acceptodds
Under review as a conference paper at ICLR 2027

Multi-Resolution Generation of Electronic Health Records

Abstract

Generative models for electronic health records aim to synthesize plausible future clinical trajectories conditioned on a patient’s prior history. These approaches support a range of downstream applications, including risk stratification and phenotype discovery. However, most existing methods generate trajectories at a predetermined resolution - on the order of hours or days - making long-horizon forecasting computational expensive and error-prone. The few approaches that incorporate multiple temporal resolutions scale poorly to multiple outcomes of interest and require costly training procedures. Thus, we present Multiple Resolution Generation of Electronic Health Records (MuRGEHR – pronounced merger), a framework for the efficient generation at multiple resolution over long horizons without sacrificing fidelity. MuRGEHR models the EHR time series as a hierarchical sequence of tokens representing different temporal resolutions. The model first constructs a coarse scaffold, which then serves as conditioning for synthesizing progressively finer resolutions. Across three datasets and multiple prediction horizons, this hierarchical strategy yields consistent gains in computational efficiency while remaining competitive with regards to fidelity on downstream clinical tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
electronic health records, healthcare, clinical prediction making

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_m4EHOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims to provide a novel multi resolution generation of electronic health records. This would be an important and difficult problem to solve. The authors aim to provide an hierachical framework and apply it to some datasets.

Critical strengths

Great Problem to work on and at least the high level motivation of an hierachical approach is nice.

Critical weaknesses

Several. I am not even sure where to start and keep it short since this work is simply not ready and should be significantly polished before it is to be submitted again.

  • Check AI hallucinated References
  • Check that they are correct
  • Check that there is a title
  • Check that there are not visual artefacts from AI tools
  • Check that the AI statement at least is coherent
  • The authors create their own baselines.
  • I do not understand the uncertainty analysis at all.

The paper has significant presentation and formatting issues.

Questions for authors

Why submit something that is clearly not ready ...

1 comment
  1. Can the reviewer point to specific places in the manuscript where they believe that hallucinated references were included?
  2. While the title was included for the OpenReview submission, it was not specifically included in the uploaded PDF. We apologize for this oversight and have updated the resubmitted PDF with the title.
  3. Can the reviewer point to specific places in the manuscript where they believe that there are visual artifacts?
  4. The AI statement was written as Yes/No to answer the precise questions asked. We recognize that the final produced AI statement is unclear, and therefore we have modified the text accordingly.
  5. We respectfully disagree with the reviewer's assessment. ETHOS (one of our baseline models) was not developed by our group. In addition, we have included two additional, previously published baselines in response 1.2.
  6. It would be helpful to point out what aspects of the uncertainty analysis is unclear as we are happy to clarify. We have made minor clarifications.
  7. We are happy to address specific concerns with respect to presentation and formatting, if the reviewer can provide more details.
Reviewer_jve6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents MuRGEHR, a generative model that represents patient records at monthly, weekly, and daily resolutions. It generates coarse summaries first and adds finer detail when needed, reducing the computation required to predict events over longer periods. Outcome probabilities are estimated from the frequency of events across generated trajectories.

The main contribution is an efficient, unified approach to generation at multiple temporal resolutions. The authors evaluate it on MIMIC-IV, EHR-SHOT, and a private heart-failure dataset, predicting abnormal laboratory results, death, and subsequent healthcare visits. Comparisons include single-resolution models, an event-stream model, and a retrained ETHOS model (compared to orginal trained on larged set of MIMIC data).

The results demonstrate faster generation. Equal-time comparisons use more trajectories for MuRGEHR without establishing that either model’s AUC has stabilized. Six-month calibration is worse than ETHOS, and the experiments use a restricted, outcome-focused vocabulary, limiting conclusions about performance on comprehensive EHR data.

