response to Reviewer uZZm (Part 2/3: Q1)

Q1: What evidence isolates the value of context-selective adaptation?

A1: The intended Setting-C control is the w/o CEA no-update condition. In the full PhysioSplit procedure, episodic meta-training uses disjoint support/query episodes. The support loss updates only the context branch, while the query loss updates the shared parameters and the learned context initializer. At deployment, CEA updates only the context branch using four support windows and then evaluates four disjoint query windows after three inner steps. Under w/o CEA, this deployment-time context update is skipped and the model remains frozen on the query windows.

The deployment update is the context-only inner loop

ϕctx(m)=ϕctx(m−1)−αin∇ϕctxLcalib(m),m=1,2,3,\phi_{\mathrm{ctx}}^{(m)}=\phi_{\mathrm{ctx}}^{(m-1)}- \alpha_{\mathrm{in}}\nabla_{\phi_{\mathrm{ctx}}}\mathcal L_{\mathrm{calib}}^{(m)}, \qquad m=1,2,3,

whereas meta-training evaluates the disjoint query objective after that update,

min⁡Φ,ϕctx∗ETi[Ltotal(Qi;Φ,ϕctx,i(3))].\min_{\boldsymbol\Phi,\phi_{\mathrm{ctx}}^*} \mathbb E_{\mathcal T_i}\left[ \mathcal L_{\mathrm{total}}(Q_i;\boldsymbol\Phi, \phi_{\mathrm{ctx},i}^{(3)})\right].

This is the training/deployment distinction implemented in Main Text Eq. 13–14 and Appendix D.

(a) Reported endpoint values.

R1-Table1: Waveform RMSE for the no-update control and PhysioSplit.

VariantSourceSet ASet BSet C
w/o CEA4.29 ± 0.097.48 ± 0.228.97 ± 0.3111.64 ± 0.52
PhysioSplit3.87 ± 0.065.43 ± 0.145.89 ± 0.176.74 ± 0.24

The Set-C numbers compare context calibration with frozen inference. The A/B entries are frozen-transfer results and are kept separate from this deployment comparison. The Setting-C result therefore stands on its own.

(b) Training and deployment separation.

Setting C compares CEA with frozen inference under the documented support budget. The conclusion concerns this context-only interface.

For the Setting-C comparison, both conditions start from the same selected checkpoint and use the same support/query episode partition: the same four support windows and the same four query windows. CEA uses the support windows for three context-update steps; the no-update control ignores the support update and evaluates the checkpoint directly on the same query windows. Model selection uses the same source validation split and source-only early-stopping rule; query results are excluded from selection.

The w/o MAML variant is a separate training ablation: it removes the episodic support/query meta-training loop and the query-driven outer update, while retaining the model architecture and source-task losses. At deployment, w/o MAML still applies CEA to the context branch using support ABP; it is therefore distinct from w/o CEA, which disables the deployment-time context update.

Revision locations.

  • Main text §5.1, p. 6; Appendix E.5, p. 13; Table 1 — describe w/o CEA as the Setting-C no-update control.
  • Main text §§3.2–3.3, pp. 4–5; Figure 2, p. 4 — separate episodic meta-training from the deployment-time context update.
  • Main text §3.3, p. 5; Table 4, p. 9 — define w/o MAML as removal of episodic meta-training while retaining deployment-time CEA.
  • Main text §5.1, p. 6; Table 1 — state that Set-C is the direct CEA versus frozen-inference comparison and A/B are frozen-transfer results.
  • Main text §5.1, p. 6; Table 1 — state that Set-C is the direct CEA versus frozen-inference comparison and A/B are frozen-transfer results.

AuthorsOpenReview ↗

Response to Reviewer uZZm (Part 3/3: Q2)

Q2: Which populations and aggregation rules underlie the external results?

A2: Settings A and B use frozen waveform transfer without target support. Setting C uses reference ABP for supervised RWW calibration. Setting D performs zero-shot phenotype classification on ED visits with ICD-10-derived labels. The corresponding units and supervision are reported separately in Section 4.1 and Table S2.

The reported waveform endpoints are 5.43 ± 0.14 mmHg on A, 5.89 ± 0.17 mmHg on B, and 6.74 ± 0.24 mmHg on C; the corresponding external degradation ratio is 1.55 (Table 1). RWW calibration uses 32 leave-one-subject-out subjects, while A/B have no target support. D uses visit-level diagnosis labels rather than ABP-derived waveform targets, so its Macro-F1 is not pooled with the waveform RMSE values.

For waveform results, the reported pooled quantity is RMSEpool=∑i∑t(y^it−yit)2∑iTi\mathrm{RMSE}_{\mathrm{pool}}=\sqrt{\frac{\sum_i\sum_t(\hat y_{it}-y_{it})^2}{\sum_iT_i}}. The unit of aggregation is the retained waveform sample. The phenotype result instead averages window probabilities within each visit before computing Macro-F1 (Appendix E.1 and E.4).

Patient-level paired A/B/C confidence intervals use predictions indexed by patient. The current tables report seed-level variation, while the paired patient estimand is defined separately by the patient-level difference and bootstrap rule in Appendix E.4.

