Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Federated Timeline Synthesis: Scalable and Private Methodology For Model Training and Deployment

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that FTS, a federated loop in which clients share only trained generative-model weights, produces a global generator whose zero-shot clinical predictions on MIMIC-IV are comparable to a model trained on real data.

desk verdict The federated synthesis idea is new and worth a referee, but the nested data splits and test-set temperature selection undercut the central empirical claim. read the letter →

arxiv 2506.23358 v1 pith:2OIIH36K submitted 2025-06-29 cs.LG cs.AI

Pawel Renc , Michal K. Grzeszczyk , Linglong Qian , Nassim Oufattole , Jeff Rasley , Arkadiusz Sitek This is my paper
classification cs.LGcs.AI
keywords federatedlearningsyntheticelectronichealthrecordspatienttimelinesgenerativetransformerszero-shotpredictionprivacy-preservingmodeltrainingMIMIC-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

Federated Timeline Synthesis (FTS) is a framework for training a generative foundation model on patient histories that stay inside their home institutions. Each hospital trains a local autoregressive transformer on its own tokenized Patient Health Timelines and sends only the trained weights to a central server; the server samples synthetic timelines from these generators and trains a Global Generator (GG) on the combined synthetic corpus. The paper's central claim is that this weight-only exchange provides strong privacy guarantees without expensive cryptographic machinery while producing a GG whose zero-shot predictions on five MIMIC-IV tasks are comparable to a model trained directly on real data. The key empirical result is that combining a small local dataset with synthetic timelines from a model trained on a large dataset (small+big_synth) nearly closes the gap to the big real-data model, whereas the small real-data model alone performs far worse. If true, this would let hospitals pool clinical knowledge without pooling records, and enable zero-shot risk prediction and simulation from a single deployed generator.

What carries the argument

The central object is the Patient Health Timeline (PHT), a language-agnostic token sequence in which each clinical event becomes discrete tokens: interval tokens for elapsed time, quantile tokens for continuous values, and hierarchical tokens for codes such as ICD and ATC. The mechanism that carries the argument is two-stage synthesis: each client trains an autoregressive transformer over its local PHTs with a negative log-likelihood objective, transmits only parameters, and the server samples pseudo-PHTs from each generator and trains the Global Generator on the union. Zero-shot inference then works by Monte Carlo simulation of future PHTs, aggregating statistics over sampled continuations to predict binary, multiclass, or regression outcomes.

What would settle it

Re-run the FTS pipeline with the small and big client splits drawn from completely disjoint patient sets from MIMIC-IV, then compare small+big_synth against the big real-data model on the same five tasks; if the combined score no longer approaches the big model's score, the reported gain comes from overlapping training data rather than federated synthesis.

Watch

Extended reading notes

Core claim

The paper claims that FTS reduces cross-institutional generative modeling to a two-stage synthesis loop: local autoregressive generators trained on private Patient Health Timelines are transmitted as weights, the server uses them to synthesize a large corpus of pseudo-patient timelines, and a Global Generator trained on that corpus inherits enough of the joint distribution to support zero-shot prediction. In the MIMIC-IV experiment, models trained on synthetic data generated by GG perform comparably to models trained on real data, and the combination small+big_synth approaches the full big-data performance, with overlapping confidence intervals for big versus big+small_synth. The paper also reports that models trained exclusively on synthetic data underperform their real-data counterparts, so the claim is approximate parity with a modest, acknowledged information-loss gap.

Load-bearing premise

The load-bearing premise is that the simulated institutions are independent, but in the experiment the 'small' split (20%) is a subset of the 'big' split (80%), so the synthetic corpora share overlapping patient histories and the aggregation results may not generalize to truly separate hospitals.

Editorial extensions

If this is right

  • A hospital with only 10 to 20 percent of the data can recover much of the performance of a large-data model by adding synthetic timelines generated by a model trained elsewhere, without sharing raw records.
  • The Global Generator can be deployed to sites that did not contribute data, supporting zero-shot prediction and cohort simulation with no task-specific retraining.
  • Because only weights are exchanged and synthesis happens centrally, communication cost is a one-time transfer per client rather than iterative gradient exchange.
  • Models trained only on synthetic timelines remain somewhat below real-data performance, so FTS as presented is a data-augmentation and knowledge-pooling tool rather than a lossless substitute for real data.

