acceptodds
Under review as a conference paper at ICLR 2027

Efficient Generative Prediction for EHR Foundation Models: The SCOPE and REACH Estimators

Abstract

Generative foundation models trained on tokenized electronic health record (EHR) timelines show promise for clinical outcome prediction via Monte Carlo sampling of simulated future trajectories. However, this approach suffers from three coupled limitations: sparse estimate distributions that poorly differentiate patient risk levels, extreme computational cost, and high sampling variance. We propose two new estimators that leverage next-token probability distributions underutilized by standard Monte Carlo: the Sum of Conditional Outcome Probability Estimator (SCOPE) and Risk Estimation from Anticipated Conditional Hazards (REACH). We prove both are unbiased, that REACH guarantees variance reduction over Monte Carlo for any model and outcome, and that REACH is a Rao-Blackwellization of any naive importance sampling scheme that preserves the non-outcome token distribution. Empirically, across clinically important outcomes in MIMIC-IV and the UChicago health system, SCOPE and REACH match -sample Monte Carlo accuracy with median token reductions of to and reductions exceeding for the rarest outcomes, with calibration preserved throughout. Because SCOPE reuses a single sampled pool across an arbitrary number of outcomes at no marginal generation cost while REACH provides a per-task variance guarantee, the two estimators are complementary in deployment and together meaningfully reduce the inference budget required for generative EHR foundation models, particularly for rare, high-impact outcomes in healthcare.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Monte Carlo, Sampling, Electronic Health Records, Transformers, Efficiency, Unbiased Estimators, Generative outcome prediction, rare event prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_oQmHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes two estimators to make risk prediction with autoregressive EHR foundation models simulation more efficient. Such models estimate risk by Monte Carlo (MC) sampling: the risk is the fraction of sampled future timelines in which the outcome token appears. The authors prove both SCOPE and REACH estimators are unbiased. They also show that REACH has lower variance than MC and SCOPE for the same number of sampled timelines

Critical strengths
  1. The work has rigorous theoretical grounding. Each estimator comes with formal proofs of unbiasedness and variance properties. The derivations and equations are detailed, so a reader can follow every step. Some core results are also machine-verified in Lean 4, which is uncommon in this area.

  2. The supplementary material contains full proofs, a notation table, worked examples and a counterexample. It also includes synthetic Markov-chain experiments, calibration and Brier-score analyses, and full result tables with confidence intervals.

Critical weaknesses
  1. The citation style does not follow the ICLR-required format.

  2. ETHOS, the core prior work, is never properly introduced. The introduction refers to “reproducing the pipeline of an existing paper (Appendix I)” without naming it. The problem statement also says: “In a standard workflow, electronic health records are represented as token sequences.” This is not accurate. EHR data are mostly stored and modeled as tables. Turning them into event-token sequences is specific to ETHOS-style generative models. Without an introduction to ETHOS, the narrative is cut off for a general reader: it is unclear why risk should be estimated by sampling future token sequences.

  3. The main motivation is the computational cost of autoregressive inference. However, there are many established ways to cut this cost without changing the estimator. For example, KV caching, quantization, prefix sharing across the 100 timelines of the same patient etc. The 1,500+ GPU-hour figure is estimated from one hour of the original ETHOS-ARES code (Appendix I). The paper never reports the cost of its own SGLang pipeline, which already includes KV and prefix caching. So it is unclear how much of the motivating cost standard engineering already removes, and how much SCOPE/REACH save on top of it. Please discuss these techniques as complementary to the proposed estimators. Please also report GPU hours and per-patient latency on the same hardware for the original ETHOS implementation, an optimized MC baseline, and SCOPE/REACH.

  4. Figure quality is low and the font sizes are too small to read.

  5. Sampling settings are missing. Temperature and top-k/top-p are not reported. Unbiasedness requires sampling exactly from the model distribution whose probabilities SCOPE and REACH use, so these settings need to be stated.

  6. The relation to classical estimators is missing, and there is no importance-sampling baseline. SCOPE is a cumulative-hazard (compensator) estimator and REACH is conditional Monte Carlo, both classical tools. Appendix Q analyses importance sampling but never runs it as a baseline.

Questions for authors
  1. How does each of the 11 outcomes map to a token (or set of tokens)? Were computed label tokens injected into the training data, and are labels derived from tokens or from the raw structured data?

  2. Under the Brier criterion, SCOPE needs more tokens than MC on 5 of 11 outcomes. How does this fit with recommending SCOPE as the default, and with “calibration preserved throughout”?

  3. Which of the theoretical results are verified in Lean 4: is it only the three listed in Section 4.1, or also Appendix H and Appendix Q?

