acceptodds
Under review as a conference paper at ICLR 2027

PULSE: Self-Supervised Clinical Phase Discovery for Multimodal ICU Disease Progression Modeling

Abstract

ICU disease progression often evolve through clinically distinct phases, characterized by different physiological states and diagnostic and treatment needs. These phases, however, are not explicitly delineated in routinely collected data and must be inferred from asynchronous multimodal observations. Observation intensity also varies substantially over time, leaving some phases densely sampled and others sparsely observed. Models that directly aggregate observed events can therefore be biased to densely sampled phases in the learned trajectory representations, potentially obscuring the underlying clinical dynamics. We propose , a self-supervised framework that explicitly discovers latent clinical phases from asynchronous multimodal observations. These phases are shared across patients and modalities and provide a common structure for representing evolving patient states. To reduce sensitivity to variations in observation intensity and timing, encourages consistent phase assignments under sampling perturbations, including event duplication, timestamp jitter, and measurement removal. We evaluate on MIMIC and eICU across five dynamic prediction tasks: AKI progression, sepsis progression, vasopressor initiation, intubation need, and remaining length of stay. achieves the best overall performance among the evaluated baselines, supporting the value of phase-aware representations for modeling and predicting evolving patient states from irregular multimodal disease trajectories.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Phase Discovery, Disease Progression Modeling, Self-supervise Learning, Multimodal longitudinal Data Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_c6jnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

PULSE is a self-supervised framework for multimodal ICU progression modeling. It combines knowledge-informed phase prototypes, equal-mass hourly optimal transport alignment, and phase-aware attention to learn shared clinical representations. Experiments on MIMIC-IV and eICU report the best mean performance in nine of ten task settings, supported by ablations, history-length analyses, and clinical case studies.

Critical strengths
  1. Coherent design for a clinically relevant problem: Knowledge-informed phase prototypes and equal-mass hourly alignment offer a well-motivated approach to modeling asynchronous ICU trajectories while addressing observation-intensity bias.
  2. Broad empirical gains: PULSE achieves the highest reported mean performance in nine of ten settings across two datasets, with particularly substantial improvements in AKI progression prediction.
Critical weaknesses
  1. Uncertainty about baseline fidelity: Appendix A.5.1 removes CTPD’s reconstruction decoder and loss, as well as parts of DiPro’s architecture and training objectives. These adaptations may be justified, but their impact is not evaluated, leaving uncertainty about whether the comparisons preserve the baselines’ competitiveness.
  2. Insufficient validation of shared clinical phases: Section 3.3 presents two retrospectively interpreted cases but does not systematically establish whether the same prototype has consistent clinical meaning across patients. This limits support for the central claim of discovering shared clinical phases beyond generic physiological state clusters.
Questions for authors

1. Baseline adaptation clarification: Appendix A.5.1 attributes the omission of DiPro’s auxiliary objectives to the lack of disease progression labels. What motivates removing CTPD’s reconstruction decoder and loss? What validation supports the competitiveness of these adapted baselines?

2. Clinical phase validation clarification: Beyond the two case studies, what evidence supports consistent clinical meanings of the same phase prototype across patients? Is there any cohort-level analysis or independent clinical assessment of this consistency?

2 comments

Thank you for recognizing the motivation and empirical results of our method and for your thoughtful comments and questions. We address each point below.

W1 & Q1: Clarification of Adoptation of the Baseline CTPD and DiPro

Thank you for the constructive comments. As described in Appendix A.5.1, all seven baselines are implemented on a shared input interface, with identical landmark examples, modality streams, and task labels, and each retains its characteristic fusion mechanism. Five baselines (UMSE, UTDE, TAMME, FuseMoE, and Flex-MoE) retain all of their original components and training objectives. Only CTPD and DiPro requires the omission of components that are not applicable to our setting:

  1. CTPD: We removed the reconstruction decoder and reconstruction loss. CTPD uses frozen pretrained representations for reconstruction supervision. Extending it to our jointly trained CXR and ECG encoders would require additional design choices regarding the reconstruction targets and decoder.

  2. DiPro: We removed progression-aware loss, since it requires additional fine-grained disease progression labels, which are unavailable in our data. We also removed the temporal-consistency loss, since it requires at least three consecutive events within the same modality, which most of our samples do not contain.


W2 & Q2: Clinical phase validation clarification

Thank you for this insightful suggestion. We evaluated the shared phase prototypes through cohort-level analysis and independent review by clinical experts.

