Pith. sign in

REVIEW 3 major objections 8 minor 48 references

Marginal resemblance in synthetic longitudinal patient data can hide destroyed covariance and individual trajectories; temporal preservation must be assessed across four categories, not by a single score.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 00:59 UTC pith:QIABODVQ

load-bearing objection A practical metric battery for temporal fidelity in synthetic longitudinal data, with honest limitations but missing external validation and a schedule-robustness check. the 3 major comments →

arxiv 2602.10643 v1 pith:QIABODVQ submitted 2026-02-11 cs.LG

Evaluation metrics for temporal preservation in synthetic longitudinal patient data

Katariina Perkonoja , Parisa Movahedi , Antti Airola , Kari Auranen , Joni Virta This is my paper
classification cs.LG
keywords temporal preservationsynthetic longitudinal dataevaluation metricskernel smoothingcovariance structurevariogrammeasurement structureresemblance evaluation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that current evaluation of synthetic longitudinal patient data is too coarse: a single fidelity score, especially one measuring marginal distributions, can look excellent while the synthetic data has destroyed the very temporal dependencies that make longitudinal data useful. It therefore proposes a battery of ten metrics, organized into four categories—population-level marginal structure, covariance structure, individual trajectories, and measurement patterns—all built on kernel smoothing so that patients with different numbers and timing of measurements can be compared. Using synthetic versions of a real critical-care cohort, the authors show that a variable can match original means and quantiles while its variogram reveals near-zero between-subject stability, or that class proportions can be preserved while minority-class transition probabilities are unreliable. The practical conclusion is that temporal preservation must be assessed jointly across all four categories, not by any single number.

Core claim

On its own terms, the paper's central claim is that no single metric adequately captures temporal preservation in synthetic longitudinal patient data, and that strong marginal-level resemblance can conceal distortions in covariance and disruptions in individual-level trajectories. The supporting discovery is a set of ten kernel-smoothed metrics spanning four categories: mean/quantile/class profiles, variance profile and variogram plus rank-order variability and local transition probabilities, subject-level trajectory plots, and measurement density/similarity/dropout divergence. In the empirical illustrations, a variable's smoothed mean and quantile profiles matched the original data while th

What carries the argument

The load-bearing machinery is kernel smoothing on a common regular time grid. For each variable, each observed value is weighted by a kernel function of the distance from its measurement time to the grid point, with weights normalized so they sum to one; because the weights are computed separately for the original and synthetic datasets, both datasets' irregular schedules are projected onto the same grid and then compared pointwise. The main diagnostic tools are the kernel-smoothed variance profile and the variogram—a curve separating between-subject variability, measurement error, and serial correlation—plus rank-order variability for individual stability, transition probabilities for discr

Load-bearing premise

The argument assumes that kernel smoothing on a common time grid, using each dataset's own measurement times to define weights, yields comparable summaries of the underlying temporal process; if real and synthetic measurement schedules sample time differently, metric discrepancies can reflect scheduling artifacts rather than genuine loss of temporal structure. The paper's manual choice of bandwidth h=6 (Section 5), combined with the claim in Remark 1 that any discrepancy at a

What would settle it

Simulate a longitudinal process with a known covariance structure, generate a synthetic counterpart that exactly preserves the process but changes only the measurement schedule (e.g., original early-dense/late-sparse, synthetic uniform); if the proposed metrics report a large discrepancy, they are detecting schedule change rather than temporal failure. Alternatively, recompute the paper's empirical illustration across a range of bandwidths (h = 2, 4, 8, 12) and check whether the qualitative conclusions—such as weight preserving marginals but failing covariance—survive; if the classification fl

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Evaluators should treat a matching mean/quantile profile as necessary but not sufficient evidence of temporal fidelity; covariance and measurement-structure metrics are required to detect the failure modes shown.
  • Generative models trained on variables with sparse or irregular measurement times will show worse temporal preservation, so quality assessments should report such variables separately rather than averaging them into one score.
  • Preprocessing choices—bin width, outlier handling, imputation, time encoding—can be the real cause of temporal distortion, so evaluation feedback should be directed at the data pipeline as well as the model.
  • The proposed metrics give a way to localize failures in time and by subject subgroup, turning 'poor fidelity' into a specific diagnosis such as 'rare classes not learned' or 'outliers inflated.'
  • Reference bands computed from resampled original data provide a natural benchmark: synthetic-to-original numbers should be compared with that internal variation, not with an absolute threshold.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A consequence the authors do not draw: if single-score fidelity benchmarks are systematically blind to covariance and trajectory loss, previously published comparative evaluations of longitudinal generators may have overstated model quality; re-benchmarking with this battery is a natural next step.
  • The same machinery could be inverted into a training objective or early-stopping criterion—penalizing variogram, rank-order, or transition-probability mismatch during generation might push models to preserve between-subject and serial structure, not just cross-sectional distributions.
  • The paper's own limitations (Section 6.1—single generator per method, univariate metrics, manual bandwidth choice) mean its empirical demonstrations should be read as proof-of-concept rather than method ranking; the metrics themselves, not the specific findings, are the contribution.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. This paper introduces ten univariate metrics for evaluating temporal preservation in synthetic longitudinal patient data, organized into marginal structure (kernel-smoothed mean, quantile, and class profiles), covariance structure (variance profiles, variograms, rank-order variability, and local transition probabilities), individual trajectories, and measurement structure (measurement density, measurement similarity, dropout divergence). The authors review existing evaluation metrics, define each new metric formally on a common time grid with dataset-specific normalized kernel weights, and illustrate the toolkit on MIMIC-III data using HALO and Health Gym GAN synthetic datasets. Their central argument is that marginal-level resemblance can conceal covariance and individual-level failures, so no single metric suffices; the examples are intended to support this claim.