Reading between the lines

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

  • If the approach scales to truly heterogeneous institutions, the PHT token vocabulary itself becomes a coordination artifact: sites must share tokenization, interval bins, and quantile definitions for the weights to be composable, so the claim of no task-specific coordination may not cover tokenization alignment.
  • A natural testable extension is to measure privacy empirically, such as membership-inference or reconstruction attempts against the Global Generator, rather than relying on the assertion that tokenization and weight-only exchange are inherently private.
  • The nested-split design means the multi-source aggregation result should be re-validated on disjoint client cohorts before deployment; if the overlap drove the gain, the real-world benefit for independent hospitals may be smaller.
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 / 5 minor

Summary. The paper proposes Federated Timeline Synthesis (FTS), a two-stage framework in which each institution trains an autoregressive transformer on tokenized patient health timelines (PHTs) and transmits only the trained weights to a central server; the server then samples a large synthetic corpus from the received generators, trains a global generator (GG) on that corpus, and deploys it for zero-shot Monte Carlo inference on downstream clinical prediction tasks. The authors evaluate FTS on MIMIC-IV by creating big (80%), small (20%), and little (10%) subsets, generating synthetic datasets at four temperatures, and training downstream models on real, synthetic, and mixed data. They report that models trained on synthetic data perform comparably to real data, and that synthetic augmentation improves low-resource performance. The paper also claims strong privacy guarantees from weight-only exchange.

Significance. If the federation mechanism worked as described, FTS would be a meaningful contribution: it avoids iterative gradient exchange, needs no task-specific heads, and provides zero-shot inference from a single generative model. The PHT tokenization builds on ETHOS and is described clearly. The paper publishes code, and the fidelity metrics (Unigram and DimWise R2) are appropriate for the tokenized setting. However, the empirical support is weakened by the nested data-split design, test-set-based temperature selection, and an overall score whose normalization is table-dependent. The paper is honest about several limitations (single dataset, no multimodal data), but the abstract and Section 1 overstate the privacy and comparability results. With disjoint client splits and proper validation, the central idea remains worth testing.

