Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Foundation Model of Electronic Medical Records for Adaptive Risk Estimation

T0 review · 2 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a single generative transformer, pretrained on tokenized electronic health records, can predict hospital mortality, ICU admission, prolonged stay, and composite risk with higher AUC than a strong tabular baseline…

desk verdict A useful dynamic risk extension of the authors' own ETHOS generative model, with strong AUCs on MIMIC-IV, but the universal superiority claim over MEDS-Tab is contradicted by their own Table S1. read the letter →

arxiv 2502.06124 v4 pith:WRZ77JKF submitted 2025-02-10 cs.LG cs.AI

Pawel Renc , Michal K. Grzeszczyk , Nassim Oufattole , Deirdre Goode , Yugang Jia , Szymon Bieganski , Matthew B. A. McDermott , Jaroslaw Was
show 4 more authors
Anthony E. Samir Jonathan W. Cunningham David W. Bates Arkadiusz Sitek
This is my paper · ORCID
classification cs.LGcs.AI
keywords GenerativeAIZero-shotinferencePatienthealthtimelinesRiskpredictionElectronicrecordsFoundationmodelEarlywarningsystemsMIMIC-IV
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 claims that one generative foundation model over tokenized electronic health records can replace multiple task-specific early warning systems. The Adaptive Risk Estimation System (ARES) uses ETHOS, a transformer pretrained to predict the next clinical event in a patient's timeline, to simulate many plausible futures and count how often a target event—death, ICU admission, or a stay longer than ten days—occurs. On the MIMIC-IV dataset, ARES reports higher AUC than the MEDS-Tab baseline for every endpoint, including composite risks, and maintains that edge across demographic subgroups. The appeal is that a single model yields dynamic, personalized, explainable probabilities that update as new data arrive, without retraining for each task.

What carries the argument

The central machinery is the Patient Health Timeline (PHT), a chronological token sequence of every coded clinical event (diagnoses, medications, labs, procedures, vitals, demographics) with time-interval tokens between events. ETHOS is a decoder-only transformer trained with next-token cross-entropy on 360 million tokens drawn from 285,622 PHTs. At inference it repeatedly completes a patient's PHT into N=100 simulated futures using top-p nucleus sampling (p=0.9); the estimated probability of an event is M/N, the fraction of futures containing that event's token. This ratio is justified in Appendix A as an unbiased Monte Carlo estimator when sampling is exact, and the method's dynamic and explainable behavior comes from re-running simulations from any earlier token and inspecting which tokens change the counts.

What would settle it

Take the same test patients and re-estimate mortality risk with a much larger number of simulations (say 10,000 instead of 100) and recompute the calibration curve and AUC for hospital mortality; if the estimates or the reported 0.940 AUC change materially, the finite-sample Monte Carlo approximation is not reliable. A complementary check is to run ARES on a second hospital system's records without retraining and see whether the claimed subgroup-robust AUC advantages and calibration persist.

Watch

Extended reading notes

Core claim

The discovery claimed is that an autoregressive transformer trained only to predict the next token of patient health timelines generalizes zero-shot to clinical risk estimation: given a patient's tokenized history, ETHOS generates N=100 future timelines (fPHTs) under nucleus sampling, and the fraction containing an outcome token is the risk probability. ARES applies this to hospital mortality (AUC 0.940), ICU admission (0.932), prolonged stay (0.853), and a composite outcome (0.906) at hospital admission, as well as to the emergency-department benchmarks of hospitalization at triage (0.946), critical outcome within 12 hours (0.945), and 72-hour ED re-presentation (0.745). The paper further claims these risks are well calibrated, robust across race and gender subgroups, and decomposable through an explainability module that highlights which events raise or lower risk.

Load-bearing premise

The central assumption is that 100 randomly generated future timelines from the model faithfully represent what actually happens to patients; if the generated futures are skewed—especially for rare events like in-hospital death—then the reported risk probabilities and their edge over other methods lose their meaning.