Cohort-level analysis (Table R13). We analyzed AKI and Sepsis trajectories in the MIMIC test set using the pretrained PULSE model without downstream fine-tuning. For each stay we summarized its phase assignments into a phase-exposure vector and tested whether adding it to standard clinical covariates improved prediction of subsequent progression (multinomial logistic-regression likelihood-ratio test, Holm-corrected). As shown in Table R13, phase assignments are significantly associated with progression for both tasks, showing that the shared phases carry predictive clinical information.

Table R13. Cohort-level association between phase assignments and subsequent AKI and Sepsis progression across MIMIC test stays (multinomial logistic-regression likelihood-ratio test, Holm-corrected).

TaskTest ICU staysHolm adjusted p
AKI4,6710.0031
Sepsis1,1212.06 × 10⁻⁵

Independent clinical assessment (Table R14 & Table R15). A clinical expert and three LLMs independently reviewed the phases generated by PULSE for 30 randomly sampled MIMIC test stays, alongside the raw EHR time series, ICU notes, discharge summaries, and phenotype labels, rating the phases on a five-point Likert scale for two aspects:

  • Disease-trajectory alignment: Whether the phases reflect the main changes in a patient's disease trajectory;
  • Clinical understanding: Whether the phases help explain the patient's disease progression and clinical course.

As shown in Table R14, most stays were rated agree or strongly agree by the clinical expert, and three independent LLM reviewers produced consistent ratings (Table R15), suggest that the discovered phases are clinically meaningful.

Table R14. Case-level review by clinical expert.

Strongly DisagreeDisagreeNeutralAgreeStrongly Agree
Disease trajectory alignment006222
Clinical understanding004251

Table R15. Case-level review by clinical expert and three independent LLM reviewers over 30 MIMIC test stays (Mean ± SD).

EvaluatorDisease trajectory alignmentClinical understanding
Clinical expert3.87 ± 0.513.90 ± 0.40
DeepSeek-v4-pro3.93 ± 0.373.90 ± 0.66
Gemini-4-argon3.80 ± 0.413.93 ± 0.25
GPT-6.1-sol3.67 ± 0.483.70 ± 0.70

Full experimental details of both analyses have been added to Appendix A.3.2.


Summary of Manucript Revisions

We have revised the manuscript accordingly to address your comments and suggestions, with all modifications highlighted in blue:

  1. Clarified the adaptations to CTPD and DiPro and their rationale. (See Appendix A.6.1)
  2. Added quantitative phase validation through cohort-level analysis and an independent clinical assessment. (See Appendix A.3.2)

Thank you again for your valuable comments and suggestions!

Dear Reviewer c6jn,

With the discussion period nearing its end, we are following up to ensure we have addressed all of your concerns. Please let us know if you have any remaining questions or further feedback.

Thank you again for your valuable time and insights.

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

The paper proposes PULSE, a self-supervised framework that discovers latent clinical phases from multimodal ICU data. Phase descriptions generated by LLM are embedded and clustered to initialize a shared phase bank and to define a clinical geometry among phases. Each hour of a stay is then assigned to phases with an unbalanced fused Gromov-Wasserstein OT that gives every hour equal mass, which the authors argue mitigates observation-intensity bias. The resulting assignments are used as pseudo-labels for hourly and event representations, and the frozen phase bank guides temporal attention during fine-tuning. The method is evaluated on MIMIC-IV and eICU across five dynamic prediction tasks, with ablations.

Critical strengths
  • I think the problem is well motivated. Observation intensity in multimodal ICU data reflects care processes as well as patient state, and the paper explicitly targets this issue rather than treating it as a side effect.
  • The evaluation is broad, covering two datasets and five dynamic prediction tasks, and PULSE performs best in most settings, with the clearest gains on disease progression tasks.
Critical weaknesses
  • The gains are not clearly attributed to the claimed mechanism. The paper argues that PULSE uses asynchronous modalities better than event-level models, but there is no PULSE variant without event modalities. Uni-EHR is a plain Transformer apparently without pretraining, so PULSE’s advantage over it may come from self-supervised pretraining and the phase structure rather than from better use of events. This is especially unclear on eICU, where the only events are notes, yet the gains over Uni-EHR are among the largest. In addition, several baselines were implemented with key components removed, so the observation that many baselines fall below Uni-EHR may partly reflect weakened implementations.
  • The phase space, which is central to the method, is underspecified and not validated. The number of phases (12) appears only in the appendix, without any rationale or sensitivity analysis. The paper does not report how the 53 LLM generated descriptions are grouped into 12 clusters, so the clinical meaning of each phase cannot be checked, and the claim that phases are anchored in clinical knowledge rests on two case studies with LLM-generated post-hoc interpretations reviewed by a clinician. The preprocessing is also unclear. The hourly EHR aggregation is not described, and if values are forward-filled, the physiological cost used in the OT would partly reintroduce observation density into the phase assignment.
Questions for authors
  1. How does PULSE perform without event modalities (EHR only, with the same pretraining)?
  2. How was the number of phases chosen, and how sensitive are the results to it?
  3. How are the hourly EHR bins constructed?
  4. YAIB defines AKI and sepsis as hourly onset prediction, not 12-hour progression. Why use 12-hour landmarks, and how do the labels relate to YAIB?
  5. The OpenReview abstract describes consistency training under sampling perturbations, which does not seem to appear in the paper. Could you clarify?
6 comments

