Pith. sign in

REVIEW 4 major objections 4 minor 160 references

EHR-MPC: Inference-Time Control for Sepsis Treatment with Generative Patient Digital Twins

T0 review · 4 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Sepsis treatment can be optimized at inference time by planning over a learned generative model of patient trajectories.

desk verdict EHR-MPC is a coherent, clearly-written framework paper with a real but modest novel instantiation; its central simulation-performance claim is not yet convincing because the counterfactual validity of the digital twin is unvalidated, but the paper is worth serious review. read the letter →

arxiv 2607.08793 v3 pith:HDAT3VFW submitted 2026-07-07 stat.ML cs.AIcs.LGcs.SYeess.SYmath.OC

Joshua Pickard , Wei Qi , Na Li , Ann Woolley , Lisa Cosimi , Roy Kishony , Deborah Hung This is my paper
classification stat.MLcs.AIcs.LGcs.SYeess.SYmath.OC
keywords sepsistreatmentmodelpredictivecontrolpatientdigitaltwingenerativeEHRinference-timeplanningreinforcementlearningoff-policyevaluationcounterfactualsimulation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes EHR-MPC, a framework that separates learning how patients evolve from deciding how to treat them. It trains a generative model of tokenized electronic health record trajectories—a patient digital twin—that predicts future clinical states under forced intervention tokens. At inference time, a model predictive controller rolls out candidate action sequences through this twin, scores the simulated trajectories against a clinical objective such as SOFA improvement or mortality risk, and executes the best first action, then replans as new data arrive. Evaluated on 36,930 ICU patients across eight hospitals, the approach matches a Q-network baseline under off-policy importance sampling and outperforms it in on-policy simulation, while allowing clinical objectives to be changed without retraining. A sympathetic reader should care because this recasts sepsis treatment optimization as reusable, inspectable, inference-time control rather than a fixed policy.

What carries the argument

The central object is the token-forced autoregressive patient digital twin combined with receding-horizon model predictive control. The twin is a transformer trained on tokenized EHR data to predict next tokens; Eq. 4 forces candidate drug-administration tokens into the context and generates future tokens from the learned conditional distribution, producing a simulated trajectory under the proposed intervention. The MPC controller enumerates candidate action sequences, scores each simulated trajectory with a reward function chosen at inference time, executes only the first action of the best sequence, and re-plans as new data arrive. This mechanism is what converts a predictive sequence mode

What would settle it

Run a prospective randomized comparison of corticosteroid administration in sepsis and compare the digital twin's predicted 24-hour SOFA and mortality trajectories against the actual outcomes; large systematic divergence would falsify the claim that the twin provides valid counterfactual simulation for planning.

Watch

Extended reading notes

Core claim

The central claim is that a generative EHR model trained only for next-token prediction can serve as an interventional patient simulator when action tokens are forced into the context, and that planning over this simulator with model predictive control yields treatment decisions competitive with or better than fixed-policy reinforcement learning. The engine is Eq. 4: the autoregressive factorization of the trajectory distribution conditioned on candidate actions, which turns a passive forecasting model into a counterfactual simulator. From this, the paper argues that learning patient dynamics and specifying clinical objectives are decoupled: the same digital twin can optimize SOFA, mortality

Load-bearing premise

The load-bearing premise is that forcing action tokens into an autoregressive model trained on observational EHR records yields trustworthy predictions of what would happen under those interventions; if the model has not learned true treatment effects, the planning advantage is an artifact of the simulator.

Editorial extensions

If this is right

  • One trained digital twin can support multiple clinical objectives, such as SOFA improvement and mortality reduction, with no retraining.
  • Clinical goals can be modified or constrained at deployment, allowing evolving standards of care or patient-specific trade-offs.
  • Counterfactual trajectories are directly inspectable, giving clinicians forecasts of future patient state rather than only a recommended action.
  • The framework applies to any sequential clinical decision problem where tokenized EHR trajectories and intervention tokens are available.
  • Planning-based policies can be evaluated on-policy in the learned simulator, avoiding some high-variance off-policy issues.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: A key untested step is whether token-forced generation in a model trained on observational data equals true interventional counterfactuals; a decisive experiment would validate the twin's counterfactual predictions against a randomized steroid trial or a natural experiment with exogenous variation in drug availability.
  • Inference: The improved on-policy performance may partly reflect the planner and evaluator sharing the same learned dynamics; the cross-epoch experiment is a partial control, and a more convincing test would plan with one digital twin and evaluate with an independently trained or prospective simulator.
  • Inference: If the framework holds, it suggests a broader path: pretrain a general EHR trajectory model once, then attach inference-time planners for many clinical goals, turning a prediction model into a reusable substrate for decision-making.
  • Inference: The dose-response sensitivity table is a natural audit point; clinicians can judge whether the monotonic steroid response the planner exploits is clinically plausible for each drug.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes EHR-MPC, a framework that separates learning a generative patient digital twin from optimizing treatment decisions. A transformer is trained on tokenized EHR trajectories from 36,930 ICU patients across eight hospitals. At inference time, candidate drug-action tokens are forced into the autoregressive model to simulate counterfactual trajectories (Eq. 4), and model predictive control selects actions by maximizing a clinically specified reward over these rollouts (Eq. 6). The authors evaluate the resulting policy against a Q-network baseline and clinician behavior using weighted importance sampling on observed data and simulation-based rollouts on the learned digital twin. They report comparable off-policy performance and improved simulation performance, along with policy divergence analyses and a dose-response sensitivity check.