2 comments
AuthorsOpenReview ↗

Official Comment to Reviewer oQmH (Critical Weaknesses 1-6a)

(Edited 11/18/2026 to elaborate on point 3) We thank reviewer oQmH for their thorough and relevant comments. We respond to the critical critiques and questions in order below:

1 - Citation Style

We apologize for this oversight where references were being formatted from an incorrect template and we have corrected it. See our reply to reviewer 3HgR for further details.

2 - Introduction of generative prediction framework

We agree that generative inference (e.g. as in ETHOS) should be more strongly motivated in this paper. In addition to the changes described in our reply to reviewer 3HgR, we have modified the earlier language in the introduction to highlight the particularities and advantages of this prediction method and where they differ from standard supervised approaches. [Lines 43-46, 51-53]

3 - Engineering and Hardware

We agree that discussing how much of this cost was removed by our inference engineering would strengthen the paper. However, we would like to emphasize that all of our existing efficiency calculations measure how much SCOPE and REACH save on top of these engineering changes, as they are all calculated in terms of tokens per patient, a hardware and pipeline agnostic unit. In order to increase the clarity of our motivation, we have expanded this motivating quote to include the figure that, using our own SGLang inference pipeline, computing 100 timelines for the MIMIC-IV held-out set of over 96k patients required 495.8 A100 GPU hours, and have also added per patient latencies for both pipelines [Lines 69-73]. These figures cannot be directly compared to the ETHOS figure in Appendix J, since the experiments of that paper utilize a different model, tokenization scheme, cohort, and so on. We believe these figures strengthen the motivation of our work as it reveals that the cost remains significant even after substantial engineering.

To further address the hardware costs specific to the estimators (created by storing and updating the estimators, potential differences in timeline length caused by the outcome avoiding distribution REACH samples from, and others), we have also added a hardware cost analysis to the appendix, which measures the per-token cost of each method on an RTX 5090. The per-token overhead is negligible for SCOPE (+0.2%, 95% CI [−0.8, +1.3]) and small for REACH (mean of +2.4%, [+1.8, +3.0] across all outcomes, worst is mortality with +9.1% [6.3%, 11.8%]), far smaller than the median 2.5× to 3.4× token reductions reported in the paper. This is a valuable additional analysis that continues to support the existing conclusions [Appendix C]. Regarding the requested same-hardware comparison against the original ETHOS implementation, we did not include ETHOS, as its implementation is closely tied to their choice of model, making a fair same-model comparison impossible.

4 - Figure Quality

We greatly appreciate this feedback. We have made considerable efforts to improve the main figures in the body of the paper. We increased font sizes where possible, re-sized and re-arranged elements, and replaced all PNGs with PDFs to ensure quality remains high regardless of scaling. [Figure 1, Figure 2, Figure 3]

5 - Sampling Settings

We agree that using any form of top p/k or temperature based sampling would bias the resulting estimator. No top p/k sampling was performed and temperature was kept as 1.0 throughout. This fact is now noted in the methods. [Lines 367-368]

6 - Relationship to classical estimators and IS baseline

a. Relationship to classical estimators

We agree that the connection to classical statistical estimators is important and can be made more explicit. The cumulative hazard observation is greatly appreciated. An additional search uncovered a paper describing a "total hazard" estimator for Markov Chains that matches SCOPE much more closely than our previous reference. The related work section has been updated to use this new reference. [Lines 240-241]

We agree that REACH is an instance of conditional Monte Carlo (Rao–Blackwellization) constructed in an augmented outcome-avoiding probability space. This connection is already described in terms of Rao-Blackwellization in Section 3, and REACH is explicitly derived as a Rao–Blackwellized estimator in Section 2.2 [Lines 205-206] and Appendix Q. Nevertheless, we agree that explicitly naming CMC improves the positioning of our contributions, and we now name this relationship explicitly in the manuscript. [Lines 234-235]

AuthorsOpenReview ↗

Official Comment to reviewer oQmH (Critical Weakness 6b and Questions)

6b. Importance-sampling baseline

Since importance sampling is a general estimation framework rather than a uniquely specified estimator, implementing an importance-sampling baseline requires selecting a proposal distribution, and the choice of proposal can substantially affect both variance and computational efficiency. We agree that empirical comparisons with well designed importance-sampling proposals could be informative.

However, our theoretical analysis already establishes a stronger result than an empirical comparison with any single proposal from a broad, natural class of importance-sampling methods. Specifically, Appendix R proves that REACH has variance no greater than any importance-sampling estimator whose proposal preserves the conditional distribution of non-outcome tokens. This class includes arbitrary history-dependent re-scalings of the outcome-token probability, including constant re-scalings, the most natural potential baseline. Consequently, adding baselines from this class would not change the established per-sample variance ordering.