Editorial extensions

If this is right

  • One zero-shot model can serve multiple endpoints at once, so hospitals would not need separate trained models for mortality, ICU admission, or length-of-stay prediction.
  • Risk estimates automatically update as new events appear, since each new token extends the patient's timeline before the next round of simulations.
  • Because all simulated futures are generated once, new endpoints can be scored by re-counting tokens in already generated timelines, without retraining.
  • If the reported AUC and calibration hold outside MIMIC-IV, ARES could be deployed as a dynamic early warning system whose five risk levels (0-20%, 20-40%, and so on) are actionable for clinicians.

Reading between the lines

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

  • Editorial inference: the 100-timeline Monte Carlo count has a standard deviation of sqrt(p(1-p)/100) around any true risk p, so for rare endpoints like hospital mortality (1.9% prevalence) the finite-sample wobble is comparable to the calibration error the paper reports; production use may need larger N.
  • Editorial inference: the reported advantage of ETHOS over MEDS-Tab on every demographic subgroup suggests generative pretraining captures patterns that tabular aggregation loses, but whether that advantage survives transfer to other hospital systems is untested and would require external validation.
  • Editorial inference: the explainability traces (e.g., intubation immediately raising ICU risk) may be more useful for retrospective understanding than for real-time alerts, since some procedures occur too late to guide intervention; prospective pilots would clarify this.
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

2 major / 4 minor

Summary. The paper presents ARES, an adaptive risk estimation system built on the previously published ETHOS generative transformer that operates on tokenized Patient Health Timelines (PHTs) from electronic health records. ARES estimates event probabilities by Monte Carlo sampling of future PHTs (N=100 per patient) and counting outcome tokens, producing dynamic risk scores for hospital mortality, ICU admission, prolonged stay, and a composite endpoint, alongside an explainability module. The authors evaluate on MIMIC-IV v2.2 with its ED extension, benchmark against traditional early warning scores and machine learning baselines including MEDS-Tab, and report AUC, calibration, and subgroup-stratified results. The paper also includes ED benchmark tasks (hospitalization at triage, critical outcome within 12 hours, and 72-hour ED re-presentation), discusses limitations, and releases source code.

Significance. If the central claims held, this would be a valuable demonstration that a single zero-shot generative model can produce dynamic, well-calibrated risk estimates for multiple clinical endpoints without task-specific retraining, with the additional benefit of personalized explainability. The paper's strengths include a large-scale tokenized corpus (over 360 million tokens), public code release, calibration reporting with Brier scores, and a candid discussion of limitations. The overall AUC results for the primary endpoints are strong and internally coherent for the unstratified comparisons. However, the paper's headline claim of universal superiority is contradicted by its own supplementary table in one non-negligible demographic subgroup, and the Monte Carlo justification does not fully cover the approximate sampling scheme used at inference. These issues are correctable but require a substantive revision of the paper's claims and presentation.

major comments (2)
  1. [§3, Table S1, Figure 4 caption] The text states that ARES 'consistently outperformed MEDS-Tab across both individual and composite endpoints, achieving higher AUC values in every case,' and Figure 4's caption says that 'ETHOS consistently outperforms MEDS-Tab across all demographic subgroups and tasks.' Table S1 contradicts this: for ICU Admission in the Unknown race subgroup, MEDS-Tab achieves AUROC 0.953 (95% CI 0.942–0.964) while ETHOS achieves 0.941 (95% CI 0.928–0.953). Because the Unknown race group constitutes 42.3% of the test set (Table S2), this is not a negligible subgroup. The confidence intervals overlap, so the accurate statement is that ETHOS is not significantly worse in this subgroup, which is materially weaker than 'higher AUC in every case.' The abstract, Results, and Figure 4 caption should be revised to report the direction and statistical significance of each subgroup comparison rather than claiming uniform superiority.
  2. [§2.1 and Appendix A] Appendix A proves that the Monte Carlo ratio M/N is an unbiased estimator of Pr(A) under the assumption that the fPHTs are drawn exactly from the model distribution p(x). In practice, inference uses nucleus sampling with top-p = 0.9 (Appendix B) and only N = 100 samples, so the proof does not apply to the actual sampling distribution. The paper should either justify that the approximate sampling distribution is sufficiently close to p(x) for the ratio to be a valid risk estimate, or characterize the potential bias, particularly for rare events such as hospital mortality (1.85% prevalence). The calibration curves in Figure S4 provide empirical evidence but do not replace a formal justification for the central probability definition.
minor comments (4)
  1. [Abstract and Table S6] The abstract and Section 3 report 285,622 PHTs, while Table S6 reports a total of 285,620 timelines; similarly, the full text states 299,721 unique patients while Table S2 reports 299,712. These numerical discrepancies should be reconciled.
  2. [§3 and Table S1] Section 3 lists task prevalences as 1.85%, 15.44%, 9.01%, and 20.39%, while Table S1 lists the hospital mortality prevalence as 1.95% and the composite prevalence as 20.41%. Please correct the inconsistent numbers.
  3. [§3 and Figure S4] Section 3 states that Brier scores were in the range 0.01–0.14, but Figure S4 shows a maximum Brier score of 0.094 (for hospitalization at triage). The stated range should be aligned with the figure.
  4. [Figure S1 caption] The caption mentions a 'percentage of discarded ambiguous inference repetitions (0.2–0.3%)' without defining what makes a repetition ambiguous; please provide a concrete definition or remove the phrase.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ARES risk estimates are empirical Monte Carlo counts evaluated on held-out MIMIC-IV labels; self-citation to ETHOS [15] is to an independent published model, not a load-bearing tautology.