Critical strengths

In appeoaches such as ETHOS or like, estimating outcome probabilities requires generating many patient trajectories, making generation speed critical. MuRGEHR’s ability to generate coarse summaries and selectively add finer detail substantially reduces this cost. This is a practical contribution because faster generation allows more precise probability estimation within a given computational budget. Authors demosntarte substential effciency gains with similar efficiency across 3 datasets.

Critical weaknesses

Models are evaluated using a small vocabulary concentrated on selected laboratory tests and clinical events. Although this provides a controlled comparison, it leaves unclear whether the reported efficiency–accuracy trade-off extends to comprehensive records containing diagnoses, medications, procedures, and richer measurements. This limits the generality of the paper’s contribution.

In equal-time comparisons, MuRGEHR generates substantially more trajectories than ETHOS. More samples can improve estimated probabilities and measured AUC without improving the underlying model. The paper does not show how AUC changes with sample count or whether it stabilizes for either method. Consequently, these results support computational efficiency but do not establish superior underlying discrimination. Worse six-month calibration further limits claims of preserving predictive quality.

Questions for authors

I think it would be important to show how AUC valries with number of generated trajectories and is AUC stabilizes. Reported predictive advantage is likely result of higher number of trajectores used.

It is unclear how ECE was calculated, how many bins ete, authors should explain this more.

It would be great to see if efficiencey-accuracy is true for larger vocabulary.

Bacuase of different tokenisation I am concern about the lengh of the context information for different methods. In particular did event-streamer and ETHOS had the same information as MuRGEHR when starting to generate?

5 comments
  1. We thank the reviewer for the comment and have included a section in the manuscript to specifically address this point (5.2). We plot the Macro AUROC as a function of generation wall-clock time in figures 4 and A.7. We notice across all datasets, that MuRGEHR achieves higher macro AUROC at shorter wall times and the difference gets closer with longer wall times. When we plot the macro AUROC for the same number of trajectories in figures A.5-6, MuRGEHR slightly underperforms ETHOS on MIMIC and EHRShot and performs similarly to ETHOS on Hospital A. Additionally we notice that MuRGEHR seems to converge beyond around 300 trajectories, whereas ETHOS is still increasing at 100 trajectories. Nevertheless, the computational cost associated with generating more trajectories for ETHOS makes it impractical and expensive.
  2. We thank the reviewer for the comment and have added the following description to the appendix (C.4). Each discharge’s predicted probability is the fraction of its generated trajectories in which the outcome of interest occurred before the time horizon of interest. The discharges are sorted by predicted probability and split into 10 equal count bins. We then calculate the ECE as the bin size weighted mean absolute difference between the mean predicted probability and the observed event rate in each bin. Additionally, for calibration analysis, we would like to point the reviewer to response .
  3. We thank the reviewer for the comment and would like to point the reviewer to response 1.3 for results on the larger vocabulary size.
  4. We thank the reviewer for the question. MuRGEHR and all of the binned models receive the 280 days of context prior to discharge in their corresponding resolutions, with MuRGEHR receiving information at all resolutions. Our event-stream model that we implemented takes the 280 most recent tokens from the 280 days of context (pre-discharge. ETHOS takes the 280 most recent tokens from the entire patient record, anchored at discharge. Thus, it is possible for the event stream model to have strictly less information than MuRGEHR, and ETHOS may have more or less information than MuRGEHR. We have included this in Appendix C2.

I appreciate authors efforts and thank for clarrifications.

Thank you for the detailed responses and additional experiments. I appreciate the authors' effort in addressing my comments.

The new AUROC-versus-trajectory-count results are particularly informative. They show that, at equal numbers of trajectories, MuRGEHR slightly underperforms ETHOS on MIMIC-IV and EHRSHOT and achieves similar performance on Hospital A. This reinforces my interpretation that the principal contribution is computational efficiency rather than improved underlying predictive performance. This is a valuable contribution in itself, but I believe the distinction should be made explicit in the paper.