More general importance-sampling proposals that modify the distribution of non-outcome tokens are not covered by this guarantee and could offer different variance–computation tradeoffs. Such proposals would require additional design choices about how to steer generated clinical trajectories. Additionally, proposals optimized for individual rare outcomes generally sacrifice the simplicity of reusing an unmodified trajectory distribution across an arbitrary set of outcomes, which is a principal motivation for SCOPE.

Questions

  1. We have expanded Appendix B with an additional table, which describes the token corresponding to each outcome, the clinical information used to generate it, and the clinical threshold used to define it. For lab and vital outcomes, the extreme bin boundaries for the relevant category were set to the clinical threshold, so that a single token captures each outcome definition (noted in Appendix B, Table 2 caption). No separate label tokens were added: each outcome token is an ordinary token produced by the tokenization of the underlying EHR data, present in the training data like any other event, and labels are defined by the appearance of that same token in the held-out timeline.
  2. It is indeed true that SCOPE requires more tokens for 5 out of 11 outcomes in the Brier score analysis on MIMIC-IV, while on the local dataset SCOPE requires no more tokens than MC for any outcome. However, it is worth noting that these BSS differences at 100 samples are all less than 0.01, meaning there is a less than 1% relative difference in Brier score; this is the sense in which calibration is preserved. We have more prominently acknowledged this fact in the body. We now recommend that SCOPE be considered alongside Monte Carlo for all outcomes, rather than as a default replacement. Since both estimators are computed from the same pool of sampled timelines at negligible additional cost, this recommendation adds no inference cost. [Lines 403-405, 430-431]
  3. Only the results listed in Section 4.1 have been verified in Lean 4: the unbiasedness of SCOPE and REACH, that REACH is a Rao-Blackwellization of MC, and that REACH consequently reduces variance relative to MC. The results in Appendices H and Q (now I and R) are supported by written proofs but have not been formally verified. We have revised the abstract, which previously could be read as stating otherwise. [Lines 25-27]
Reviewer_eJMcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses and interesting and important problem for foundation models of electronic health records.

Critical strengths

The problem is of practical importance, the paper is nicely written. The paper is polished. The ideas and technical derivations are very good and non trivial. Nicely done that they do two estimators in one paper and not just try to papers. Many positive things are speaking for that paper.

Critical weaknesses

The paper addresses a problem of practical importance but then is more concerned about technicalities than maybe solving the problem. Or why are the supervised outcome prediction methods not included. There is some justification for it in the paper but that feels not really like the full answer. My assumption (please correct me) is that these methods might perform significantly better. But even if they do I do not think they would take something away from the paper. I just think it would be a nice comparison to have even if they are better and I think it is feasible to do so. Or at least justify it much better why that is not done.

Then why not use one more model family beyond Llama? Why not Qwen or at least one other model?

Questions for authors

Why did you go for ICLR rather than a medical journal?

see above i would already vote for accept and I think it is a strong paper

1 comment
AuthorsOpenReview ↗

Official Comment to Reviewer eJMc

We thank Reviewer eJMc for their clear and helpful review of our paper.

With regard to the first comment, we agree that we should more strongly motivate the paper's focus on the generative method. In our response to reviewer 3HgR, we describe the references we have added that demonstrate the efficacy of these methods compared to traditional supervised methods and described other advantages of this technique. [Lines 65-67]

The major reason why only one model family is considered in this paper is that each additional model family duplicates the entire computational cost. We do not believe this cost to be justified based on other results that show little difference between these model families for our intended use. In a recently accepted paper, we performed an experiment on this pipeline that compared Qwen and a Llama version of the model on hundreds of combinations of tokenizaiton schemes, downstream outcomes, and metrics. This failed to yield even a single statistically significant difference after BH correction, with the lowest corrected p-value being 0.354. While we cannot cite the paper directly at this time due to anonymization, upon acceptance, this number will be cited directly from the other paper.

We chose to submit this paper for publication in ICLR rather than a medical journal because the general question that SCOPE and REACH aim to answer: "What is the probability that a certain token will appear in the auto-regressively generated result of some prompt?" is one with wide reaching potential applications across many fields including recommender systems [1,2] and finance [3]. In light of this feedback, we have added discussion of application to other fields as potential future work discussed at the end of the paper [Lines 507-510]. Furthermore, a number of medically focused papers foundational to this work, most prominently Context Clues by Wornow et al. [4] have been published at ICLR.