full rationale

The central claim that ARES/ETHOS outperforms MEDS-Tab on MIMIC-IV held-out endpoints is an empirical evaluation, not a derivation that reduces to its inputs. Risk probabilities are computed as M/N, the fraction of N=100 generated fPHTs containing the outcome token (Section 2.1), and are then compared against ground-truth labels from the held-out 10% split via AUROC and calibration curves. This is a standard zero-shot evaluation protocol: the model is pretrained on the 90% train/validation split and tested on the disjoint 10% split, with MEDS-Tab and other baselines trained on the same splits. The unbiasedness proof in Appendix A assumes exact sampling from p(x), whereas inference uses nucleus sampling (Appendix B); this is a theoretical gap about the estimator's calibration under approximate sampling, not a circular reduction of the performance claim to its inputs. The paper also contains an internal reporting inconsistency: Table S1 shows MEDS-Tab AUROC 0.953 vs ETHOS 0.941 for ICU admission in the Unknown-race subgroup, contradicting the statement 'higher AUC values in every case'; this is a correctness concern, not circularity. Self-citations to [15] describe the ETHOS architecture, tokenization, and training, which are components used here rather than conclusions derived in this paper; those prior results are independently published, code-released, and externally evaluable, so they do not constitute load-bearing circular evidence.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central performance claim depends on a small number of hand-chosen sampling and tokenization parameters, on the assumption that the generative distribution is clinically meaningful, and on the completeness of structured data; no new physical entities are introduced.

free parameters (6)
  • N (number of sampled future timelines per patient) = 100
    Monte Carlo sample count used for risk estimation; larger N reduces variance but increases compute. Chosen by hand (Section 2.1).
  • Nucleus sampling probability (top-p) = 0.9
    Sampling threshold controlling diversity of generated fPHTs; chosen heuristically (Appendix B).
  • Quantile bin boundaries for continuous values = 10 quantiles (Q1-Q10)
    Lab values and continuous variables are binned into ten quantiles for tokenization; boundaries are data-dependent fits on the training set (Section 2.3).
  • Time-interval token cutpoints = 19 bins from <5m to >=6mo
    Inter-event intervals are discretized into hand-chosen cutpoints, shaping temporal semantics of the PHTs (Section 2.3, Table S11).
  • Risk level thresholds = 20%, 40%, 60%, 80%
    Predicted probabilities are mapped to five ordinal risk levels; cutpoints are chosen, not learned (Section 1, Figures 1 and 2).
  • Prolonged stay outcome threshold = >90th percentile (also reported as >10 days)
    The prolonged-stay outcome is defined as length of stay >90th percentile in the Results but as >10 days in Figure 2 and Table S1; the ambiguity affects the outcome label.
assumptions (4)
  • domain assumption Cross-entropy training on next-token prediction produces a density over token sequences that reflects clinically meaningful future event probabilities.
    The entire ARES system relies on treating the generative next-token distribution as a source of clinically valid risk estimates (Section 2.1, Appendix B).
  • domain assumption Nucleus sampling with p=0.9 draws samples close enough to the model distribution for the Monte Carlo estimator to be unbiased.
    Appendix A proves unbiasedness assuming exact sampling from p(x), but inference uses approximate top-p sampling (Appendix B, step 3); this gap is not quantified.
  • domain assumption The set of tokenized structured EHR events (labs, meds, procedures, diagnoses, vitals, demographics) is sufficient to predict the target outcomes.
    Clinical notes and other unstructured data are excluded (Section 2.2), so predictive signal is limited to structured fields.
  • ad hoc to paper Patients excluded after tokenization (those with no usable structured data) do not bias the benchmark comparison.
    The paper excludes about 14,000 patients with little structured data without analyzing their characteristics (Results, first paragraph).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Foundation Model of Electronic Medical Records for Adaptive Risk Estimation." pith.science (2026). https://pith.science/paper/WRZ77JKF