Significance. If the central claims hold, EHR-MPC would be a genuinely useful framework: it decouples dynamics learning from reward specification, enables inference-time adaptation to new clinical objectives without retraining, and produces interpretable counterfactual trajectory forecasts. The use of a large multi-center EHR cohort, next-token forecasting evaluation, and a non-simulator off-policy estimator are strengths. However, the headline improvement over RL baselines rests almost entirely on simulation-based evaluation using the same learned dynamics family that is used for planning. The paper's own limitations section concedes that the training process does not guarantee correct system identification. Unless the counterfactual validity of Eq. 4 is established or the simulation claim is substantially reframed, the empirical contribution is closer to a self-consistency demonstration than to validated treatment optimization.

major comments (4)
  1. [§5.2, Figs. 3–4] The abstract's 'improved simulation performance' is supported by on-policy rollouts under the same learned digital twin p_theta that is used for planning in Eq. 6. The robustness check (Fig. 4) decouples planning and evaluation using early (40% epochs) and later checkpoints from the same training run, but these checkpoints share architecture, tokenization, training data, and confounding structure. This does not rule out the possibility that MPC is exploiting systematic biases in the learned simulator family. A claim of improved treatment needs either an external evaluation simulator from a different model family or mechanistic source, or a clear statement that the result is a within-model comparison.
  2. [§3.2, Eq. (4)] Eq. (4) is the load-bearing step: forcing action tokens into an autoregressive EHR model and decoding the next-token distribution is interpreted as a counterfactual trajectory under intervention. The model is trained on observational EHR data, and the paper's Limitations state that 'the training process does not provide guarantees of correct system identification.' No causal identifiability conditions, unobserved-confounding sensitivity analysis, or external counterfactual validation is provided. Consequently, the dose-response pattern in Table 4 (steroid mortality risk falling from 0.38 to 0.12 as injected tokens increase from 1 to 20) could reflect a learned association or a reward-hacking artifact rather than physiology. The authors should either state explicitly that simulation performance is a model-internal metric or provide validation against known interventional effects (e.g., ra
  3. [§5.2, Fig. 3 (left)] The off-policy WIS evaluation—the only result that does not rely on the digital twin—shows that EHR-MPC and the Q-network achieve similar point estimates with overlapping confidence intervals. This supports 'comparable' but not 'improved' performance. The abstract's dual claim therefore leans almost entirely on the circular simulation result. If the simulation result is not externally grounded, the paper's stronger conclusion should be softened or removed.
  4. [§3.3 vs. Appendix B.3, Eq. (6), Algorithm 1] Eq. (6) defines the MPC objective as an argmax over an expectation with respect to p_theta, but Algorithm 1 approximates this by greedy (argmax) decoding of a single trajectory per candidate action sequence. The greedy approximation is not justified, and no analysis is given for how this affects planning quality or whether stochastic rollouts would change the selected actions. Since all experimental results use this approximate planner, the gap between the formal objective and the implemented method should be discussed and, if possible, quantified.
minor comments (4)
  1. [§3.3 / Algorithm 1] The notation for action indices is inconsistent: Eq. (6) uses a_{kt:kt+h}, while Algorithm 1 loops τ = k_t to k_{t+h} and forces a_τ. Clarify that the candidate action sequence is defined on token positions, not time steps, and that the time-to-token mapping depends on realized events.
  2. [§5.2, Figs. 3–4] The legend in the 'Policy Comparison' panels is unclear: 'Best policy' is not defined, and the y=x line should be explained in the caption. Also state whether the ΔSOFA values are daily changes or cumulative.
  3. [Appendix B.4] The WIS behavior policy is estimated with a multinomial logistic regression on last-token representations. Since the same representations are used for the Q-network and MPC, this ties the off-policy estimator to the learned feature space. Report sensitivity to the temperature τ and clip c_max, as these are free parameters that can affect the conclusions.
  4. [Table 4] The dose-response table reports point estimates without confidence intervals. Given the high variability of generated trajectories, bootstrapped intervals across patients or generation seeds would make the monotonicity claim more credible.

Circularity Check

3 steps flagged · score 6.0 of 10

Simulation-based 'improved performance' is partly self-consistency: MPC optimizes under pθ and the evaluator rolls out the same fitted model family.

  1. self definitional [§3.3 Eq. 6 and §5.2 'On-Policy Evaluation via Digital Twin Simulator']
    "a∗t = arg max akt:kt+h E x̂kt:kt+h ∼pθ(·|x1:kt ,akt:kt+h ) [ R(x̂kt:kt+h) ] . ... We assess learned policies using the digital twin as a simulator. ... Across simulations, EHR-MPC consistently outperforms Q-network policies aimed to improve daily SOFA scores and decrease mortality risk"

    Eq. 6 defines the EHR-MPC policy as the argmax of expected reward under pθ. The simulation-based evaluation then scores that same policy by rolling out pθ (a later checkpoint of the same training run). A policy chosen to maximize R under pθ will, by construction, look better under pθ than policies such as Q-network or clinician that were not optimized for pθ's reward. Thus 'improved simulation performance' is a self-consistency property of the fitted model, not an independent confirmation that the counterfactual trajectories are valid. The off-policy WIS result is the only evaluation that does not share this loop.

  2. other [§5.2 'Robustness to Model Misspecification']
    "To address this, we decouple the planning model from the evaluation model using checkpoints from different training epochs: MPC planning uses an early-epoch checkpoint (40% of training epochs), while policy scoring and simulation use a later, more-converged checkpoint."

    The claimed decoupling is only between two checkpoints of the same fitted pθ: same architecture, tokenizer, training data, and observational confounding structure. Any systematic bias learned by the model family, e.g., steroid tokens appearing to lower mortality due to confounding, is shared by both checkpoints. The robustness experiment therefore does not break the self-consistency loop; it only prevents exact exploitation of one parameter setting.

1 more flagged steps
  1. fitted input called prediction [§5.1 'Dose-Response Sensitivity' and Table 4]
    "we construct counterfactual 24-hour trajectories by injecting repeated drug tokens into the observed context as if the drug were administered at a higher dose and then generate future tokens by sampling from the digital twin pθ. ... Corticosteroid tokens induce a consistent, monotonic decrease in predicted mortality risk across increasing injection levels."