Thank you for your constructive feedback and the recognition of our motivation and experiments. We address your concerns in detail below:

W1.1 & Q1: Contribution of asynchronous modalities and EHR-only PULSE ablation

Thank you for your insightful comment. To attribute the gains, we ran an Uni-EHR PULSE that keeps the phase pretraining but drops the event modalities (Table R8).

  1. Uni-EHR PULSE removes the event modalities while keeping the pretraining and the phase structures; across both datasets it is at parity with Uni-EHR Baseline for all tasks other than the LoS task.

  2. Full PULSE (multimodal) achieves significant performance improvements for most of the tasks, compared with Uni-EHR PULSE.

Therefore, for most tasks, the performance gains are mainly attributed to the better use of event modalities. For LoS task, the pre-training contributes more to the performance improvement.

Table R8. Ablation study on Uni-EHR PULSE pre-training.

DatasetModelsAKI F1Sepsis F1Vaso PR-AUCIntub. PR-AUCLoS Kappa
MIMICUni-EHR Baseline0.458±0.0030.341±0.0030.560±0.0030.911±0.0040.402±0.006
Uni-EHR PULSE0.466±0.0040.330±0.0080.554±0.0010.918±0.0030.416±0.002
Full PULSE0.478±0.0020.380±0.0100.576±0.0050.920±0.0040.411±0.008
eICUUni-EHR Baseline0.433±0.0060.365±0.0130.496±0.0050.903±0.0020.475±0.005
Uni-EHR PULSE0.471±0.0120.364±0.0120.487±0.0070.896±0.0020.491±0.004
Full PULSE0.494±0.0090.381±0.0190.534±0.0100.932±0.0020.492±0.004

W1.2: Clarification of Baseline Adaptations.

Thank you for the constructive comments. As described in Appendix A.5.1, all seven baselines are implemented on a shared input interface, with identical landmark examples, modality streams, and task labels, and each retains its characteristic fusion mechanism. Five baselines (UMSE, UTDE, TAMME, FuseMoE, and Flex-MoE) retain all of their original components and training objectives. Only CTPD and DiPro requires the omission of components that are not applicable to our setting:

  1. CTPD: We removed the reconstruction decoder and reconstruction loss. CTPD uses frozen pretrained representations for reconstruction supervision. Extending it to our jointly trained CXR and ECG encoders would require additional design choices regarding the reconstruction targets and decoder.

  2. DiPro: We removed progression-aware loss, since it requires additional fine-grained disease progression labels, which are unavailable in our data. We also removed the temporal-consistency loss, since it requires at least three consecutive events within the same modality, which most of our samples do not contain.


W2.1 & Q2: Number of Phases Rationale and Sensitivity Analysis

We appreciate the reviewer pointing this out. Given that soft phasing already supports a large expressive space, we selected a moderate K = 12 empirically. The sensitivity analysis over K (Table R9) shows that PULSE remains better than or competitive with the best baseline across all evaluated K. In addition, K = 12 achieves the best AKI and Sepsis F1, while larger K provides further gains on more complex tasks such as intubation, vasopressor need, and LoS. This pattern suggests that disease-specific prediction tasks are well represented by a moderate number of phases, whereas tasks involving multiple interacting conditions benefit from finer-grained phase representations.

Table R9. Sensitivity of PULSE to the number of phases KK on MIMIC. The final row reports the best baseline results for comparison.

Number of phasesAKI F1Sepsis F1Vaso PRAUCIntub. PRAUCLoS Kappa
80.468±0.0050.375±0.0080.573±0.0130.919±0.0030.398±0.010
120.478±0.0020.380±0.0100.576±0.0050.920±0.0040.411±0.008
160.463±0.0050.370±0.0080.579±0.0100.919±0.0010.418±0.005
200.463±0.0050.356±0.0120.598±0.0020.925±0.0020.417±0.007
240.459±0.0060.370±0.0100.595±0.0020.924±0.0040.425±0.010
Best baseline reults0.458±0.0030.361±0.0030.586±0.0030.911±0.0030.402±0.006

W2.2: Details of Phase Bank Construction and Clincial Interpretation

We thank the reviewer for this request and now report the clustering in full detail.