@misc{pith2026250206124,
  author       = {Pith},
  title        = {Pith review of: Foundation Model of Electronic Medical Records for Adaptive Risk Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WRZ77JKF}},
  note         = {Machine review of arXiv:2502.06124}
}
read the original abstract

Hospitals struggle to predict critical outcomes. Traditional early warning systems, like NEWS and MEWS, rely on static variables and fixed thresholds, limiting their adaptability, accuracy, and personalization. We previously developed the Enhanced Transformer for Health Outcome Simulation (ETHOS), an AI model that tokenizes patient health timelines (PHTs) from EHRs and uses transformer-based architectures to predict future PHTs. ETHOS is a versatile framework for developing a wide range of applications. In this work, we develop the Adaptive Risk Estimation System (ARES) that leverages ETHOS to compute dynamic, personalized risk probabilities for clinician-defined critical events. ARES also features a personalized explainability module that highlights key clinical factors influencing risk estimates. We evaluated ARES using the MIMIC-IV v2.2 dataset together with its Emergency Department (ED) extension and benchmarked performance against both classical early warning systems and contemporary machine learning models. The entire dataset was tokenized resulting in 285,622 PHTs, comprising over 360 million tokens. ETHOS outperformed benchmark models in predicting hospital admissions, ICU admissions, and prolonged stays, achieving superior AUC scores. Its risk estimates were robust across demographic subgroups, with calibration curves confirming model reliability. The explainability module provided valuable insights into patient-specific risk factors. ARES, powered by ETHOS, advances predictive healthcare AI by delivering dynamic, real-time, personalized risk estimation with patient-specific explainability. Although our results are promising, the clinical impact remains uncertain. Demonstrating ARES's true utility in real-world settings will be the focus of our future work. We release the source code to facilitate future research.

Figures

Figures reproduced from arXiv: 2502.06124 by the authors.

Figure 1
Figure 1. Workflow of the Adaptable Risk Estimation Score (ARES) Framework. This figure illustrates the ARES framework, developed on the ETHOS model, for dynamic and explainable risk evaluation. Panel 1 depicts the tokenization of a patient’s entire health history into structured events represented as a sequence of tokens (PHTs), incorporating standardized coding systems such as ATC for medications, ICD-PCS for procedures, an… view at source ↗
Figure 2
Figure 2. Timeline of a Patient’s Hospital Stay and Hypothetical Risk Predictions by ARES. This figure illustrates the timeline of a patient’s hospital stay, from admission to discharge around Day 14, demonstrating how ARES dynamically adjusts its predictions based on the patient’s evolving clinical status and medical history. By Day 5, ARES predicts a high risk of ICU admission, which is subsequently confirmed as the patient… view at source ↗
Figure 3
Figure 3. ARES Risk Trajectories. This figure illustrates risk trajectories for nearly 1000 tokens preceding patient death, as monitored by ARES, which evaluates the probability of death, ICU admission, prolonged hospital stay, and a composite risk score. The lower panel provides a color-coded representation of risk with the actual time since the ED presentation. In contrast, the upper panel highlights three 5-token regions i… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: AUC Comparison Between ETHOS and MEDS-Tab Across Demographic Subgroups and Prediction Tasks. AUC scores with 95% confidence intervals are shown for ETHOS (orange) and MEDS-Tab (gray) across four prediction tasks: Hospital Mortality, ICU Admission, Prolonged Stay, and C…
Figure 5
Figure 5. Figure 5: Predictive results for the ED benchmark tasks. Fewer methods appear in the ED re-presentation task (right) because score-based approaches, designed specifically to estimate in-hospital deterioration, are not applicable once the patient has left the ED. Ethos consistent…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Quantifying surprise in clinical care: Detecting highly informative events in electronic health records with foundation models

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Self-information from a next-token EHR foundation model identifies clinically surprising tokens and events whose counts predict mortality and long length-of-stay, and whose removal degrades representation-based progno...