    This dose-response 'sensitivity' is generated by rolling out pθ and scored by a task-specific prediction head trained on the same frozen backbone representations. The monotone steroid response is therefore a property of the learned model, possibly reflecting confounding in observational treatment assignment, not an external counterfactual validation. Presenting it as evidence of 'local controllability' that supports MPC is circular to the extent that the model is used to validate its own intervention behavior.

full rationale

The off-policy WIS evaluation is genuinely external: it uses observed trajectories and outcomes rather than rollouts of pθ, and it supports only 'comparable' performance. The headline 'improved simulation performance,' however, is substantially self-referential: Eq. 6 optimizes actions under pθ and the on-policy evaluation rolls out pθ, while the robustness split between two checkpoints of the same training run leaves the shared model family inside the loop. The paper's own Limitations concede that 'the training process does not provide guarantees of correct system identification,' which is exactly the condition under which the simulation result could be a self-consistency artifact. The dose-response check is likewise internal. I found no load-bearing self-citation: Pickard et al. 2025 is only a related-work example, so the circularity is not a citation-chain problem but a fitted-model-evaluating-itself problem. Hence partial circularity: score 6.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical entities are introduced; the 'digital twin' is a trained model, not a postulated object. The ledger is dominated by the counterfactual-validity assumption and by hand-chosen MPC/WIS/SOFA hyperparameters that the central empirical claims depend on.

free parameters (5)
  • MPC planning horizon h = 24 hours (six [4 hour] tokens)
    Chosen by hand in Section 3.3 and Appendix B.3; determines the length of simulated counterfactual trajectories and directly shapes the reward signal.
  • WIS temperature and importance-weight clip = tau=0.5, c_max=10
    Appendix B.4: softmax temperature and clipping are user-chosen and materially affect the off-policy value estimates.
  • Action candidate set = 4 action classes; up to 5 representative drug tokens per class; N candidate sequences in Algorithm 1 not reported
    The search space is constructed by hand; the number of sampled candidates N is not stated, so the optimization procedure is under-specified.
  • SOFA imputation rules = dPaO2 = 3(SpO2-90)+60 mmHg; FiO2 hierarchy; fixed vasopressor score of 2
    Appendix A.1: ad hoc approximations for missing PaO2, FiO2, and vasopressor dose; these directly affect the reward used for both optimization and evaluation.
  • Greedy decoding instead of expectation in Eq. 6 = argmax decoding, one rollout per candidate
    Algorithm 1/B.3 replaces the expectation over p_theta in Eq. 6 with a single greedy trajectory; the approximation error is unquantified.
assumptions (6)
  • domain assumption Token-forced next-token generation p(x_{k_t:k_t+h} | x_{1:k_t}, a_{k_t:k_t+h}) in Eq. 4 produces counterfactual trajectories under candidate treatments.
    This is the load-bearing causal assumption. The model is trained on observational EHR data with no stated unobserved-confounding or system-identification conditions. The authors state in Limitations that the training process provides no guarantees of correct system identification.
  • domain assumption A transformer trained for next-token prediction on tokenized EHR records captures patient dynamics well enough that MPC planned over it improves real outcomes.
    Section 5.1 infers 'empirical conditions for local controllability' from forecasting accuracy and dose-response sensitivity, but these do not establish that simulated improvements transfer to real patients.
  • ad hoc to paper Greedy argmax decoding of one trajectory per candidate action approximates the expectation over p_theta in Eq. 6.
    Algorithm 1 and Appendix B.3 use greedy decoding and a single rollout per candidate while Eq. 6 defines an expectation; this discrepancy is not analyzed.
  • domain assumption Outcome-prediction heads trained on frozen real-token representations generalize to generated/token-forced trajectories with distribution shift.
    The simulation evaluation scores generated trajectories with task-specific heads (Appendix B.2). No validation shows head accuracy is preserved under model-generated inputs.
  • domain assumption The SOFA and mortality reward functions, as computed from raw EHR fields, are valid proxies for clinical benefit.
    Appendix A.1 includes ad hoc imputations (PaO2 from SpO2, FiO2 hierarchy, fixed vasopressor score 2) that affect the reward used for policy optimization and evaluation.
  • domain assumption Aggregating 512-token context windows approximates Markovian ICU dynamics.
    Appendix B.1: long trajectories are split into non-overlapping windows; this architectural choice is assumed adequate for capturing short-horizon clinical dynamics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EHR-MPC: Inference-Time Control for Sepsis Treatment with Generative Patient Digital Twins." pith.science (2026). https://pith.science/paper/HDAT3VFW

@misc{pith2026260708793,
  author       = {Pith},
  title        = {Pith review of: EHR-MPC: Inference-Time Control for Sepsis Treatment with Generative Patient Digital Twins},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HDAT3VFW}},
  note         = {Machine review of arXiv:2607.08793}
}
read the original abstract

Sepsis is a leading cause of mortality, yet optimal treatment policies remain contested. Existing reinforcement learning (RL) approaches learn fixed strategies for sepsis treatment, limiting adaptability to changing clinical objectives during inference. We propose EHRMPC, a framework that decouples learning patient dynamics from optimizing treatment by training a patient digital twin in the form of a generative electronic health record (EHR) model. The digital twin predicts clinical trajectories under interventions and enables model predictive control (MPC) to optimize treatments via inference-time planning over simulations. We evaluate EHR-MPC on a multicenter ICU sepsis cohort spanning 8 hospitals in the Mass General Brigham health system using both off-policy importance sampling and on-policy simulation-based evaluation. Relative to RL baselines, EHR-MPC achieves comparable off-policy performance and improved simulation performance. Unlike RL, this work frames sepsis treatment optimization as inference-time control over learned patient dynamics, establishing a general framework for decision making with generative clinical models.

Figures

Figures reproduced from arXiv: 2607.08793 by the authors.