Construction of the prototype bank. We L2-normalized each description's 768-dimensional MedEmbed-base CLS embedding, subtracted the shared mean direction, and renormalized. We then applied spherical k-means with cosine similarity, deterministic farthest-point initialization, and up to 50 iterations. The resulting K synthetic centroids define the fixed medical phase-similarity graph and are used to initialize the K learnable phase tokens.

Clinical interpretation. We have generated a complete cluster-membership report for all 53 phases discriptions. Table 20 in the appendix lists each member and a descriptive summary of the members' shared theme. One representative cluster is the Cluster 5 (Acute respiratory illness and deterioration). Its nine members include COPD exacerbation, acute lower respiratory illness, pneumonia, pleural compromise, and developing or established respiratory failure. Their shared theme is respiratory distress, seen in impaired oxygenation/ventilation and escalating respiratory support.


W2.3: The Clinical Phase Validation Clarification

Thank you for encouraging a deeper validation of the clinical phases. We have added two-level quantitative evaluations, both confirming their clinical meaningfulness:

Case-level review (Table R10 & R11). A clinical expert and three LLMs independently reviewed the phases generated by PULSE for 30 randomly sampled MIMIC test stays, alongside the raw EHR time series, ICU notes, discharge summaries, and phenotype labels, rating the phases on a five-point Likert scale for two aspects:

  • Disease-trajectory alignment: Whether the phases reflect the main changes in a patient's disease trajectory;
  • Clinical understanding: Whether the phases help explain the patient's disease progression and clinical course.

As shown in Table R10, most stays were rated agree or strongly agree by the clinical expert, and three independent LLM reviewers produced consistent ratings (Table R11), suggest that the discovered phases are clinically meaningful.

Table R10. Case-level review by clinical expert.

Strongly DisagreeDisagreeNeutralAgreeStrongly Agree
Disease trajectory alignment006222
Clinical understanding004251

Table R11. Case-level review by clinical expert and three independent LLM reviewers over 30 MIMIC test stays (Mean ± SD).

EvaluatorDisease trajectory alignmentClinical understanding
Clinical expert3.87 ± 0.513.90 ± 0.40
DeepSeek-v4-pro3.93 ± 0.373.90 ± 0.66
Gemini-4-argon3.80 ± 0.413.93 ± 0.25
GPT-6.1-sol3.67 ± 0.483.70 ± 0.70

Cohort-level analysis (Table R12). We analyzed AKI and Sepsis trajectories in the MIMIC test set using the pretrained PULSE model without downstream fine-tuning. For each stay we summarized its phase assignments into a phase-exposure vector and tested whether adding it to standard clinical covariates improved prediction of subsequent progression (multinomial logistic-regression likelihood-ratio test, Holm-corrected). As shown in Table R12, phase assignments are significantly associated with progression for both tasks, showing that the shared phases carry predictive clinical information.

Table R12. Cohort-level association between phase assignments and subsequent AKI and Sepsis progression across MIMIC test stays (multinomial logistic-regression likelihood-ratio test, Holm-corrected).

TaskTest ICU staysHolm adjusted p
AKI4,6714,6710.00310.0031
Sepsis1,1211,1212.06×10−52.06 × 10^{-5}

Full experimental details of both analyses have been added to Appendix A.3.2.


W2.4 & Q3: EHR Preprocessing and Concerns of Bias Introduced by Forward-filled

We thank the reviewer for raising this point and address the two parts separately.

  1. EHR preprocessing. Hourly bins follow DiPro's preprocessing: irregular EHR records are assigned to one-hour intervals relative to ICU admission, the most recent record is kept when several fall in the same hour, and missing values are forward-filled. Each variable is then standardized using statistics computed only from observed values.

  2. The observation mask prevents forward-fill bias. Thank you for pointing out the potential impact of forward-filled values. In our implementation, we used a binary observation mask to denote if the value is directly observed (mask = 1) or imputed (mask = 0). When computing the physiological cost, we use this observation mask to exclude the forward-filled values to prevent observation density from leaking into phase assignment. We have included relevant details in the revised paper.

Q4: AKI Disease Progression Labels and the Rationale for 12-Hour Dynamic Landmarks

We thank the reviewer for this question and address the two parts respectively.

  1. We reuse YAIB's underlying labels but define a different target. For AKI, we define the 12-hour progression labels as improved / unchanged / worsened based on the change in maximum KDIGO stage from the preceding to the subsequent 12-hour window. For sepsis, we analogously define 12-hour progression by the change in peak SOFA score from the preceding to the subsequent window, requiring a difference of at least two points for improved or worsened.