Reference graph

Works this paper leans on

51 extracted references · 35 canonical work pages · cited by 1 Pith paper

  1. [1]

    National Health Expenditure Data: Historical

    Centers for Medicare & Medicaid Services. National Health Expenditure Data: Historical . https://www. cms . gov / data - research / statistics - trends - and - reports / national - health - expenditure - data/historical. Accessed: 2025-05-03. 2024

    work page 2025

  2. [2]

    Health Care from a Global Perspective, 2022: Accelerating Spending, Worsening Outcomes

    Munira Z Gunja, Evan D Gumas, and Reginald D Williams II.U.S. Health Care from a Global Perspective, 2022: Accelerating Spending, Worsening Outcomes. Jan. 2023. (Visited on 01/26/2025)

    work page 2022

  3. [3]

    Hospital-based emergency care: at the breaking point

    Committee on the Future of Emergency Care in the United States Health System. Hospital-based emergency care: at the breaking point . Washington, D.C., DC: National Academies Press, 2007

    work page 2007

  4. [4]

    Managing emergency department crowding through improved triaging and resource allocation

    Kum Khiong Yang et al. “Managing emergency department crowding through improved triaging and resource allocation”. en. In: Oper . Res. Health Care10 (Sept. 2016), pp. 13–22. DOI: 10.1016/j.orhc.2016.05.001

    work page doi:10.1016/j.orhc.2016.05.001 2016

  5. [5]

    Modified early warning score-based clinical decision support: cost impact and clinical outcomes in sepsis

    Devin J Horton et al. “Modified early warning score-based clinical decision support: cost impact and clinical outcomes in sepsis”. en. In: JAMIA Open 3 (2 July 2020), pp. 261–268. DOI: 10.1093/jamiaopen/ooaa014

    work page doi:10.1093/jamiaopen/ooaa014 2020

  6. [6]

    Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis

    Roy Adams et al. “Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis”. en. In: Nat. Med. 28 (7 July 2022), pp. 1455–1460. DOI: 10.1038/s41591-022-01894-0

    work page doi:10.1038/s41591-022-01894-0 2022

  7. [7]

    Early warning scores with and without artificial intelligence

    Dana P Edelson et al. “Early warning scores with and without artificial intelligence”. en. In: JAMA Netw. Open 7 (10 Oct. 1, 2024), e2438986. DOI: 10.1001/jamanetworkopen.2024.38986

    work page arXiv 2024

  8. [8]

    Early warning scores for detecting deterioration in adult hospital patients: systematic review and critical appraisal of methodology

    Stephen Gerry et al. “Early warning scores for detecting deterioration in adult hospital patients: systematic review and critical appraisal of methodology”. en. In: BMJ 369 (May 20, 2020), p. m1501. DOI: 10.1136/bmj.m1501

    work page doi:10.1136/bmj.m1501 2020

Show all 51 references
  1. [9]

    The impact of a machine learning early warning score on hospital mortality: A multicenter clinical intervention trial

    Christopher J Winslow et al. “The impact of a machine learning early warning score on hospital mortality: A multicenter clinical intervention trial”. en. In: Crit. Care Med. 50 (9 Sept. 1, 2022), pp. 1339–1347. DOI: 10.1097/CCM.0000000000005492

    2022 doi

  2. [10]

    Automated identification of adults at risk for in-hospital clinical deterioration

    Gabriel J Escobar et al. “Automated identification of adults at risk for in-hospital clinical deterioration”. en. In: N. Engl. J. Med. 383 (20 Nov. 12, 2020), pp. 1951–1960. DOI: 10.1056/NEJMsa2001090

    2020 doi

  3. [11]

    External validation and comparison of a general ward deterioration index between diversely different health systems