Figure 1
Figure 1. Digital Twin MPC for Treatment Optimization. (left) A patient generates data that is observed by a clinician, tokenized, and sent to the digital twin. (simulation) Using observed data, a digital twin performs forecasts on the state of the patient as if several candidate treatments were applied. Each forecast is evaluated according to several clinical outcomes. (optimization) Based upon clinical goals defined at infe… view at source ↗
Figure 2
Figure 2. Representations and Forecasting. Prediction of outcomes from patient embeddings and random forest classifiers. ministrative (admit, transfer, discharge, etc.) events were mapped to dedicated vocabulary entries. The tokenized dataset contained 469 million content tokens, 90.3% of which in￾dicated observational data rather than interventions or other data. The median patient trajectory sequence length contained 20,232… view at source ↗
Figure 3
Figure 3. Offline Evaluation of EHR-MPC Treatment Policies. The learned EHR-MPC policy is evaluated with WIS (left) and simulated patient outcomes (middle, right). The improvement of patient outcomes according to EHR-MPC and Q-networks is shown per patient (right). weighted importance sampling (WIS). Although pθ is trained on the full ICU cohort, evalu￾ation is restricted to sepsis patients. Across both objectives, EHR-MPC an… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: MPC Robustness to Model Misalignment. Performance of MPC and Q-networks are evaluated on the digital twin simulator when the MPC controller uses a less accurate version of the patient digital twin. right). We examine the robustness of this result to model misspecificat…
Figure 5
Figure 5. Figure 5: Policy Divergence for Mortality Optimization. (Top) The distribution of recom￾mended drugs is shown for each of the three policies across locations. (Bottom) The JSD divergence between these distributions is shown. selecting interventions that optimize clinical goals. …
Figure 6
Figure 6. Figure 6: Policy Variation for Mortality Risk Minimization. EHR-MPC DQN Clin || MPC Clin || DQN EHR-MPC || DQN Actions Vasopressor Antibiotics Steroids JSD Total MGH BWH SLM NWH WDH BWF CDH MVH [PITH_FULL_IMAGE:figures/full_fig_p027_6.png]
Figure 6
Figure 6. Figure 6: Policy Variation for Mortality Risk Minimization. EHR-MPC DQN Clin || MPC Clin || DQN EHR-MPC || DQN Actions Vasopressor Antibiotics Steroids JSD Total MGH BWH SLM NWH WDH BWF CDH MVH [PITH_FULL_IMAGE:figures/full_fig_p028_6.png]
Figure 7
Figure 7. Figure 7: Policy Distributions and Variation for SOFA Minimization. JSD Clinical Policy EHR-MPC Policy DQN Policy [PITH_FULL_IMAGE:figures/full_fig_p027_7.png]
Figure 7
Figure 7. Figure 7: Policy Distributions and Variation for SOFA Minimization. JSD Clinical Policy EHR-MPC Policy DQN Policy [PITH_FULL_IMAGE:figures/full_fig_p028_7.png]
Figure 8
Figure 8. Figure 8: Policy Variation for SOFA Minimization. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 8
Figure 8. Figure 8: Policy Variation for SOFA Minimization. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_8.png]

Review history (3 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

160 extracted references

  1. [1]

    Name of the conference , pages =

    My awesome Machine Learning model will change the way we do healthcare , author =. Name of the conference , pages =

  2. [2]

    Proceedings of the National Academy of Sciences , volume=

    Dynamic sensor selection for biomarker discovery , author=. Proceedings of the National Academy of Sciences , volume=. 2025 , publisher=

    2025

  3. [3]

    Acta Anaesthesiologica Scandinavica , publisher =

    The first intensive care unit in the world: Copenhagen 1953 , author =. Acta Anaesthesiologica Scandinavica , publisher =

    1953

  4. [4]

    Critical care medicine , publisher =

    Hydrocortisone reverses refractory septic shock , author =. Critical care medicine , publisher =

  5. [5]

    Critical care medicine , publisher =

    Stress doses of hydrocortisone reverse hyperdynamic septic shock: a prospective, randomized, double-blind, single-center study , author =. Critical care medicine , publisher =

  6. [6]

    Critical care medicine , publisher =

    Reversal of late septic shock with supraphysiologic doses of hydrocortisone , author =. Critical care medicine , publisher =

  7. [7]

    The Lancet , publisher =

    Adjunct prednisone therapy for patients with community-acquired pneumonia: a multicentre, double-blind, randomised, placebo-controlled trial , author =. The Lancet , publisher =

  8. [8]

    Corticosteroid treatment and intensive insulin therapy for septic shock in adults: a randomized controlled trial , author =

Show all 160 references
  1. [9]

    Canadian Medical Association Journal , publisher =

    Low-dose hydrocortisone in patients with cirrhosis and septic shock: a randomized controlled trial , author =. Canadian Medical Association Journal , publisher =

  2. [10]

    Critical care medicine , publisher =

    Hydrocortisone effects on cardiovascular variability in septic shock: a spectral analysis approach , author =. Critical care medicine , publisher =

  3. [11]

    Critical Care , publisher =

    Corticosteroids in septic shock: a systematic review and network meta-analysis , author =. Critical Care , publisher =

  4. [12]

    Clinical infectious diseases , publisher =

    Safety and efficacy of corticosteroids for the treatment of septic shock: a systematic review and meta-analysis , author =. Clinical infectious diseases , publisher =

  5. [13]

    JAMA internal medicine , volume = 179, number = 2, pages =

    Association of corticosteroid treatment with outcomes in adult patients with sepsis: a systematic review and meta-analysis , author =. JAMA internal medicine , volume = 179, number = 2, pages =

  6. [14]

    Jama , volume = 324, number = 13, pages =

    Association between administration of systemic corticosteroids and mortality among critically ill patients with COVID-19: a meta-analysis , author =. Jama , volume = 324, number = 13, pages =

  7. [15]

    Critical care medicine , publisher =

    Corticosteroids in sepsis: an updated systematic review and meta-analysis , author =. Critical care medicine , publisher =

  8. [16]