My concern regarding comparable historical information also remains. Providing MuRGEHR with 280 days of patient history while restricting ETHOS to 280 tokens does not necessarily provide comparable clinical information. Depending on event density and tokenization, these inputs may represent substantially different amounts of patient history. I would appreciate clarification of why ETHOS was not provided with all available events from the same observation window, subject to its context-length limitations.

My concern regarding comparable historical information also remains. Providing MuRGEHR with 280 days of patient history while restricting ETHOS to 280 tokens does not necessarily provide comparable clinical information. Depending on event density and tokenization, these inputs may represent substantially different amounts of patient history. I would appreciate clarification of why ETHOS was not provided with all available events from the same observation window, subject to its context-length limitations.

Overall, I recognize the computational contribution of the work. However, I believe that claims about predictive performance should be carefully stated, and the limitations of the baseline comparisons made explicit. I want to keep my original score.

Reviewer_fG2HOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes MuRGEHR, a multi-resolution generative model for EHR trajectories. A 504-day window anchored at discharge is discretized into time bins at three resolutions (month, week, and day), and each bin is tokenized as the sum of the embeddings of the codes that occur in it. Generation proceeds from coarse to fine, and each finer resolution predicts a residual correction that is added to the logits of its coarser parent. With a per-chunk attention mask and greedy allocation, the model covers a given horizon with the fewest possible tokens. The authors evaluate the model on MIMIC-IV, EHRSHOT, and a private dataset (Hospital A) for abnormal labs, death, readmission, and other outcomes. They report that MuRGEHR achieves AUROC comparable to ETHOS (Renc et al., 2024) with only about five generated tokens per trajectory, and that it matches or exceeds ETHOS under an equal wall-clock budget.

Critical strengths

Well-motivated problem. The generation cost of event-stream models grows with event density, while fixed-grid models are tied to a single predetermined resolution. Both limitations are real obstacles for rollout-based zero-shot prediction, and handling multiple resolutions within a single model is a reasonable motivation.

Coarse readouts are used directly at inference. As the authors point out, prior hierarchical generative models do not directly evaluate their coarse-level outputs or use them for prediction. This paper uses the coarse-resolution outputs themselves for prediction and consistently builds the design around a per-chunk mask that minimizes the number of generated tokens. Generating more than 20× fewer tokens per trajectory than ETHOS is of practical value.

Critical weaknesses

W1. The benefit of the multi-resolution design is not sufficiently supported

  • No task requires fine resolution: The evaluation horizons (28, 30, 42, and 168 days) are mostly multiples of 7 or 28, and day tokens are used only for two days in the 30-day horizon. The contributions claim that "clinically salient intervals can be synthesized at high resolution," but no method is presented for selecting clinically important intervals to generate at high resolution, and all performance results are obtained with greedy allocation, which favors the coarsest resolution. Demonstrating the benefit of multiple resolutions would require evaluations such as short horizons (e.g., 1, 3, or 10 days), the accuracy of event timing, and the ordering of events.
  • Key baselines are missing: The most direct comparison for the core claim of "a single generative model over nested resolutions" is MoRGEN (Oufattole et al., 2026), which combines per-resolution models. It is missing from the baselines even though it has already been published at ICML 2026. In addition, the paper claims to be "achieving competitive performance on clinical tasks while using significantly less compute resources," but it does not compare against models that predict probabilities directly without rollout, such as the time-to-event model MOTOR (Steinberg et al., 2024) or task-specific supervised baselines. Because these models predict with a single forward pass and no generation, they are the strongest competitors in terms of efficiency. Without them, the efficiency–performance trade-off cannot be judged in its full context.

W2. The reduced vocabulary limits the generality of the conclusions

According to Appendix B, the vocabulary was reduced to 74–430 codes by keeping only the evaluated outcomes, admissions and discharges, five labs, death and birth codes, and age and demographic information. This setting raises two concerns. First, real-world EHRs contain thousands of codes or more, with much higher event density. It is unclear whether bin-level sum pooling and independent per-code sigmoids still work in such settings. Second, because the vocabulary was chosen in advance to match the evaluated outcomes, the key advantage of generative models, namely that endpoints need not be specified in advance, is diluted.