    Brandon C Cummings et al. “External validation and comparison of a general ward deterioration index between diversely different health systems”. en. In: Crit. Care Med. 51 (6 June 1, 2023), pp. 775–786. DOI: 10.1097/ CCM.0000000000005837

    2023

  4. [12]

    The National Early Warning Score: from concept to NHS implementation

    Bryan Williams. “The National Early Warning Score: from concept to NHS implementation”. en. In: Clin. Med. 22 (6 Nov. 2022), pp. 499–505. DOI: 10.7861/clinmed.2022-news-concept

    2022 doi

  5. [13]

    Validation of a modified Early Warning Score in medical admissions

    C P Subbe et al. “Validation of a modified Early Warning Score in medical admissions”. en. In: QJM 94 (10 Oct. 2001), pp. 521–526. DOI: 10.1093/qjmed/94.10.521

    2001 doi

  6. [14]

    Attention is all you need

    Ashish Vaswani et al. “Attention is all you need”. In: Adv. Neural Inf. Process. Syst. 30 (2017)

    2017

  7. [15]

    Zero shot health trajectory prediction using transformer

    Pawel Renc et al. “Zero shot health trajectory prediction using transformer”. en. In:NPJ Digit. Med. 7 (1 Sept. 19, 2024), p. 256. DOI: 10.1038/s41746-024-01235-0

    2024 doi

  8. [16]

    TransformEHR: transformer-based encoder-decoder generative model to enhance prediction of disease outcomes using electronic health records

    Zhichao Yang et al. “TransformEHR: transformer-based encoder-decoder generative model to enhance prediction of disease outcomes using electronic health records”. en. In: Nat. Commun. 14 (1 Nov. 29, 2023), p. 7857. DOI: 10.1038/s41467-023-43715-z

    2023 doi

  9. [17]

    Hi-BEHRT: Hierarchical Transformer-Based Model for Accurate Prediction of Clinical Events Using Multimodal Longitudinal Electronic Health Records

    Yikuan Li et al. “Hi-BEHRT: Hierarchical Transformer-Based Model for Accurate Prediction of Clinical Events Using Multimodal Longitudinal Electronic Health Records”. en. In: IEEE J Biomed Health Inform 27 (2 Feb. 2023), pp. 1106–1117. DOI: 10.1109/JBHI.2022.3224727

    2023

  10. [18]

    Large language models surpass human experts in predicting neuroscience results

    Xiaoliang Luo et al. “Large language models surpass human experts in predicting neuroscience results”. en. In: Nat. Hum. Behav. (Nov. 27, 2024). DOI: 10.1038/s41562-024-02046-9

    2024 doi

  11. [19]

    Large language models in medicine

    Arun James Thirunavukarasu et al. “Large language models in medicine”. en. In: Nat. Med. 29 (8 Aug. 2023), pp. 1930–1940. DOI: 10.1038/s41591-023-02448-8

    2023 doi

  12. [20]

    Foresight-a generative pretrained transformer for modelling of patient timelines using electronic health records: a retrospective modelling study

    Zeljko Kraljevic et al. “Foresight-a generative pretrained transformer for modelling of patient timelines using electronic health records: a retrospective modelling study”. en. In: Lancet Digit Health 6 (4 Apr. 2024), e281– e290. DOI: 10.1016/S2589-7500(24)00025-6 . 10 Foundat...

    2024 doi

  13. [21]

    Event Stream GPT: A data pre-processing and modeling library for generative, pre-trained transformers over continuous-time sequences of complex events

    Matthew B A McDermott et al. “Event Stream GPT: A data pre-processing and modeling library for generative, pre-trained transformers over continuous-time sequences of complex events”. In: Adv. Neural Inf. Process. Syst. abs/2306.11547 (June 20, 2023). DOI: 10.48550/arXiv.2306.11547

    2023 doi arXiv

  14. [22]

    MOTOR: A time-to-event foundation model for structured medical records

    Ethan Steinberg et al. “MOTOR: A time-to-event foundation model for structured medical records”. In:arXiv [cs.LG] (Jan. 8, 2023)

    2023

  15. [23]

    BEHRT: Transformer for electronic health records

    Yikuan Li et al. “BEHRT: Transformer for electronic health records”. en. In: Sci. Rep. 10 (1 Apr. 28, 2020), p. 7155. DOI: 10.1038/s41598-020-62922-y