    New England Journal of Medicine , publisher =

    Hydrocortisone plus fludrocortisone for adults with septic shock , author =. New England Journal of Medicine , publisher =

  9. [17]

    New England Journal of Medicine , publisher =

    Adjunctive glucocorticoid therapy in patients with septic shock , author =. New England Journal of Medicine , publisher =

  10. [18]

    Critical care medicine , publisher =

    The interaction of vasopressin and corticosteroids in septic shock: a pilot randomized controlled trial , author =. Critical care medicine , publisher =

  11. [19]

    Critical care medicine , publisher =

    Interaction of vasopressin infusion, corticosteroid treatment, and mortality of septic shock , author =. Critical care medicine , publisher =

  12. [20]

    Jama , volume = 316, number = 5, pages =

    Effect of early vasopressin vs norepinephrine on kidney failure in patients with septic shock: the VANISH randomized clinical trial , author =. Jama , volume = 316, number = 5, pages =

  13. [21]

    Cochrane Database of Systematic Reviews , publisher =

    Corticosteroids for treating sepsis in children and adults , author =. Cochrane Database of Systematic Reviews , publisher =

  14. [22]

    New England Journal of Medicine , publisher =

    Hydrocortisone therapy for patients with septic shock , author =. New England Journal of Medicine , publisher =

  15. [23]

    Jama , volume = 316, number = 17, pages =

    Effect of hydrocortisone on development of shock among patients with severe sepsis: the HYPRESS randomized clinical trial , author =. Jama , volume = 316, number = 17, pages =

  16. [24]

    Bmj , publisher =

    Corticosteroids for severe sepsis and septic shock: a systematic review and meta-analysis , author =. Bmj , publisher =

  17. [25]

    Jama , publisher =

    Effect of treatment with low doses of hydrocortisone and fludrocortisone on mortality in patients with septic shock , author =. Jama , publisher =

  18. [26]

    low-dose

    Immunologic and hemodynamic effects of “low-dose” hydrocortisone in septic shock: a double-blind, randomized, placebo-controlled, crossover study , author =. American journal of respiratory and critical care medicine , publisher =

  19. [27]

    American Journal of Respiratory and Critical Care Medicine , volume = 138, number = 1, pages =

    Ineffectiveness of high-dose methylprednisolone in preventing parenchymal lung injury and improving mortality in patients with septic shock , author =. American Journal of Respiratory and Critical Care Medicine , volume = 138, number = 1, pages =

  20. [28]

    New England journal of medicine , publisher =

    Effect of high-dose glucocorticoid therapy on mortality in patients with clinical signs of systemic sepsis , author =. New England journal of medicine , publisher =

  21. [29]

    New England Journal of Medicine , publisher =

    The effects of high-dose corticosteroids in patients with septic shock: a prospective, controlled study , author =. New England Journal of Medicine , publisher =

  22. [30]

    New England Journal of Medicine , publisher =

    A controlled clinical trial of high-dose methylprednisolone in the treatment of severe sepsis and septic shock , author =. New England Journal of Medicine , publisher =

  23. [31]

    Intensive care medicine , publisher =

    Surviving Sepsis Campaign: international guidelines for management of sepsis and septic shock 2026 , author =. Intensive care medicine , publisher =

    2026

  24. [32]

    Advances in Neural Information Processing Systems , volume = 36, pages =

    Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models , author =. Advances in Neural Information Processing Systems , volume = 36, pages =

  25. [33]

    Critical care explorations , publisher =

    Accuracy of international classification of diseases, 10th revision codes for identifying sepsis: a systematic review and meta-analysis , author =. Critical care explorations , publisher =

  26. [34]

    Handbook of sepsis , publisher =

    Handbook of sepsis , author =. Handbook of sepsis , publisher =

  27. [35]

    Automatica , publisher =

    Model predictive control: Theory and practice—A survey , author =. Automatica , publisher =

  28. [36]

    2012 ieee 51st ieee conference on decision and control (cdc) , pages =

    An introduction to event-triggered and self-triggered control , author =. 2012 ieee 51st ieee conference on decision and control (cdc) , pages =

    2012

  29. [37]

    npj Digital Medicine , publisher =

    Large language models forecast patient health trajectories enabling digital twins , author =. npj Digital Medicine , publisher =

  30. [38]

    GigaScience , publisher =

    Foundation model of electronic medical records for adaptive risk estimation , author =. GigaScience , publisher =

  31. [39]

    Critical care medicine , publisher =

    Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021 , author =. Critical care medicine , publisher =

    2021

  32. [40]

    Evaluating reinforcement learning algorithms in observational health settings , author =

  33. [41]

    The international journal of advanced manufacturing technology , publisher =

    Review on model predictive control: An engineering perspective , author =. The international journal of advanced manufacturing technology , publisher =

  34. [42]

    Eligibility traces for off-policy policy evaluation , author =

  35. [43]

    Robustness in identification and control , publisher =

    Robust model predictive control: A survey , author =. Robustness in identification and control , publisher =

  36. [44]

    Signal analysis and prediction , publisher =

    System identification , author =. Signal analysis and prediction , publisher =

  37. [45]

    International Conference on Machine Learning , pages =

    Counterfactual off-policy evaluation with gumbel-max structural causal models , author =. International Conference on Machine Learning , pages =

  38. [46]

    International conference on machine learning , pages =

    Estimating individual treatment effect: generalization bounds and algorithms , author =. International conference on machine learning , pages =

  39. [47]

    , author =

    Controllability of nonlinear systems. , author =

  40. [48]

    IEEE Transactions on automatic control , publisher =

    Nonlinear controllability and observability , author =. IEEE Transactions on automatic control , publisher =

  41. [49]

    New England Journal of Medicine , publisher =

    The path to personalized medicine , author =. New England Journal of Medicine , publisher =

  42. [50]

    npj Digital Medicine , publisher =

    Evolutionary digital twin framework for optimal aminoglycoside dosing in neonates with suspected sepsis , author =. npj Digital Medicine , publisher =