  2. The 12-hour horizon reflects a clinically meaningful reassessment interval. It matches the KDIGO staging thresholds: stage 2 requires urine output below 0.5 mL/kg/h for at least 12 hours, and stage 3 requires anuria for at least 12 hours[1]. And it matches response-driven reassessment in sepsis and acute respiratory failure[2,3]. We therefore use 12-hour landmarks to evaluate near-term, dynamically updated risk, rather than a single forecast from a fixed early window.

References:

[1] Khwaja et al. "KDIGO clinical practice guidelines for acute kidney injury." Nephron Clinical Practice, 2012.

[2] Evans et al. "Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021." Intensive care medicine, 2021.

[3] Rochwerg, et al. "Official ERS/ATS clinical practice guidelines: noninvasive ventilation for acute respiratory failure." European Respiratory Journal, 2017.


Q5: OpenReview Metadata Abstract Discrepancy

Thank you for catching this discrepancy. The submitted PDF reflects our final method, while the OpenReview metadata abstract was an earlier draft. We apologize for this oversight and will correct the metadata to match the final method described in the PDF.


Summary of Manucript Revisions

In response to your comments and suggestions, we have made the following updates. Changes to the manuscript are highlighted in blue:

  1. Added the Uni-EHR PULSE ablation and accompanying analysis. (See Section 3.4, “Ablation Study.”)
  2. Clarified the adaptations to CTPD and DiPro and their rationale. (See Appendix A.6.1)
  3. Added a sensitivity analysis of the number of phases. (See Section 3.5)
  4. Detailed the construction and clinical interpretation of the phase bank. (See Appendix A.6)
  5. Added quantitative clinical phase validation at both the case and cohort levels. (See Appendix A.3.2)
  6. Expanded the EHR preprocessing details in Appendix A.3 (“Details of Data Preprocessing”) and clarified how the observation mask is incorporated into the physiological cost in Section 2.3 (the “Equal-mass phase assignment” paragraph).
  7. Updated the OpenReview metadata abstract to match the manuscript.

Thank you once again for your insightful review and feedback!

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

This paper describes PULSE (Self-Supervised Clinical Phase Discovery for Multimodal ICU Disease Progression Modeling), a framework designed to discover shared, latent clinical phases from asynchronous multimodal intensive care unit (ICU) trajectories. The paper tackles two core clinical machine learning challenges:

  1. Unsupervised discovery of latent clinical phases without explicit phase annotations.
  2. Mitigation of observation-intensity bias, where event-level aggregation disproportionately weights periods with frequent clinical documentations or diagnostic procedures. PULSE operates in two main stages: Stage 1 (Self-Supervised Pretraining): Initializes a shared phase prototype bank B anchored by clinical knowledge descriptions generated from a medical LLM (Med-Gemini) and embedded via MedEmbed. It then aligns fused hourly representations (Z^fused) with the phase prototypes using Unbalanced Fused Gromov-Wasserstein (FGW) optimal transport, enforcing uniform source mass (mu_l = 1/L) across all hourly bins. The transport couplings serve as pseudo-labels to train phase distillation losses (L_phase and L_event). Stage 2 (Downstream Fine-Tuning): Fine-tunes the encoders and fusion modules while keeping the phase bank frozen, injecting learned phase assignments into a phase-aware temporal attention layer for downstream dynamic prediction tasks. The authors evaluate PULSE across five dynamic tasks on MIMIC-IV and eICU (AKI progression, Sepsis progression, Vasopressor initiation, Intubation need, and remaining Length of Stay) against eight baselines.
Critical strengths
  1. Important Formulation of Observation-Intensity Bias: Identifying how irregular, event-level multimodal data aggregation causes models to conflate clinical workflow/sampling intensity with patient physiological state is a critical insight for healthcare ML. The baseline finding that an EHR-only model (Uni-EHR) outperforms complex multimodal baselines in 47 out of 70 dataset-task settings highlights this issue.
  2. Methodological Novelty in Optimal Transport Application: Using Unbalanced Fused Gromov-Wasserstein (FGW) optimal transport to decouple observation intensity from phase contribution is theoretically sound and well-formulated. Constraining the source marginal to uniform mass (mu_l = 1/L) while relaxing the target marginal with a KL penalty ensures that every hour contributes equally to phase allocation regardless of event density.
  3. Knowledge-Informed Phase Regularization: Leveraging LLM-derived clinical phase descriptions to induce a pairwise clinical prior geometry matrix (C^P) provides structured regularization (L_prior), preventing the unconstrained representation collapse common in pure data-driven clustering/state-space models.
  4. Comprehensive Empirical Evaluation: Testing across two large multi-center/single-center databases (MIMIC-IV, eICU), five distinct prediction tasks, varying input history lengths (h in [36, 96] hours), and multiple ablations demonstrates rigorous benchmarking.
Critical weaknesses
  1. Conceptual Tension: Substituting "Observation-Intensity Bias" with "Uniform-Time Bias" The Argument: The paper argues that event-level aggregation biases models toward densely sampled periods, proposing instead that every hourly bin receive equal source mass (mu_l = 1/L) in optimal transport. Why it is Weak: In ICU clinical reality, patient physiology does not evolve uniformly across time. An hour characterized by sudden septic collapse, acute hemodynamic instability, or cardiac arrest contains far higher physiological information density than ten consecutive hours of stable, sedated recovery. By enforcing mu_l = 1/L, PULSE treats one hour of uneventful sleep identically to one hour of acute resuscitation in the OT coupling formulation. This introduces a unexamined uniform-time bias.