    2020 doi

  16. [24]

    Event-Based Contrastive Learning for medical time series

    Hyewon Jeong et al. “Event-Based Contrastive Learning for medical time series”. In: arXiv [cs.LG] (Dec. 15, 2023)

    2023

  17. [25]

    MIMIC-IV, a freely accessible electronic health record dataset

    Alistair E W Johnson et al. “MIMIC-IV, a freely accessible electronic health record dataset”. en. In: Sci Data 10 (1 Jan. 3, 2023), p. 1. DOI: 10.1038/s41597-022-01899-x

    2023 doi

  18. [26]

    Mimic-iv

    Alistair Johnson et al. “Mimic-iv”. In: PhysioNet. Available online at: https://physionet. org/content/mimiciv/2.2/(accessed Oct 1, 2023) (2023)

    2023

  19. [27]

    MEDS-DEV: Establishing Reproducibility and Comparability in Health AI

  20. [28]

    Medical Event Data Standard (MEDS): Facilitating Machine Learning for Health

    Bert Arnrich et al. “Medical Event Data Standard (MEDS): Facilitating Machine Learning for Health”. In: ICLR 2024 Workshop on Learning from Time Series F or Health. 2024

    2024

  21. [29]

    Benchmarking emergency department prediction models with machine learning and public electronic health records

    Feng Xie et al. “Benchmarking emergency department prediction models with machine learning and public electronic health records”. en. In: Sci. Data 9 (1 Oct. 27, 2022), p. 658. DOI: 10.1038/s41597-022-01782-9

    2022 doi

  22. [30]

    The ability of the National Early Warning Score (NEWS) to discriminate patients at risk of early cardiac arrest, unanticipated intensive care unit admission, and death

    Gary B Smith et al. “The ability of the National Early Warning Score (NEWS) to discriminate patients at risk of early cardiac arrest, unanticipated intensive care unit admission, and death”. en. In: Resuscitation 84 (4 Apr. 2013), pp. 465–470. DOI: 10.1016/j.resuscitation.2012.12.016

    2013 doi

  23. [31]

    A multimodal biomedical foundation model trained from fifteen million image–text pairs

    Sheng Zhang et al. “A multimodal biomedical foundation model trained from fifteen million image–text pairs”. en. In: NEJM AI 2 (1 Jan. 2025). DOI: 10.1056/aioa2400640

    2025 doi

  24. [32]

    Rapid Emergency Medicine score: a new prognostic tool for in-hospital mortality in nonsurgical emergency department patients

    T Olsson, A Terent, and L Lind. “Rapid Emergency Medicine score: a new prognostic tool for in-hospital mortality in nonsurgical emergency department patients”. en. In: J. Intern. Med. 255 (5 May 2004), pp. 579–587. DOI: 10.1111/j.1365-2796.2004.01321.x

    2004

  25. [33]

    Derivation of a cardiac arrest prediction model using ward vital signs

    Matthew M Churpek et al. “Derivation of a cardiac arrest prediction model using ward vital signs”. en. In: Crit. Care Med. 40 (7 July 2012), pp. 2102–2108. DOI: 10.1097/CCM.0b013e318250aa5a

    2012 doi

  26. [34]

    The Emergency Severity Index triage algorithm version 2 is reliable and valid

    David R Eitel et al. “The Emergency Severity Index triage algorithm version 2 is reliable and valid”. en. In:Acad. Emerg. Med. 10 (10 Oct. 2003), pp. 1070–1080. DOI: 10.1197/s1069-6563(03)00350-6

    2003 doi

  27. [35]

    Multi-layer representation learning for medical concepts

    Edward Choi et al. “Multi-layer representation learning for medical concepts”. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining . KDD ’16: The 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining ...

    2016

  28. [36]

    Long short-term memory

    S Hochreiter and J Schmidhuber. “Long short-term memory”. en. In: Neural Comput. 9 (8 Nov. 15, 1997), pp. 1735–1780. DOI: 10.1162/neco.1997.9.8.1735

    1997 doi

  29. [37]

    MEDS-Tab: Automated tabularization and baseline methods for MEDS datasets

    Nassim Oufattole et al. “MEDS-Tab: Automated tabularization and baseline methods for MEDS datasets”. In: arXiv [cs.LG] (Oct. 31, 2024)

    2024

  30. [38]

    XGBoost: A Scalable Tree Boosting System

    Tianqi Chen and Carlos Guestrin. “XGBoost: A Scalable Tree Boosting System”. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (San Francisco, California, USA). KDD ’16. New York, NY , USA: Association for Computing Machiner...