  43. [51]

    Nature computational science , publisher =

    Digital twins in medicine , author =. Nature computational science , publisher =

  44. [52]

    Controlling generative models with continuous factors of variations , author =

  45. [53]

    Bootstrapping with models: Confidence intervals for off-policy evaluation , author =

  46. [54]

    Weighted importance sampling for off-policy learning with linear function approximation , author =

  47. [55]

    Annals of surgery , publisher =

    Steroids in the treatment of clinical septic shock , author =. Annals of surgery , publisher =

  48. [56]

    Heliyon , publisher =

    Development and validation of the SickKids Enterprise-wide Data in Azure Repository (SEDAR) , author =. Heliyon , publisher =

  49. [57]

    npj Digital Medicine , publisher =

    Clinically-guided models or foundation models? predicting cervical spondylotic myelopathy from electronic health records , author =. npj Digital Medicine , publisher =

  50. [58]

    NPJ digital medicine , publisher =

    A multi-center study on the adaptability of a shared foundation model for electronic health records , author =. NPJ digital medicine , publisher =

  51. [59]

    EHR-R1: A Reasoning-Enhanced Foundational Language Model for Electronic Health Record Analysis , author =

  52. [60]

    Machine learning for health , pages =

    CEHR-BERT: Incorporating temporal information from structured EHR data to improve prediction tasks , author =. Machine learning for health , pages =

  53. [61]

    VIReC Insights , volume = 3, number = 1, pages =

    Veterans health information systems and technology architecture (VistA) as a research tool , author =. VIReC Insights , volume = 3, number = 1, pages =

  54. [62]

    Journal of Biomedical Informatics , publisher =

    EHR-BERT: A BERT-based model for effective anomaly detection in electronic health records , author =. Journal of Biomedical Informatics , publisher =

  55. [63]

    Nature , publisher =

    Health system-scale language models are all-purpose prediction engines , author =. Nature , publisher =

  56. [64]

    NPJ digital medicine , publisher =

    A large language model for electronic health records , author =. NPJ digital medicine , publisher =

  57. [65]

    Nature , publisher =

    Large language models encode clinical knowledge , author =. Nature , publisher =

  58. [66]

    Nature medicine , publisher =

    Toward expert-level medical question answering with large language models , author =. Nature medicine , publisher =

  59. [67]

    JAmIA Open , publisher =

    Standardized Health data and Research Exchange (SHaRE): promoting a learning health system , author =. JAmIA Open , publisher =

  60. [68]

    International journal of epidemiology , publisher =

    Data resource profile: clinical practice research datalink (CPRD) , author =. International journal of epidemiology , publisher =

  61. [69]

    Scientific reports , publisher =

    BEHRT: transformer for electronic health records , author =. Scientific reports , publisher =

  62. [70]

    NPJ digital medicine , publisher =

    Med-BERT: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction , author =. NPJ digital medicine , publisher =

  63. [71]

    Nejm Ai , publisher =

    Towards generalist biomedical AI , author =. Nejm Ai , publisher =

  64. [72]

    Bioinformatics , publisher =

    BioBERT: a pre-trained biomedical language representation model for biomedical text mining , author =. Bioinformatics , publisher =

  65. [73]

    Clinicalbert: Modeling clinical notes and predicting hospital readmission , author =

  66. [74]

    NPJ digital medicine , publisher =

    Zero shot health trajectory prediction using transformer , author =. NPJ digital medicine , publisher =

  67. [75]

    Foundation models for electronic health records: representation dynamics and transferability , author =

  68. [76]

    npj digital medicine , publisher =

    The shaky foundations of large language models and foundation models for electronic health records , author =. npj digital medicine , publisher =

  69. [77]

    New England journal of medicine , publisher =

    Early goal-directed therapy in the treatment of severe sepsis and septic shock , author =. New England journal of medicine , publisher =

  70. [78]

    American journal of respiratory and critical care medicine , publisher =

    The timing of early antibiotics and hospital mortality in sepsis , author =. American journal of respiratory and critical care medicine , publisher =

  71. [79]

    Understanding intermediate layers using linear classifier probes , author =

  72. [80]

    Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers , pages =

    Using the output embedding to improve language models , author =. Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers , pages =

  73. [81]

    Language models are unsupervised multitask learners , author =

  74. [82]

    Distilling the knowledge in a neural network , author =

  75. [83]

    IEEE Transactions on Signal Processing , publisher =

    On maintaining linear convergence of distributed learning and optimization under limited communication , author =. IEEE Transactions on Signal Processing , publisher =

  76. [84]

    International conference on machine learning , pages =

    Model-agnostic meta-learning for fast adaptation of deep networks , author =. International conference on machine learning , pages =

  77. [85]

    Nature medicine , publisher =

    Early prediction of circulatory failure in the intensive care unit using machine learning , author =. Nature medicine , publisher =

  78. [86]

    Intensive care medicine , publisher =

    Machine learning for the prediction of sepsis: a systematic review and meta-analysis of diagnostic test accuracy , author =. Intensive care medicine , publisher =

  79. [87]

    Critical care medicine , publisher =

    Sharing ICU patient data responsibly under the society of critical care medicine/European society of intensive care medicine joint data science collaboration: the Amsterdam university medical centers database (AmsterdamUMCdb) example , author =. Critical care medicine , publisher =

  80. [88]

    Scientific data , publisher =

    The eICU Collaborative Research Database, a freely available multi-center database for critical care research , author =. Scientific data , publisher =

  81. [89]

    Scientific data , publisher =

    MIMIC-IV, a freely accessible electronic health record dataset , author =. Scientific data , publisher =

  82. [90]

    Scientific data , publisher =

    MIMIC-III, a freely accessible critical care database , author =. Scientific data , publisher =

  83. [91]

    2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society , pages =

    Open-access MIMIC-II database for intensive care research , author =. 2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society , pages =

    2011

  84. [92]

    circulation , publisher =

    PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals , author =. circulation , publisher =

  85. [93]

    Intensive care medicine , publisher =

    The SOFA (Sepsis-related Organ Failure Assessment) score to describe organ dysfunction/failure: On behalf of the Working Group on Sepsis-Related Problems of the European Society of Intensive Care Medicine (see contributors to the project in the appendix) , author =. Intensive ...