W3. Calibration consistently degrades at long horizons

At the 6-month horizon, the macro ECE of MuRGEHR is substantially worse than that of the event-stream model and ETHOS on all three datasets.

6-month macro ECE (↓)monthevent streamETHOSMuRGEHR (100)MuRGEHR (equal time)
MIMIC (Table A.10).038.021.042.059.059
EHRSHOT (Table A.11).099.045.047.098.099
Hospital A (Table A.12).024.025.017.062.062

In contrast, at 28 days the ECE of MuRGEHR is lower than or comparable to that of the event-stream model and ETHOS. This problem is difficult to attribute to sampling noise or to a general limitation of grid-based approaches, for two reasons. First, the ECE barely changes even in the equal wall-clock setting, where the number of trajectories is increased to 643–1098 (MIMIC .059→.059, EHRSHOT .098→.099, Hospital A .062→.062). This indicates a systematic bias in the probabilities output by the model rather than sampling variance. Second, among the single-resolution grid models, the week and day models are also poorly calibrated at 6 months, but on MIMIC and Hospital A, MuRGEHR is worse even than the month model with the same architecture (.059 vs. .038 and .062 vs. .024).

Since risk estimation is the main intended use, calibration is as important as AUROC, yet the main text does not discuss these results. One might argue that post-hoc recalibration can fix this, but recalibration requires labeled data for each endpoint. This conflicts with the motivation in the Introduction of estimating risk through rollouts alone without specifying endpoints in advance. The authors should analyze why calibration degrades at long horizons, for example with reliability diagrams, and clarify whether the probabilities are overestimated or underestimated.


References

  • Oufattole, N., McDermott, M., and Stultz, C. MoRGen: Mixture-of-Resolutions Generative Forecasting for Irregularly Sampled Medical Time-Series Data. In International Conference on Machine Learning (ICML), 2026.
  • Renc, P., Jia, Y., Samir, A. E., Was, J., Li, Q., Bates, D. W., and Sitek, A. Zero shot health trajectory prediction using transformer. npj Digital Medicine, 7(1), 2024.
  • Steinberg, E., Fries, J. A., Xu, Y., and Shah, N. MOTOR: A time-to-event foundation model for structured medical records. In International Conference on Learning Representations (ICLR), 2024.
Questions for authors

See weaknesses above.

8 comments
  1. We thank the reviewer for this helpful comment. MuRGEHR’s principal advance is the efficient outcome estimation over long horizons, and we have clarified and emphasized this in the manuscript. Many challenging clinical questions involve such horizons, including six month mortality and 30-day heart failure readmission. At shorter horizons, our results indicate that MuRGEHR remains competitive with fine-resolution models, rather than consistently outperforming them. We have added evaluations at 1, 3, and 10 days (Tables A.6-8, which are summarized below). On MIMIC, MuRGEHR does not outperform the day model or ETHOS at 100 trajectories. Under matched wall-clock time, it generates only slightly more trajectories than ETHOS, consistent with the limited efficiency advantage over a 10 day rollout. On Hospital A, all three models perform similarly at 100 trajectories across the evaluated horizons. With 197 trajectories under matched wall-clock time, MuRGEHR outperforms the other models at all three horizons; the larger trajectory count reflects the use of an A100 rather than a V100. On EHRSHOT, the day model performs worse than ETHOS at every horizon and worse than MuRGEHR at 3 and 10 days. At 100 trajectories, MuRGEHR’s performance approaches that of ETHOS as the horizon increases. Taken together, these results show that a single multi-resolution model can remain competitive at short horizons, while its main advantage lies in longer-horizon estimation.

Macro AUROC at short horizons. Per-outcome results are in Tables A.6–A.8.