  1. Actions speak louder than words, Zhai, et al., ICML, https://dl.acm.org/doi/10.5555/3692070.3694484
  2. GenRec: An LLM-Backed Recommendation Ranker at Netflix, Li, et al., https://doi.org/10.48550/arXiv.2608.10257
  3. TransactionGPT: Toward Foundational Transaction Modeling, Dou et al., KDD, https://dl.acm.org/doi/10.1145/3770855.3818496
  4. Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data, Wornow et al., ICLR, https://proceedings.iclr.cc/paper_files/paper/2025/hash/2a246ec4437808525a088abae744c2d3-Abstract-Conference.html
Reviewer_3HgROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work aims to improve improve the efficiency of outcome probability estimators that make use of EHR foundation models. It introduces: two estimators, SCOPE and REACH, which leverage next-token probabilities to reduce the number of generated trajectories required compared to standard MC estimation. REACH is an unbiased estimator and the authors provide guarantees of variance reduction. The authors evaluate both methods on several clinical outcomes across two EHR datasets, showing important reductions in inference cost while a predictive performance is maintained.

Critical strengths

The proposed methods are intuitive and simple to implement, and REACH comes with an interesting application of variance-reduction techniques and proper theoretical analysis. The theoretical developments look sound. The empirical evaluation is quite broad, covering multiple clinical outcomes with two datasets, and shows substantial computational savings. Overall, the paper is clearly written.

Critical weaknesses

I have two major concerns with respect to this paper:

  • Significance: I feel like the paper should argue in a much more convincing way why the addressed problem is important, i.e., why is it interesting to be able to estimate outcomes from foundation models instead of using specific models for that, which may be more effective and efficient. In which cases would that be optimal?
  • I am not convinced that the proposed method is sufficiently novel. Scope is very simple, and Reach is a clever trick that may have some practical usefulness if the problem is really significant, but I am not sure it has enough substance, even if the variance analysis is helpful.

Related to the significance concern, why is the probability of an event under simulated trajectories, terminated according to the model's stopping conditions, a useful surrogate for the patient's event risk?

Finally, the paper does not follow the citation style of ICLR, and it seems like authors have modified it to show only numbers instead of first author name and year to save space.

Questions for authors

See the weaknesses.

2 comments
AuthorsOpenReview ↗

Official Comment to Reviewer 3HgR

We thank Reviewer 3HgR for their constructive and insightful review of our paper.

We agree that the paper should more prominently discuss the significance of the generative prediction method at the heart of this paper. As of this rebuttal, numerous prior papers that have used generative inference have compared performance to single task supervised baselines. These papers have generally found that this method often exceeds the performance of supervised baselines [Waxler: Appendix J, Redekop: Discussion, Renc: Figure 5]. We have updated the introduction to reference this existing body of evidence and also described a number of the unique aspects and advantages of the generative prediction method, including multi-outcome prediction, the lack of task-specific training, the performance on label-scarce outcomes, and interpretability. With respect to efficiency, the main cost of this approach is the inference expense of sampling many trajectories, which is precisely what SCOPE and REACH reduce. We believe this change improves the motivation of the paper for readers who are less familiar with the generative prediction framework. [Lines 51-67]

While both estimators appear simple, we do not believe that the apparent simplicity of these estimators contradicts their novelty or significance. The apparent simplicity of each method masks that constructing each unbiased estimator required a careful formal derivation from the original indicator variable (subtle details such as the stopping condition of SCOPE and the outcome avoiding sampling of REACH are not obvious without this derivation and are required for the estimators to be unbiased). Beyond the estimators themselves, we show that REACH reduces variance relative to both MC and SCOPE for any model and stopping condition, that it is a Rao-Blackwellization of any importance sampling scheme preserving the non-outcome token distribution, and that the efficiency gains of both estimators are explainable (R² > 0.93) from spontaneity and prevalence alone.

As for why the probability of an event under simulated trajectories is often a useful surrogate for predicting the patient's event risk, the key is that an optimal model under the next token cross-entropy objective ensures that the true conditional next token distribution matches the observed conditional next token distribution. Propagating this property forward by auto-regressive inference then shows that an optimal model is not only able to predict the next token distribution faithfully but also any sequence level event. In practice, limited data, unobserved noise, and other factors make training a truly optimal model impossible. But the fact remains that the next token distribution method optimizes the model toward being calibrated for all sequence level events, which are the outcomes examined in this paper. With respect to the stopping condition, our observed outcomes are defined as occurring at least 24h after the patient is admitted and before discharge. Naturally, when labeling generated timelines, we label them in the same manner. Finally, the citation style accidentally used a numeric format from an incorrect template. We apologize for this mistake and have updated the manuscript to use the proper ICLR 2027 format.

  1. Generative Medical Event Models Improve with Scale, Waxler et al.
  2. Zero-shot medical event prediction using a generative pretrained transformer on electronic health records, Redekop et al.
  3. Foundation Model of Electronic Medical Records for Adaptive Risk Estimation, Renc et al.

I agree now with the authors justification for the significance of the problem. Regarding the novelty issue, I think there are arguments to defend the two proposed methods are substantial enough, although I am not entirely convinced. I will raise my score, and hope the other reviewers and AC will help determine whether the contribution is novel enough to justify acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.