  86. [94]

    Archives of Surgery , publisher =

    APACHE 1978-2001: the development of a quality assurance system based on prognosis: milestones and personal reflections , author =. Archives of Surgery , publisher =

    1978

  87. [95]

    influenzae , author =

    On the antibacterial action of cultures of a penicillium, with special reference to their use in the isolation of B. influenzae , author =

  88. [96]

    Jama , volume = 315, number = 8, pages =

    The third international consensus definitions for sepsis and septic shock (Sepsis-3) , author =. Jama , volume = 315, number = 8, pages =

  89. [97]

    Critical care medicine , publisher =

    2001 sccm/esicm/accp/ats/sis international sepsis definitions conference , author =. Critical care medicine , publisher =

    2001

  90. [98]

    Chest , publisher =

    Definitions for sepsis and organ failure and guidelines for the use of innovative therapies in sepsis , author =. Chest , publisher =

  91. [99]

    Intensive care medicine , publisher =

    Historical perspective of the word “sepsis” , author =. Intensive care medicine , publisher =

  92. [100]

    Advances in neural information processing systems , volume = 33, pages =

    Conservative q-learning for offline reinforcement learning , author =. Advances in neural information processing systems , volume = 33, pages =

  93. [101]

    Advances in neural information processing systems , volume = 34, pages =

    Bellman-consistent pessimism for offline reinforcement learning , author =. Advances in neural information processing systems , volume = 34, pages =

  94. [102]

    Exploration from a Primal-Dual Lens: Value-Incentivized Actor-Critic Methods for Sample-Efficient Online RL , author =

  95. [103]

    Value-incentivized preference optimization: A unified approach to online and offline rlhf , author =

  96. [104]

    International conference on machine learning , pages =

    Is pessimism provably efficient for offline rl? , author =. International conference on machine learning , pages =

  97. [105]

    Reinforcement learning via fenchel-rockafellar duality , author =

  98. [106]

    Nature Machine Intelligence , publisher =

    The need for uncertainty quantification in machine-assisted medical decision making , author =. Nature Machine Intelligence , publisher =

  99. [107]

    Trends in immunology , publisher =

    Towards systems immunology of critical illness at scale: from single cell ‘omics to digital twins , author =. Trends in immunology , publisher =

  100. [108]

    IEEE Open Journal of Engineering in Medicine and Biology , publisher =

    A reinforcement learning model for optimal treatment strategies in intensive care: assessment of the role of cardiorespiratory features , author =. IEEE Open Journal of Engineering in Medicine and Biology , publisher =

  101. [109]

    Journal of the Society for Industrial and Applied Mathematics, Series A: Control , publisher =

    Mathematical description of linear dynamical systems , author =. Journal of the Society for Industrial and Applied Mathematics, Series A: Control , publisher =

  102. [110]

    Journal of the Society for Industrial and Applied Mathematics, Series A: Control , publisher =

    Controllability and observability in multivariable control systems , author =. Journal of the Society for Industrial and Applied Mathematics, Series A: Control , publisher =

  103. [111]

    Proceedings first international conference on automatic control, Moscow, USSR , pages =

    On the general theory of control systems , author =. Proceedings first international conference on automatic control, Moscow, USSR , pages =

  104. [112]

    Journal of Clinical Medicine , publisher =

    Development of a reinforcement learning algorithm to optimize corticosteroid therapy in critically ill patients with sepsis , author =. Journal of Clinical Medicine , publisher =

  105. [113]

    PLoS computational biology , publisher =

    Structural identifiability of dynamic systems biology models , author =. PLoS computational biology , publisher =

  106. [114]

    Communications medicine , publisher =

    Sepsis endotypes identified by host gene expression across global cohorts , author =. Communications medicine , publisher =

  107. [115]

    Nature immunology , publisher =

    The pathophysiology of sepsis and precision-medicine-based immunotherapy , author =. Nature immunology , publisher =

  108. [116]

    Learning to treat sepsis with multi-output gaussian process deep recurrent q-networks , author =

  109. [117]

    Nature medicine , publisher =

    The artificial intelligence clinician learns optimal treatment strategies for sepsis in intensive care , author =. Nature medicine , publisher =

  110. [118]

    Journal of the American Medical Informatics Association , publisher =

    Sepsis prediction, early detection, and identification using clinical text for machine learning: a systematic review , author =. Journal of the American Medical Informatics Association , publisher =

  111. [119]

    science , publisher =

    Dynamic programming , author =. science , publisher =

  112. [120]

    Reinforcement learning: An introduction , author =

  113. [121]

    Journal of Cognitive Neuroscience , volume = 11, number = 1, pages =

    Reinforcement learning , author =. Journal of Cognitive Neuroscience , volume = 11, number = 1, pages =

  114. [122]

    Optimizing medical treatment for sepsis in intensive care: from reinforcement learning to pre-trial evaluation , author =

  115. [123]

    Human-Centric Intelligent Systems , publisher =

    Offline safe reinforcement learning for sepsis treatment: Tackling variable-length episodes with sparse rewards , author =. Human-Centric Intelligent Systems , publisher =

  116. [124]

    Offline reinforcement learning with uncertainty for treatment strategies in sepsis , author =

  117. [125]

    PLOS Digital Health , publisher =

    Unifying cardiovascular modelling with deep reinforcement learning for uncertainty aware control of sepsis treatment , author =. PLOS Digital Health , publisher =

  118. [126]