DatasetHorizonDayEvent streamETHOSMuRGEHR (100)MuRGEHR (matched wall-clock)
MIMIC-IV1 day.818.806.813.768.779 (110 traj.)
3 days.840.810.823.790.801
10 days.835.819.828.802.807
EHRSHOT1 day.799.794.865.760.764 (120 traj.)
3 days.751.768.811.794.795
10 days.757.773.810.806.807
Hospital A1 day.967.958.967.968.969 (197 traj.)
3 days.910.897.908.909.914
10 days.898.890.893.902.908
  1. We thank the reviewer for highlighting these important comparisons. We have added comparison with MoRGEN and MOTOR in Appendix A.3 (Tables A9-10, which are summarized below). Both methods were originally presented and evaluated in supervised settings, rather than zero-shot approaches. We clarify the evaluation setting and limitations of each comparison below.

    MoRGEN uses rollouts from generative models operating at different temporal resolutions to train a classifier for pre-specified tasks, combining the trajectories through a supervised mixture of experts. On MIMIC-IV, MoRGen achieves a slightly higher macro AUROC than MuRGEHR (at matched wall-time to ETHOS) at 30 days, and the two are tied at 6 months. On EHRSHOT, MuRGEHR with 100 trajectories slightly underperforms MoRGEN at 30d and exceeds MoRGEN with matched wall-clock time; on the 6 month horizon, MuRGEHR outperforms MoRGEN at both 100 trajectories and matched wall-clock time. On Hospital A, MuRGEHR outperforms MorGEN at 100 trajectories and matched wall-clock time at all horizons. We notice that MoRGEN is better calibrated especially at long horizons (Table A.11), likely due to its supervised mixture-of-experts on each task.

    MOTOR is originally trained on codes disjoint from the evaluation codes, followed by a linear probe or fine-tuning trained on the evaluation codes. For our zero-shot comparison, we instead include the evaluation codes in MOTOR’s training set. This adaptation was neither proposed nor evaluated in the original paper, and we make this distinction explicit when reporting the results.

    At 100 trajectories, MuRGEHR outperforms MOTOR on EHRSHOT at both the 30-day and six-month horizons, and on Hospital A at the 30-day horizon. Additional trajectories are needed to outperform MOTOR on MIMIC at both horizons and to match its performance on Hospital A at six months.

    We acknowledge that a single-pass model such as MOTOR has lower inference cost. However, MOTOR addresses a narrower prediction target: the time to the first occurrence of a given code. In our evaluations, it is also less well calibrated on the large datasets across all evaluated horizons (Table A11). In contrast, autoregressive models produce trajectories that can subsequently be queried for properties such as event recurrences, counts, and ordering. We have therefore clarified that our efficiency claim is relative to other autoregressive models used for zero-shot predictions, not to single-pass supervised predictors. Macro AUROC at 30 days and 6 months. Per-outcome results are in Tables A.9–A.10. MoRGen was retrained on our datasets and is supervised on each task; all other methods are zero-shot. Matched wall-clock = the time ETHOS takes to generate 100 trajectories.

DatasetHorizonMuRGEHR (100)MuRGEHR (matched wall-clock)ETHOS (100)MOTOR (zero-shot)MoRGen (supervised, 100)
MIMIC-IV30d.829.840 (643).836.833.844
6m.823.828 (643).828.827.828
EHRSHOT30d.807.825 (707).820.784.814
6m.807.814 (707).815.753.797
Hospital A30d.884.890 (1,098).883.875.871
6m.846.851 (1,098).846.851.838

Macro ECE (lower is better) at 100 trajectories. See Table A.11 and Tables A.25–A.27.

DatasetMOTOR 30dMoRGen 30dMuRGEHR 30dMOTOR 6mMoRGen 6mMuRGEHR 6m
MIMIC-IV.052.012.015.064.023.059
EHRSHOT.055.023.035.095.029.098
Hospital A.048.024.019.075.034.062
  1. We thank the reviewer for raising this important point. To address this concern, we have added evaluations with substantially larger vocabularies. We retain the original codes and add all codes observed in at least 1,000 training patients for MIMIC and Hospital A, and at least 100 training patients for EHRSHOT. This yields vocabularies of 12,370, 24,441, and 5,736 codes for MIMIC, Hospital A, and EHRSHOT, respectively. The results are reported in Appendix A.4 and summarized below.