Δi=RMSEi,comparator−RMSEi,PhysioSplit,Δˉ=1N∑iΔi.\Delta_i=\mathrm{RMSE}_{i,\mathrm{comparator}}-\mathrm{RMSE}_{i,\mathrm{PhysioSplit}}, \qquad \bar\Delta=\frac{1}{N}\sum_i\Delta_i.

MC-MED construction. The MC-MED database contains 70,545 unique patients and 118,385 ED visits. The primary diagnosis fields are Dx_ICD9, Dx_ICD10, and Dx_name in visits.csv; MRN identifies patients and CSN identifies visits (Appendix E.1). Setting D is evaluated at the visit level: window probabilities are averaged within each retained labeled visit before Macro-F1. I10–I16 identify hypertension, I95.* identifies hypotension, visits without either family define the normotension reference group, and visits containing both families are excluded. These database totals are distinct from the post-filter ECG/PPG analysis cohort.

Revision locations.

  • Main text §4.1, p. 6; Appendix E.1, p. 11; Appendix E.4, p. 12; Table S2, p. 12 — report A/B, C, and D as separate estimands.
  • Appendix E.1, p. 11; Table S2, p. 12; Appendix E.4, p. 12 — distinguish source inventory from the visit denominator; distinguish paired intervals from seed variation.

We thank Reviewer uZZm for highlighting the attribution and evaluation-unit distinctions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their thoughtful and constructive feedback. We appreciate their recognition of the cross-cohort transfer results and the methodological value of connecting representation structure with deployment-time adaptation.

We have revised the main manuscript and supplementary material to clarify the adaptation comparisons, expand the delay analyses, and improve the description of the external evaluations. Below, we summarize our core contributions and the principal updates. Our individual responses address each reviewer’s comments and provide the corresponding revision locations.


Core Contributions of Our Work

  1. Delay-Aware Multimodal Fusion: PhysioSplit incorporates a bounded fractional ECG–PPG arrival-delay proxy into a state-space encoder to account for sub-sample temporal offsets during multimodal fusion.

  2. Context/Event Partition and Selective Adaptation: The model separates temporally pooled context from event features and uses the context pathway as the individual calibration interface. Deployment-time CEA updates this pathway while keeping the event branch and shared timing representation fixed.

  3. Cross-Cohort Evaluation: With VitalDB as the sole training source, PhysioSplit achieves frozen-transfer RMSEs of 5.43 and 5.89 mmHg on MIMIC-III and UCI, compared with 8.03 and 9.67 mmHg for SiamQuality. The evaluation also covers small-support calibration on RWW and zero-shot phenotype recognition on MC-MED.


Updates of experimental results during Rebuttal

  • Main Text, Table 1; Appendix E.5: Clarified the Setting-C comparison using the same selected checkpoint, support/query episodes, and model-selection rule. Context-only CEA reduces query RMSE from 11.64 ± 0.52 to 6.74 ± 0.24 mmHg, a 42.1% reduction. CEA performs three context-update steps, while the control performs zero updates on the same checkpoint and evaluates identical query windows (uZZm, yxpk).

  • Appendix E.7: Presented delay-bound sensitivity results for 100–400, 80–420, and 150–350 ms, with Setting-C RMSEs of 6.74 ± 0.24, 6.55 ± 0.30, and 7.00 ± 0.45 mmHg, respectively (RezM).

  • Representation Analyses: Expanded the explanation connecting reliability measurements and frozen probes to the context/event partition. Context probes predict age and BMI with R² = 0.531 and 0.489, compared with 0.094 and 0.078 for event probes; the RWW SBP probe achieves AUROC 0.712 from context and 0.597 from event (uZZm, yxpk).


Updates of in-depth discussions during Rebuttal

  • Main Text; Appendix E.5: Distinguished deployment-time CEA from episodic meta-training. Clarified that w/o CEA disables the Setting-C context update, whereas w/o MAML removes episodic meta-training while retaining deployment-time CEA. Explained the separate interpretation of frozen A/B transfer and calibrated Setting C (uZZm, yxpk).

  • Appendix E.3: Specified reference ABP from four support windows as the supervision used in the evaluated CEA procedure, identified the parameters updated during calibration, and described frozen inference when target supervision is unavailable (RezM).

  • Main Text, Eq. 2–3; Appendix D.1: Expanded the CDSE indexing discussion, distinguishing PPG-history reads from causal ECG-delay pairing. Provided integer and fractional examples that specify the delayed signal, read indices, and output timestamps, including ECG[50]–PPG[70] pairing and a 20.5-sample delay (yxpk).

  • Main Text; Appendix E.4: Defined the R-peak-to-PPG-foot reference interval and explained the delay-agreement results as consistency with signal landmarks (RezM).

  • Appendix E.1–E.4: Clarified cohort descriptions, MC-MED patient/visit terminology, waveform aggregation, and visit-level probability averaging. Distinguished variation across three training runs from patient-level paired uncertainty (uZZm, RezM).


These revisions strengthen the presentation of PhysioSplit’s architectural design, experimental evidence, and deployment procedures. The main manuscript and supplementary material have been updated accordingly, with the revisions highlighted in blue.


We thank all three reviewers again for their valuable feedback and consideration.

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

Related papers

Discussion (0)

Sign in to comment.