major comments (4)
  1. [Section 3, Training Data Division] The federated simulation does not use disjoint client data. The text states that big (80%), small (20%), and little (10%) are subsets of the same orig cohort, with 'little (10%, a subset of small)'. Because small is contained in big, any synthetic corpus generated by the big generator (big_synth) contains information about the patients held by the small client. Consequently, the results for 'small+big_synth' and 'big_synth+small_synth' in Table 6 and Fig. 3 cannot be interpreted as cross-institution transfer or aggregation of independent generators. The defining mechanism of FTS—server-side aggregation of generators trained on disjoint client data—is therefore not empirically tested. The authors should re-run the experiments with disjoint splits (e.g., a separate 20% holdout and an 80% that excludes it) or with independent hospital datasets, and report whether the gains persist.
  2. [Section 3, Synthetic Data Generation Tuning] The generation temperature for all synthetic corpora was selected on the test split. The text says stages (3) and (4) 'are carried out on test', and the synthetic-data tuning procedure evaluates all generated datasets on the test split. The final evaluation in Table 6 is also on the same test split. This is a test-set selection step: temperature 1.0 was chosen because it maximized test performance, and the same test set is then used to report the gains of synthetic-data-trained models. The reported advantage of synthetic data over real-data baselines is therefore optimistically biased, and the magnitude of the bias is unknown. The temperature should be selected on val1/val2 (as was done for inference temperature) or via cross-validation, with test results reported for all candidate temperatures.
  3. [Section E, Overall Score Computation; Tables 5-6] The 'Overall Score' is a min-max normalized inverse-variance weighted sum, with normalization computed across the models present in each table. As a result, the same model receives different overall scores when placed in different comparison sets: the big_synth model has identical per-task metrics in Table 5 and Table 6 (DRG accuracy 0.648, R2 0.559, readmission AUC 0.753, ICU AUC 0.899, mortality AUC 0.909), yet its overall score is 0.418 in Table 5 and 0.813 in Table 6. This makes the overall score unsuitable for cross-table comparisons and for the statement that models trained on synthetic data 'perform comparably' to real-data models: in Table 6, the best fully synthetic model (big_synth) and the aggregated synthetic model (big_synth+small_synth) score 0.813 and 0.792 respectively against big's 1.000, with DRG accuracy dropping from 0.740 to 0.638. The authors should report raw metrics and a fixed-reference normalization, or avoid absolute 'comparable' language.
  4. [Section 1, Significance of Federated Timeline Synthesis; Abstract] The paper claims 'strong privacy guarantees' and 'strong privacy guarantees without expensive cryptographic machinery' from the fact that only model weights are exchanged. No privacy analysis, threat model, membership-inference evaluation, or differential-privacy mechanism is provided. Generative transformers can memorize training examples, and model-weight sharing is known to enable inversion and extraction attacks. Without an empirical privacy evaluation or a clearly scoped statement of what 'strong' means, this claim is unsupported. The authors should either add a privacy evaluation (e.g., membership inference or distance-to-record tests) or substantially weaken the privacy claim in the abstract and introduction.
minor comments (5)
  1. [Section 2.1; Section 1] There are typos: 'optimzied' should be 'optimized' in Section 2.1, and Section 1 contains 'balancing and and fairness'.
  2. [Figure 2] The label 'k1 S GG Deploy Aggregate k2 k3 kK' is difficult to parse; please redraw with clear arrows and labels for the server and deployment path.
  3. [Table 5 caption; Table 6] The Table 5 caption says 'We did not generate synthetic data for little because the model overfitted', yet Table 6 includes the row 'little_synth'. Please reconcile this inconsistency.
  4. [Table 4] The overall-score confidence intervals for different inference temperatures overlap heavily (e.g., 0.968 [0.688, 1.000] for 0.9 and 0.846 [0.563, 1.000] for 1.0), so the claim that 0.9 'yields the best results' is not statistically supported; report pairwise tests or a calibration-based justification.
  5. [Section 2.1] The temporal ordering notation uses (τp,i, sp,i) but the secondary key sp,i is defined only in words; please formalize the tie-breaking rule or use a single index.

Circularity Check

2 steps flagged · score 4.0 of 10

Nested 'institutions' and test-set temperature tuning make the federated-synthesis gains partly self-referential.

  1. self definitional [Section 3, 'Training Data Division' and Section 3, 'Evaluation of Federated Timeline Synthesis Scenarios']
    "We assume K clients, each holding a disjoint set of tokenized Patient Health Timelines (PHTs). ... we define the following subsets of orig: big (80%), small (20%), and little (10%, a subset of small). These partitions are used in subsequent experiments to emulate institutions with varying levels of data availability. ... aggregating synthetic data from multiple sources (big_synth+small_synth) outperforms using the real small dataset alone."

    The experimental 'institutions' are not disjoint: small is a subset of big by construction, so the generator trained on big has already seen every patient assigned to the small client. Consequently, synthetic PHTs drawn from the big model encode the small client's own records. The two headline results that motivate the federated mechanism (small+big_synth and big_synth+small_synth) are therefore inflated by self-augmentation of the small data reappearing through the big generator, rather than by knowledge transfer from an independent institution. The claimed cross-institution aggregation effect reduces to the nested split definition.

  2. fitted input called prediction [Section 3, 'Synthetic Data Generation Tuning' and Section 3, 'Evaluation of Federated Timeline Synthesis Scenarios']
    "Stages (1) and (2) are evaluated on val1 and val2 to prevent overfitting to the test set, while stages (3) and (4) are carried out on test. ... We evaluate all generated datasets on the test split and compute the overall score across all DTs. ... The results indicate that a generation temperature of 1.0 yields the best performance."