    2016

  31. [39]

    SciPy 1.0: fundamental algorithms for scientific computing in Python

    Pauli Virtanen et al. “SciPy 1.0: fundamental algorithms for scientific computing in Python”. en. In:Nat. Methods 17 (3 Mar. 2020), pp. 261–272. DOI: 10.1038/s41592-019-0686-2

    2020 doi

  32. [40]

    Scikit-learn: Machine learning in python journal of machine learning research

    F Pedregosa et al. “Scikit-learn: Machine learning in python journal of machine learning research”. In: Journal of machine learning research 12 (2011), pp. 2825–2830

    2011

  33. [41]

    Interpretability and fairness evaluation of deep learning models on MIMIC-IV dataset

    Chuizheng Meng et al. “Interpretability and fairness evaluation of deep learning models on MIMIC-IV dataset”. en. In: Sci. Rep. 12 (1 May 3, 2022), p. 7166. DOI: 10.1038/s41598-022-11012-2

    2022 doi

  34. [42]

    The early warning paradox

    Hugh Logan Ellis et al. “The early warning paradox”. en. In: NPJ Digit. Med. 8 (1 Feb. 3, 2025), p. 81. DOI: 10.1038/s41746-024-01408-x

    2025 doi

  35. [43]

    AutoScore: a machine learning–based automatic clinical score generator and its application to mortality prediction using electronic health records

    Feng Xie et al. “AutoScore: a machine learning–based automatic clinical score generator and its application to mortality prediction using electronic health records”. In: JMIR medical informatics 8.10 (2020), e21798. 11 Foundation Model of EMR for Adaptive Risk Estimation Suppl...

    2020

  36. [44]

    samples {x(1), x(2),

    Draw N i.i.d. samples {x(1), x(2), . . . ,x(N )} from the model p(x)

  37. [45]

    Define an indicator function I(x(i) ∈ A), which is 1 if the sample x(i) lies in A, and 0 otherwise

  38. [46]

    successes

    Estimate Pr(A) by the ratio ˆPr(A) = 1 N NX i=1 I x(i) ∈ A . In other words, ˆPr(A) is simply the fraction of samples whose corresponding timelines satisfy event A indicated as M/N in the text. B. Unbiasedness If the samples x(i) are drawn exactly from p(x), then for each samp...

  39. [47]

    Story.” Each Patient Health Timeline (PHT) is a long sequence of discrete tokens, think of each token as a “word

    Treating Patient History as a “Story.” Each Patient Health Timeline (PHT) is a long sequence of discrete tokens, think of each token as a “word” in a clinical narrative. Static tokens (e.g., demographics) set the scene, categorical tokens (e.g., diagnoses) are like domain-spec...

  40. [48]

    zero-shot

    Zero-Shot Generative Inference. Because ETHOS has been trained to model the joint distribution of these tokens, at inference time it can continue any PHT without further training. We provide the model with the tokens of a patient’s past, and then let it generate new tokens one...

  41. [49]

    completions

    Exploring Possible Futures via Sampling. To capture uncertainty, we perform many independent “completions” of the PHT. Each completion (an fPHT) is analogous to asking, “What could happen next?” and letting the model write the next chapter of the patient’s story. We use nucleu...

  42. [50]

    Some tokens are can be treated as outcome (e.g., mortality)

    From Generated Tokens to Risk Estimates. Some tokens are can be treated as outcome (e.g., mortality). To compute inpatient mortality, we simply count how many of the N simulated fPHTs include that token while patient is in hospital (before discharge token). The fraction M/N th...

  43. [51]

    End of timeline

    Toward Controlled Cohort Generation. Because ETHOS is a pure generator, one can modify the sampling distribution, reweighting certain token types (e.g., age or comorbidity tokens), to synthesize patient cohorts with desired characteristics. This capability opens the door to fa...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.