  2. Fragility and Over-Reliance on LLM Prior Geometry (K=12) The Argument: The phase prototypes are initialized and regularized by K cluster centers derived from Med-Gemini descriptions of clinical phenotypes embedded via MedEmbed.
    Why it is Weak: (i) Arbitrary & Global Prototype Space (K=12): The manuscript uses a single global phase bank of K=12 prototypes shared across all patients and diseases (e.g., Sepsis, CABG, AKI, Traumatic Brain Injury). It is conceptually implausible that 12 global states adequately capture the nuanced progression phases of every ICU condition simultaneously. (ii) LLM & Embedding Sensitivity: Med-Gemini text prompts can vary significantly with minor template alterations. The paper provides no sensitivity analysis regarding the choice of LLM (e.g., Med-Gemini vs. GPT-4o vs. Llama-3-70B-Instruct) or text embedding model (MedEmbed vs. BioLinkBERT/Clinical-Longformer). (iii) Lack of Quantitative Phase Validation: In Section 3.3, clinical phase alignment is validated purely qualitatively via LLM (Codex 5.6) narrative generation and informal clinician inspection. This subjective approach lacks quantitative rigor.

  3. Algorithmic Complexity and Stability of Pretraining The Argument: PULSE uses an iterative block-coordinate framework during pretraining, solving a 30-step FGW optimal transport problem per stay to recompute pseudo-label assignments T* alongside model weight updates. Why it is Weak: FGW optimal transport is non-convex and sensitive to hyperparameter tuning (alpha, beta, sigma_time, lambda_pi, lambda_prior, lambda_event). The manuscript does not discuss training stability, convergence behavior, or computational overhead (wall-clock time per epoch compared to standard baselines). There is an unaddressed risk of alternating optimization oscillation between the continuous representations Z^fused and the discrete-transport coupling T*.

  4. Selective Superiority in Main Benchmark The Argument: PULSE achieves the "best overall performance among evaluated baselines". Why it is Weak: On MIMIC-IV Vasopressor need, UTDE achieves 0.586 AUPRC vs. PULSE's 0.576 AUPRC. On eICU Vasopressor need, UTDE (0.524 AUPRC) and TAMME (0.517 AUPRC) are very close to or competitive with PULSE (0.534 AUPRC). The performance gains on intervention tasks (Vasopressor, Intubation) and Length of Stay (LoS) are relatively modest (0.8%–3.4%) compared to disease progression tasks. The text should explicitly analyze why short-term intervention prediction benefits less from global phase alignment than long-term disease progression.

Questions for authors

Listed above in weakness.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors

We thank the reviewer for the thoughtful and constructive comments, and we appreciate the positive feedback of our work. Below, we address each point in detail:

W1: Solving Uniform-Time Bias

We agree that ICU physiology is non-uniform, and we would like to clarify what μℓ=1/L\mu_\ell=1/L does. It only gives every hour the same budget to distribute over the phases, so that an hour does not dominate the prediction simply because it contains more recorded events. It does not determine which phase an hour is assigned to, nor how important that hour is for prediction.

  1. Phase assignment is determined by the content of each hour. The assignment R∗R^* is governed by the fused features together with the physiological and temporal costs. Consequently, an hour of acute resuscitation and an hour of stable recovery receive different phase assignments, although both carry equal mass.

  2. The importance of each hour for prediction is learned during fine-tuning. The temporal attention weights each hour according to the downstream task, so that clinically critical hours can receive greater weight than stable ones.

Equal source mass therefore removes the influence of observation counts on phase assignment, while the distinction between critical and stable hours is preserved in both the phase assignments and the downstream prediction.