    The generation temperature (1.0) used to create every synthetic corpus in Table 6 and Figure 3 is selected by evaluating synthetic datasets on the test split, and the final federated-scenario evaluation is also performed on the test split. The headline scenario results are therefore read off the same evaluation set used to tune the generation hyperparameter, so the reported gains are partially fitted to the test set rather than being held-out predictions.

full rationale

The core generative-training equations are self-contained: local generators minimize a standard negative log-likelihood over local PHTs, the server samples a synthetic corpus and trains GG on it, and zero-shot inference is Monte Carlo simulation from the trained generator. Those steps do not reduce by definition to each other, and the downstream evaluation on held-out MIMIC-IV test patients is a genuine transfer test. The circularity is in the evaluation design rather than in the method's arithmetic. First, the simulated clients are nested subsets of one cohort, so the 'large institution' already contains the 'small institution's' patients; the central federated aggregation evidence (small+big_synth, big_synth+small_synth) is contaminated by construction and does not demonstrate disjoint-client federation. Second, the generation temperature is tuned on the test split and then the headline scenario results are reported on that same split, making the final numbers partly fitted inputs rather than predictions. The ETHOS self-citation in Appendix A is provenance for the timeline representation and is not scored as load-bearing circularity because the FTS pipeline is independently evaluated here.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claims rest on several assumptions that are either unverified or violated by the experimental design: that model weights do not leak patient data, that synthetic timelines preserve joint statistics, and that the simulated institutions are independent. The hyperparameters for inference and generation temperature are selected from the data and should be treated as free parameters.

free parameters (4)
  • Inference temperature = 0.9
    Selected on the val2 split for zero-shot generation during downstream evaluation (Sec. 3, Inference Temperature Selection).
  • Synthetic generation temperature = 1.0
    Selected on the test split (Sec. 3, Synthetic Data Generation Tuning), a source of data leakage.
  • Number of simulated trajectories N = Not reported
    Zero-shot inference probabilities are estimated from N generated fPHTs (Eq. 6), but N is not given.
  • Tokenization hyperparameters = Q=10, fixed time bins
    Quantile and interval bins are chosen by hand from prior work (ETHOS), not optimized here.
assumptions (3)
  • domain assumption Model weights exchanged between clients and server do not reveal identifiable patient information
    The paper's privacy claim rests on this, but no differential privacy or empirical leakage test is provided.
  • domain assumption Synthetic timelines generated from client models preserve the joint distribution needed for downstream tasks
    The fidelity metrics used (unigram and dimension-wise R2) test only marginal distributions.
  • domain assumption Simulated clients hold disjoint data
    The experiments use nested subsets (big, small, little) of MIMIC, so this condition is violated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Federated Timeline Synthesis: Scalable and Private Methodology For Model Training and Deployment." pith.science (2026). https://pith.science/paper/2OIIH36K

@misc{pith2026250623358,
  author       = {Pith},
  title        = {Pith review of: Federated Timeline Synthesis: Scalable and Private Methodology For Model Training and Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2OIIH36K}},
  note         = {Machine review of arXiv:2506.23358}
}
read the original abstract

We present Federated Timeline Synthesis (FTS), a novel framework for training generative foundation models across distributed timeseries data applied to electronic health records (EHR). At its core, FTS represents patient history as tokenized Patient Health Timelines (PHTs), language-agnostic sequences encoding temporal, categorical, and continuous clinical information. Each institution trains an autoregressive transformer on its local PHTs and transmits only model weights to a central server. The server uses the generators to synthesize a large corpus of trajectories and train a Global Generator (GG), enabling zero-shot inference via Monte Carlo simulation of future PHTs. We evaluate FTS on five clinically meaningful prediction tasks using MIMIC-IV data, showing that models trained on synthetic data generated by GG perform comparably to those trained on real data. FTS offers strong privacy guarantees, scalability across institutions, and extensibility to diverse prediction and simulation tasks especially in healthcare, including counterfactual inference, early warning detection, and synthetic trial design.

Figures

Figures reproduced from arXiv: 2506.23358 by the authors.