    MuRGEHR maintains strong predictive performance with these larger vocabularies (Tables A12-14), which are reproduced below. On MIMIC, at 100 trajectories, it performs best at every horizon except 28 days, where it is slightly behind the event-stream model; with additional trajectories (corresponding to equal wall-clock time), it performs best at all evaluated horizons. On EHRSHOT, it outperforms all baselines at 100 trajectories, and the advantage increases with additional trajectories. On Hospital A, it performs best among the evaluated baselines, including ETHOS, at 100 trajectories at every horizon except six months, where the month model performs slightly better. The larger vocabularies do, however, substantially reduce MuRGEHR’s latency advantage: in the time ETHOS takes to generate 100 trajectories, MuRGEHR generates 156 on MIMIC, 222 on EHRSHOT, and 228 on Hospital A. This reduction reflects the dense implementation of each bin’s bag representation. A sparse implementation offers scope for further optimization; we measured approximate speedups of 1.5x on EHRSHOT, 2.5x on MIMIC, and 4x on Hospital A. These results support MuRGEHR’s predictive performance with larger vocabularies while highlighting the importance of implementation efficiency at this scale.

Macro AUROC with the expanded vocabularies (MIMIC-IV: 12,370 codes; EHRSHOT: 5,736; Hospital A: 24,441). Per-outcome results are in Tables A.12–A.14. "–" = horizon not representable at that resolution.

DatasetHorizonMonthWeekDayEvent streamETHOSMuRGEHR (100)MuRGEHR (matched wall-clock)
MIMIC-IV30d––.800.833.784.834.835
6m.818.809.784.810.752.821.823
EHRSHOT30d––.734.731.773.789.796
6m.777.739.724.705.745.784.790
Hospital A30d––.870.860.874.884.888
6m.852.836.815.814.817.849.851
  1. We thank the reviewer for the comment. We agree that calibration is very important and have included a section in the manuscript discussing this point (5.1). We also concur with the reviewer that it is not a sampling noise problem. In our case, a multi-resolution approach coupled with independent sigmoid decoding results in over-estimating the occurrence of some codes over long time horizons. For example, under independent sigmoid decoding, P(empty) = ∏(1−p_c), which shrinks quickly when many codes have moderate probability. We plot reliability diagrams at 28 days and 168 days for MIMIC (Fig 2). These data demonstrate that risk-overestimation is greater at the 6-month horizon relative to the 28-day horizon.

    To explicitly demonstrate that multi-resolution and the other binned models tend to overestimate code-occurrence, we compare the model’s predicted probability that the next bin is empty with the observed fraction of empty bins. Given the true patient history via teacher forcing, the probability that the next bin is empty is far lower than the real rate of empty bins (Fig. 3). Similar trends hold for the other datasets shown in Fig. A1-4.

    With respect to the etiology of MuRGEHR’s over-predictions over long time horizons, we note that MuRGEHR was trained where each bin (independent of resolution) contributes the same amount to the loss function; e.g., the month resolution head receives only 3% of the training signal. While this does not yield worse discrimination, we hypothesize that this is what leads to the overpredictions that compound over the six-month rollout. Possible solutions may be reweighting the month head or including an empty bin as part of the vocabulary. We leave this for future work.

    We recognize that over-estimating risk at long time horizons is a limitation of the approach, and we highlight this in our limitations section (Section D). It should be noted, however, that for the most deleterious outcomes (e.g. death, heart failure etc.) over-estimating risk is less harmful than underestimating risk, as under-estimating risk of such adverse events may delay beneficial therapies to those patients at highest risk. The trade-off being that risk over-estimation may lead to unnecessary follow-up procedures for some patients.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.