W2.1: Discussion on Phase Bank Granularity (KK) and Empirical Sensitivity Analysis

Thank you for the constructive comments. The KK prototypes are not intended to enumerate the phases of each disease. They form a shared basis of general clinical states, and each hour is represented as a soft mixture Rℓj∗R^*_{\ell j} over this basis. Disease-specific progression is therefore expressed through the composition and temporal ordering of the shared prototypes, rather than through dedicated prototypes for each condition.

Empirically, we further conduct a sensitivity analysis, and the results (Table R1) show that PULSE remains competitive across the evaluated values of KK (8-24) for most tasks.

Table R1. Sensitivity of PULSE to the number of phases KK on MIMIC. The final row reports the best baseline results for comparison.

Number of phasesAKI F1Sepsis F1Vaso. AUPRCIntub. AUPRCLoS Kappa
K=8K = 80.468 ± 0.0050.375 ± 0.0080.573 ± 0.0130.919 ± 0.0030.398 ± 0.010
K=12K = 120.478 ± 0.0020.380 ± 0.0100.576 ± 0.0050.920 ± 0.0040.411 ± 0.008
K=16K = 160.463 ± 0.0050.370 ± 0.0080.579 ± 0.0100.919 ± 0.0010.418 ± 0.005
K=20K = 200.463 ± 0.0050.356 ± 0.0120.598 ± 0.0020.925 ± 0.0020.417 ± 0.007
K=24K = 240.459 ± 0.0060.370 ± 0.0100.595 ± 0.0020.924 ± 0.0040.425 ± 0.010
Best Baseline0.458 ± 0.0030.361 ± 0.0030.586 ± 0.0030.911 ± 0.0030.402 ± 0.006

W2.2: LLM and Text Embedding Model Sensitivity

We conducted additional sensitivity experiments on the MIMIC dataset using different LLMs (Med-Gemini, GPT-4o, Llama-3-70B-Instruct) and embedding models (MedEmbed, BioClinicalBERT, Clinical-Longformer). As shown in Table R2, PULSE remains stable with minor variations across different LLMs and embedding models, and consistently outperforms or remains competitive with the best baseline results.

Table R2. Sensitivity of PULSE to different LLMs and text embedding models on MIMIC.

ModelLLMEmbedding ModelAKI F1Sepsis F1Vaso AUPRCIntub. AUPRCLoS Kappa
PULSE (Default)Med-GeminiMedEmbed0.478 ± 0.0020.380 ± 0.0100.576 ± 0.0050.920 ± 0.0040.411 ± 0.008
M1GPT-4oMedEmbed0.470 ± 0.0120.382 ± 0.0170.576 ± 0.0200.921 ± 0.0080.398 ± 0.010
M2Llama-3-70B-InstructMedEmbed0.479 ± 0.0040.370 ± 0.0090.592 ± 0.0040.924 ± 0.0020.411 ± 0.014
M3Med-GeminiBioClinicalBERT0.484 ± 0.0030.383 ± 0.0100.588 ± 0.0070.921 ± 0.0030.407 ± 0.005
M4Med-GeminiClinical-Longformer0.471 ± 0.0050.378 ± 0.0080.583 ± 0.0110.923 ± 0.0000.403 ± 0.003
Best Baseline——0.458 ± 0.0030.361 ± 0.0030.586 ± 0.0030.911 ± 0.0030.402 ± 0.006

W2.3: Quantitative Phase Validation

We thank the reviewer for the constructive comments. We further conduct case-level review and cohort-level analysis.

Case-level review (Table R3). A clinical expert independently reviewed the phases generated by PULSE for 30 randomly sampled MIMIC test stays, alongside the raw EHR time series, ICU notes, discharge summaries, and phenotype labels, rating the phases on a five-point Likert scale for two aspects:

  • Disease-trajectory alignment: Whether the phases reflect the main changes in a patient's disease trajectory;
  • Clinical understanding: Whether the phases help explain the patient's disease progression and clinical course.

As shown in Table R3, most stays were rated agree or strongly agree, suggest that the discovered phases are clinically meaningful.

Table R3. Case-level review by clinical expert.

Strongly DisagreeDisagreeNeutralAgreeStrongly Agree
Disease trajectory alignment006222
Clinical understanding004251

Cohort-level analysis (Table R4). We analyzed AKI and Sepsis trajectories in the MIMIC test set using the pretrained PULSE model without downstream fine-tuning. For each stay we aggregate its phase assignments into a static vector and tested whether adding it to standard clinical covariates improves prediction of subsequent progression (multinomial logistic-regression likelihood-ratio test, Holm-corrected). As shown in Table R4, phase assignments are significantly associated with progression for both tasks, showing that the shared phases carry predictive clinical information.