Figure 1
Figure 1. Tokenization and model training introduce two layers of anonymization. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Federated Time￾line Synthesis (FTS) workflow. Clients (k1 to kK) train lo￾cal generative transformers on their PHTs and send trained generative models to a cen￾tral server. The global gen￾erator (GG) is trained on the generated output of models. The trained GG can then be deployed to both contributing and unseen sites (S) for zero￾shot inference. Significance of Federated Timeline Synthesis By converting het￾erogene… view at source ↗
Figure 3
Figure 3. Overall score for downstream tasks across various training [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of unigram code distributions between real and synthetic data for the [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Calibration curves for three binary downstream tasks (30-day readmission, ICU admission, [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Calibration curves for three binary downstream tasks for model trained on data generated at [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages

  1. [1]

    Synthesizing electronic health records using improved generative adversarial networks

    Mrinal Kanti Baowaly, Chia-Ching Lin, Chao-Lin Liu, and Kuan-Ta Chen. Synthesizing electronic health records using improved generative adversarial networks. Journal of the American Medical Informatics Association, 26(3):228–241, 2019

    work page 2019

  2. [2]

    Fedsyn: Synthetic data generation using federated learning

    Monik Raj Behera, Sudhir Upadhyay, Suresh Shetty, Sudha Priyadarshini, Palka Patel, and Ker Farn Lee. Fedsyn: Synthetic data generation using federated learning. arXiv preprint arXiv:2203.05931, 2022

    arXiv 2022

  3. [3]

    Generating multi-label discrete patient records using generative adversarial networks

    Edward Choi, Siddharth Biswal, Bradley Malin, Jon Duke, Walter F Stewart, and Jimeng Sun. Generating multi-label discrete patient records using generative adversarial networks. In Machine learning for healthcare conference, pages 286–305. PMLR, 2017

    work page 2017

  4. [4]

    Survey of medical applications of federated learning

    Geunho Choi, Won Chul Cha, Se Uk Lee, and Soo-Yong Shin. Survey of medical applications of federated learning. Healthcare Informatics Research, 30(1):3–15, 2024

    work page 2024

  5. [5]

    Clinicalbert: Modeling clinical notes and predicting hospital readmission

    Kai Huang, Jan Altosaar, and Rajesh Ranganath. Clinicalbert: Modeling clinical notes and predicting hospital readmission. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 207–212. IEEE, 2019

    work page 2019

  6. [6]

    Emerging trends in federated learning: From model fusion to federated x learning

    Shaoxiong Ji, Yue Tan, Teemu Saravirta, Zhiqin Yang, Yixin Liu, Lauri Vasankari, Shirui Pan, Guodong Long, and Anwar Walid. Emerging trends in federated learning: From model fusion to federated x learning. International Journal of Machine Learning and Cybernetics, 15(9):3769–3790, 2024

    work page 2024

  7. [7]

    Mimic-iv, a freely accessible electronic health record dataset

    Alistair EW Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J Pollard, Sicheng Hao, Benjamin Moody, Brian Gow, et al. Mimic-iv, a freely accessible electronic health record dataset. Scientific data, 10(1):1, 2023

    2023

  8. [8]

    Scaffold: Stochastic controlled averaging for federated learning

    Sai Praneeth Karimireddy, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. Scaffold: Stochastic controlled averaging for federated learning. In Inter- national conference on machine learning, pages 5132–5143. PMLR, 2020

    work page 2020

Show all 37 references
  1. [9]

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

    Zeljko Kraljevic, Dan Bean, Anthony Shek, Rebecca Bendayan, Harry Hemingway, Joshua Au Yeung, Alexander Deng, Alfred Balston, Jack Ross, Esther Idowu, et al. Foresight—a generative pretrained transformer for modelling of patient timelines using electronic health records: a ret...