    Communications medicine , publisher =

    Deep reinforcement learning extracts the optimal sepsis treatment policy from treatment records , author =. Communications medicine , publisher =

  119. [127]

    Improving sepsis treatment strategies by combining deep and kernel-based reinforcement learning , author =

  120. [128]

    Machine learning for healthcare conference , pages =

    Continuous state-space models for optimal sepsis treatment: a deep reinforcement learning approach , author =. Machine learning for healthcare conference , pages =

  121. [129]

    Artificial intelligence in medicine , publisher =

    Clinical knowledge-guided deep reinforcement learning for sepsis antibiotic dosing recommendations , author =. Artificial intelligence in medicine , publisher =

  122. [130]

    Biomedical Engineering Letters , publisher =

    Optimizing sepsis treatment strategies via a reinforcement learning model , author =. Biomedical Engineering Letters , publisher =

  123. [131]

    2020 IEEE International conference on healthcare informatics (ICHI) , pages =

    Safe reinforcement learning for sepsis treatment , author =. 2020 IEEE International conference on healthcare informatics (ICHI) , pages =

    2020

  124. [132]

    NPJ Digital Medicine , publisher =

    A value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis , author =. NPJ Digital Medicine , publisher =

  125. [133]

    Model-based reinforcement learning for sepsis treatment , author =

  126. [134]

    PLoS One , publisher =

    Superhuman performance on sepsis MIMIC-III data by distributional reinforcement learning , author =. PLoS One , publisher =

  127. [135]

    Machine Learning for Healthcare Conference , pages =

    Reinforcement learning for sepsis treatment: A continuous action space solution , author =. Machine Learning for Healthcare Conference , pages =

  128. [136]

    Deep reinforcement learning for sepsis treatment , author =

  129. [137]

    Diagnostics , publisher =

    Sepsis trajectory prediction using privileged information and continuous physiological signals , author =. Diagnostics , publisher =

  130. [138]

    Scientific Reports , publisher =

    Continuous sepsis trajectory prediction using tensor-reduced physiological signals , author =. Scientific Reports , publisher =

  131. [139]

    Critical Care , publisher =

    Effect of corticosteroids on the clinical course of community-acquired pneumonia: a randomized controlled trial , author =. Critical Care , publisher =

  132. [140]

    The Pediatric infectious disease journal , publisher =

    Evaluation of early corticosteroid therapy in management of pediatric septic shock in pediatric intensive care patients: a randomized clinical study , author =. The Pediatric infectious disease journal , publisher =

  133. [141]

    Journal of Emergency Practice and Trauma , publisher =

    The effect of physiologic dose of intravenous hydrocortisone in patients with refractory septic shock: a randomized control trial , author =. Journal of Emergency Practice and Trauma , publisher =

  134. [142]

    Jama , volume = 183, number = 6, pages =

    The effectiveness of hydrocortisone in the management of severe infections , author =. Jama , volume = 183, number = 6, pages =

  135. [143]

    American journal of respiratory and critical care medicine , publisher =

    Hydrocortisone infusion for severe community-acquired pneumonia: a preliminary randomized study , author =. American journal of respiratory and critical care medicine , publisher =

  136. [144]

    Sao Paulo Medical Journal , publisher =

    Early dexamethasone treatment for septic shock patients: a prospective randomized clinical trial , author =. Sao Paulo Medical Journal , publisher =

  137. [145]

    Journal of Thoracic Disease , volume = 10, number =

    Steroids for sepsis: yes, no or maybe , author =. Journal of Thoracic Disease , volume = 10, number =

  138. [146]

    Advances in Neural Information Processing Systems , volume = 36, pages =

    Ehrshot: An ehr benchmark for few-shot evaluation of foundation models , author =. Advances in Neural Information Processing Systems , volume = 36, pages =

  139. [147]

    Journal of biomedical informatics , publisher =

    Language models are an effective representation learning technique for electronic health record data , author =. Journal of biomedical informatics , publisher =

  140. [148]

    IEEE Transactions on biomedical engineering , publisher =

    Robust predictive control strategy applied for propofol dosing using BIS as a controlled variable during anesthesia , author =. IEEE Transactions on biomedical engineering , publisher =

  141. [149]

    IEEE Transactions on biomedical engineering , publisher =

    Advanced model-based control studies for the induction and maintenance of intravenous anaesthesia , author =. IEEE Transactions on biomedical engineering , publisher =

  142. [150]

    Closed-loop artificial pancreas using subcutaneous glucose sensing and insulin delivery and a model predictive control algorithm: preliminary studies in Padova and Montpellier , author =

  143. [151]

    Annals of Operations Research , publisher =

    Simulation optimization: a review of algorithms and applications , author =. Annals of Operations Research , publisher =

  144. [152]

    Scientific Reports , publisher =

    Coordinated distributed model predictive control for multi energy carrier systems , author =. Scientific Reports , publisher =

  145. [153]

    Evaluating real-world robot manipulation policies in simulation , author =

  146. [154]

    Mira: Medical time series foundation model for real-world health data , author =

  147. [155]

    NPJ digital medicine , publisher =

    A primer on reinforcement learning in medicine for clinicians , author =. NPJ digital medicine , publisher =

  148. [156]

    Healthcare , volume = 13, number = 14, pages = 1752, organization =

    Reinforcement learning and its clinical applications within healthcare: A systematic review of precision medicine and dynamic treatment regimes , author =. Healthcare , volume = 13, number = 14, pages = 1752, organization =

  149. [157]

    npj Digital Medicine , publisher =

    An interpretable RL framework for pre-deployment modeling in ICU hypotension management , author =. npj Digital Medicine , publisher =

  150. [158]

    Statistics in medicine , publisher =

    Reinforcement learning design for cancer clinical trials , author =. Statistics in medicine , publisher =

  151. [159]

    Jama , volume = 321, number = 20, pages =

    Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis , author =. Jama , volume = 321, number = 20, pages =

  152. [160]

    Multimodal medical code tokenizer , author =

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.