Table R4. Cohort-level association between phase assignments and subsequent AKI and Sepsis progression across MIMIC test stays (multinomial logistic-regression likelihood-ratio test, Holm-corrected).

TaskTest ICU staysHolm adjusted p
AKI4,6710.0031
Sepsis1,1212.06 × 10⁻⁵

We have included this results as well as the complete experiment details in Appendix A.3.2 of the revised manuscript.

W3: Optimization stability, Hyperparameter Sensitivity and Computational Overhead

Training Stability Analysis (Table R5). We thank the reviewer for raising the concern about oscillation in the alternating optimization. We evaluated the stability of PULSE pretraining across five random seeds. As shown in Table R5, pretraining converges consistently: the final losses are comparable across seeds, and they do not collapse onto a subset of phases (the average number of active phases remains approximately 11.8 out of 12). Regarding the interaction between the representations and the transport coupling, we would like to clarify that the pseudo-labels are continuous soft assignments R∗R^* rather than discrete labels. Changes in the representations therefore lead to gradual adjustments of R∗R^* instead of abrupt re-assignments. In addition, the target geometry CPC^{\mathcal{P}} is fixed throughout pretraining, and Lprior\mathcal{L}_{\mathrm{prior}} keeps the phase bank anchored to it, which further stabilizes the alternating updates.

Table R5. Training stability of PULSE pretraining across five random seeds.

SeedFinal total lossFinal phase lossFinal event lossNumber of active phases / 12
20271.44101.13151.547711.788
20281.30051.00961.454611.813
20291.22170.94901.363211.828
20301.25910.91481.721311.806
20311.44541.14941.479911.863
Mean ± SD1.3335 ± 0.10391.0309 ± 0.10581.5134 ± 0.133711.820 ± 0.028

Hyperparameter Sensitivity (Table R6). We evaluated OT-related hyperparameters on the MIMIC AKI task. PULSE's performance is comparatively stable under changes to λevent\lambda_{\mathrm{event}} and λprior\lambda_{\mathrm{prior}}, with moderate sensitivity to β\beta and σtime\sigma_{\mathrm{time}}. The most sensitive parameter is α\alpha, which balances feature similarity and relational structure in the FGW objective. Higher values emphasize structural consistency in phase assignment; the default α=0.8\alpha=0.8 lies in a favorable region, with little change upon further increase. λπ\lambda_{\pi} is also sensitive: larger values more strongly regularize the phase marginal toward the uniform prior, preventing collapse onto a few phases.

Table R6. Sensitivity of AKI prediction to one-factor-at-a-time hyperparameter perturbations.

ParameterValueAKI F1ParameterValueAKI F1
α\alpha0.70.450 ± 0.004λπ\lambda_{\pi}0.0250.457 ± 0.008
0.80.478 ± 0.0020.050.478 ± 0.002
0.90.476 ± 0.0020.10.466 ± 0.007
β\beta0.350.466 ± 0.004λprior\lambda_{\mathrm{prior}}50.472 ± 0.004
0.50.478 ± 0.002100.478 ± 0.002
0.650.465 ± 0.003200.474 ± 0.006
λevent\lambda_{\mathrm{event}}0.10.472 ± 0.003σtime\sigma_{\mathrm{time}}30.464 ± 0.001
0.20.478 ± 0.00260.478 ± 0.002
0.40.469 ± 0.003120.466 ± 0.007

Computational Overhead (Table R7). We compare the computational efficiency and inference cost on the MIMIC dataset, the AKI task. PULSE exhibits the second-lowest GFLOPs and the lowest measured forward time, alongside higher predictive accuracy. PULSE benefits from two key computational advantages: (1) The OT solver is only used in pretraining stage and is omitted during downstream fine-tuning and inference. (2) The pretraining stage is task-agnostic and performed only once for all downstream tasks. Therefore, downstream training and inference stay fast and lightweight.

Table R7. Computational efficiency comparison between PLUSE and baseline models.

Model / measurement scopeParameters (M)Forward GFLOPs/sampleWall-clock (s/epoch)Forward time (ms/sample)AKI F1
PULSE — downstream40.1018.822554.131.4150.478 ± 0.002
DiPro39.0089.140548.171.4820.439 ± 0.003
FlexMoE39.2488.618757.021.9920.434 ± 0.004
CTPD45.5579.400874.155.4210.441 ± 0.001
UMSE51.4009.559646.711.7570.436 ± 0.003
TAMME43.1019.510647.671.4450.426 ± 0.003
FuseMoE269.0289.169626.141.7360.432 ± 0.001
UTDE43.2608.931666.501.9450.446 ± 0.003

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.