    2024

  2. [10]

    Biobert: a pre-trained biomedical language representation model for biomedical text mining

    Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Jiwon So, Jaewook Kang, Seokho Hwang, Minbyul Kim, Youhan Lee, Min Seo, et al. Biobert: a pre-trained biomedical language representation model for biomedical text mining. Bioinformatics, 36(4):1234–1240, 2020

    2020

  3. [11]

    Federated optimization in heterogeneous networks

    Tian Li, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. Federated optimization in heterogeneous networks. Proceedings of Machine learning and systems, 2:429–450, 2020

    2020

  4. [12]

    Behrt: transformer for electronic health records

    Yikuan Li, Shishir Rao, José Roberto Ayala Solares, Abdelaali Hassaine, Rema Ramakrishnan, Dexter Canoy, Yajie Zhu, Kazem Rahimi, and Gholamreza Salimi-Khorshidi. Behrt: transformer for electronic health records. Scientific reports, 10(1):7155, 2020

    2020

  5. [13]

    Trading off scalability, privacy, and performance in data synthesis

    Xiao Ling, Tim Menzies, Christopher Hazard, Jack Shu, and Jacob Beel. Trading off scalability, privacy, and performance in data synthesis. IEEE Access, 12:26642–26654, 2024

    2024

  6. [14]

    Federated learning for generating synthetic data: a scoping review

    Claire Little, Mark Elliot, and Richard Allmendinger. Federated learning for generating synthetic data: a scoping review. International Journal of Population Data Science, 8(1):2158, 2023

    2023

  7. [15]

    Recent advances on federated learning: A systematic survey

    Bingyan Liu, Nuoyan Lv, Yuanchun Guo, and Yawen Li. Recent advances on federated learning: A systematic survey. Neurocomputing, page 128019, 2024

    2024

  8. [16]

    Communication-efficient learning of deep networks from decentralized data

    Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273–1282. PMLR, 2017

    2017

  9. [17]

    The eicu collaborative research database, a freely available multi-center database for critical care research

    Tom J Pollard, Alistair EW Johnson, Jesse D Raffa, Leo A Celi, Roger G Mark, and Omar Badawi. The eicu collaborative research database, a freely available multi-center database for critical care research. Scientific data, 5(1):1–13, 2018

    2018

  10. [18]

    How deep is your guess? a fresh perspective on deep learning for medical time-series imputation

    Linglong Qian, Hugh Logan Ellis, Tao Wang, Jun Wang, Robin Mitra, Richard Dobson, and Zina Ibrahim. How deep is your guess? a fresh perspective on deep learning for medical time-series imputation. IEEE Journal of Biomedical and Health Informatics, 2025

    2025

  11. [19]

    Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction

    Laila Rasmy, Yang Xiang, Ziqian Xie, Cui Tao, and Degui Zhi. Med-bert: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction. NPJ digital medicine, 4(1):86, 2021

    2021

  12. [20]

    Zero shot health trajectory prediction using transformer

    Pawel Renc, Yugang Jia, Anthony E Samir, Jaroslaw Was, Quanzheng Li, David W Bates, and Arkadiusz Sitek. Zero shot health trajectory prediction using transformer. NPJ Digital Medicine, 7(1):256, 2024

    2024

  13. [21]

    Motor: a time-to-event foundation model for structured medical records

    Ethan Steinberg, Jason Fries, Yizhe Xu, and Nigam Shah. Motor: a time-to-event foundation model for structured medical records. arXiv preprint arXiv:2301.03150, 2023. 10

    2023 arXiv

  14. [22]

    Synthesize high-dimensional longitudinal electronic health records via hierarchical autoregressive language model

    Brandon Theodorou, Cao Xiao, and Jimeng Sun. Synthesize high-dimensional longitudinal electronic health records via hierarchical autoregressive language model. Nature communications, 14(1):5305, 2023

    2023

  15. [23]

    Patrick J Thoral, Jan M Peppink, Ronald H Driessen, Eric JG Sijbrands, Erwin JO Kompanje, Lewis Kaplan, Heatherlee Bailey, Jozef Kesecioglu, Maurizio Cecconi, Matthew Churpek, et al. Sharing icu patient data responsibly under the society of critical care medicine/european soci...

    2021

  16. [24]

    Differentially private synthetic medical data generation using convolutional gans

    Amirsina Torfi, Edward A Fox, and Chandan K Reddy. Differentially private synthetic medical data generation using convolutional gans. Information Sciences, 586:485–500, 2022

    2022

  17. [25]

    Attention is all you need

    Ashish Vaswani and et al. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  18. [26]

    Amol A Verma, Yishan Guo, Janice L Kwan, Lauren Lapointe-Shaw, Shail Rawal, Terence Tang, Adina Weinerman, Peter Cram, Irfan A Dhalla, Stephen W Hwang, et al. Patient characteristics, resource use and outcomes associated with general internal medicine hospital care: the genera...

    2017

  19. [27]

    Generation of synthetic electronic health records using a federated gan

    John Weldon, Tomas Ward, and Eoin Brophy. Generation of synthetic electronic health records using a federated gan. arXiv preprint arXiv:2109.02543, 2021

    2021 arXiv

  20. [28]

    The shaky foundations of large language models and foundation models for electronic health records

    Michael Wornow, Yizhe Xu, Rahul Thapa, Birju Patel, Ethan Steinberg, Scott Fleming, Michael A Pfeffer, Jason Fries, and Nigam H Shah. The shaky foundations of large language models and foundation models for electronic health records. npj digital medicine, 6(1):135, 2023

    2023

  21. [29]

    A large language model for electronic health records

    Xi Yang, Aokun Chen, Nima PourNejatian, Hoo Chang Shin, Kaleb E Smith, Christopher Parisien, Colin Compas, Cheryl Martin, Anthony B Costa, Mona G Flores, et al. A large language model for electronic health records. NPJ digital medicine, 5(1):194, 2022

    2022

  22. [30]

    Ehr-safe: generating high-fidelity and privacy-preserving synthetic electronic health records

    Jinsung Yoon, Michel Mizrahi, Nahid Farhady Ghalaty, Thomas Jarvinen, Ashwin S Ravi, Peter Brune, Fanyu Kong, Dave Anderson, George Lee, Arie Meir, et al. Ehr-safe: generating high-fidelity and privacy-preserving synthetic electronic health records. NPJ digital medicine, 6(1):...

    2023

  23. [31]

    Federated learning: Overview, strategies, applications, tools and future directions

    Betul Yurdem, Murat Kuzlu, Mehmet Kemal Gullu, Ferhat Ozgur Catak, and Maliha Tabassum. Federated learning: Overview, strategies, applications, tools and future directions. Heliyon, 2024

    2024

  24. [32]

    Generating clinically realistic ehr data via a hierarchy-and semantics-guided transformer

    Guanglin Zhou and Sebastiano Barbieri. Generating clinically realistic ehr data via a hierarchy-and semantics-guided transformer. arXiv preprint arXiv:2502.20719, 2025. 11 A Timeline Implementation for Federated Synthetic EHR Generation Several timeline representations have be...

    2025 arXiv

  25. [33]

    The prediction is made based on the entire available patient history up to the point of admission

    DRG Prediction (Multiclass Classification): The model generates a single token repre- senting the most likely Diagnosis-Related Group (DRG) code associated with a patient’s hospital stay. The prediction is made based on the entire available patient history up to the point of a...

  26. [34]

    The model regresses the score based on historical clinical data up to the time of assessment

    SOFA Score Prediction (Regression): This task involves predicting the Sequential Organ Failure Assessment (SOFA) score, a continuous measure quantifying the extent of a patient’s organ dysfunction. The model regresses the score based on historical clinical data up to the time ...

  27. [35]

    The generation starts from the last token indicating hospital discharge and continues forward in time

    30-day Readmission (Binary Classification): The model predicts whether a patient will be readmitted to the hospital or die within 30 days of discharge. The generation starts from the last token indicating hospital discharge and continues forward in time. Both readmission and i...

  28. [36]

    Generation begins from the last token corresponding to hospital admission

    ICU Admission (Binary Classification) : This task predicts whether a patient will be admitted to the Intensive Care Unit (ICU) or die following a hospital admission. Generation begins from the last token corresponding to hospital admission. Both ICU admission and in-hospital d...

  29. [37]

    Count”) and the corresponding unique-token count (“N

    In-Hospital Mortality (Binary Classification) : The model predicts whether a patient will die during the hospital stay. Generation starts from the last token related to hospital admission. Only death is treated as a positive label, making this a more specific and challenging b...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.