Significance. The paper addresses a genuine gap: most synthetic-data evaluations focus on cross-sectional fidelity and ignore temporal dependencies, especially in unbalanced longitudinal data. The proposed framework is practical, non-parametric, and accompanied by open-source R code and reproducible analysis scripts. The explicit reference-based calibration for measurement metrics (Metrics 9–10) is a good practice, and the authors are appropriately cautious about not comparing generators. If the design-bias confound is resolved and the metrics are validated against known degradations, the paper would be a useful methodological contribution to synthetic-data evaluation.

major comments (3)
  1. [§2.2, Remark 1; §5.2] The normalized weights in Definition 1 use each dataset's own measurement times. For this local-constant smoother, the asymptotic bias includes the design-density term μ_j'(t) f_j'(t)/f_j(t), so identical processes with different schedules yield different estimated profiles. The weight example in §5.2 therefore conflates genuine covariance loss with schedule/pair-lag differences: the synthetic data has more single-observation subjects, which changes the effective pair set in Metric 5, and the rank-order comparison in Fig. 8(c) inherits the same weights. Remark 1's claim that any discrepancy for any h is evidence of a difference is too strong. Please add (i) a bandwidth sensitivity analysis around h=6, (ii) an original-vs-original or schedule-matched calibration, and (iii) a known-degradation experiment showing the metrics detect real covariance distortions when schedules are controlled.
  2. [§4.2, Metric 6; Figs. 5, 8(c), 10(c), 13(b)] The metric R_ij is defined as the average first difference of quantile-bin indices, so it measures directional rank drift, not rank variability or rank-order stability. A subject oscillating between adjacent bins has expectation near zero and would be classified as stable despite high volatility. The text says values near zero indicate 'little systematic change,' but earlier frames the metric as measuring stability/volatility. None of the covariance-failure conclusions based on this metric are directly supported. Use an absolute difference or variance of rank changes for stability, or rename and reinterpret the current definition as drift.
  3. [§5, Figs. 3–17] The kernel-smoothed profile metrics are reported without uncertainty bands or original-vs-original calibration. For sparse variables such as AST (observed at 1.5% of time points), pointwise discrepancies may be sampling noise, and the manual bandwidth h=6 may affect which differences appear. The paper provides reference values only for Metrics 9–10. Add bootstrap pointwise bands or resampled-original reference envelopes for at least the key examples (weight, respiratory rate, AST) so that 'preserved' versus 'failed' judgments have a defined baseline.
minor comments (8)
  1. [§1] Typo: 'increasinly' should be 'increasingly'.
  2. [§2.2, Remark 1] Typo: 'bandwdith' should be 'bandwidth'.
  3. [Definition 2 / Metric 2] Subscript notation is inconsistent: F^(K)_jz(t) appears in Definition 2, while F^(K)_jq(t) appears in Metric 2. Clarify that q is the quantile level and avoid reusing q as an index.
  4. [Metric 7] Notation K_h(t, t−d_ijk(t)) is confusing; since it reduces to K_h(d_ijk(t)), write it directly as a function of the distance d.
  5. [§5] Grammar: 'The author's of HALO reported' should be 'The authors of HALO reported'.
  6. [§5.2] Typo: 'aspartate aminotrasferase' should be 'aspartate aminotransferase'.
  7. [Metric 8] The definition uses exact equality I_ijt on a continuous time grid, which may make the measurement density degenerate for general continuous-time data. State the intended discretization (e.g., rounding or a tolerance).
  8. [§5, Metrics 9–10] Report standard deviations or ranges across the 100 iterations, not only means, to support the reference comparisons.

Circularity Check

1 steps flagged

Self-contained metric proposal; only mild operational circularity in defining temporal preservation as agreement of the proposed summaries.

specific steps
  1. self definitional [Section 2.3, Definition 3 (and Discussion, Section 6)]
    "Applying M j to the original and synthetic data yields Mj = M((xijk, tijk)), M˜j = M((˜xijk, ˜tijk)). We say that the synthetic data exhibit univariate temporal preservation for variable j if ˜Mj ≈ Mj."

    Temporal preservation is defined as closeness of the chosen summary M_j (embodied in Metrics 1–10) between original and synthetic data. Consequently, the later conclusion that these metrics 'are designed for evaluating temporal preservation' is partly true by construction: there is no external, independently defined ground-truth construct of temporal preservation against which the metrics are validated. This is a mild operational tautology rather than a statistical reduction, and the paper's illustrative findings—e.g., HALO weight marginal agreement coexisting with variogram discrepancy—remain empirical and are not forced by the definitions.

full rationale

This paper is a metrics proposal, not a fitted-model derivation. The metrics in Section 4 are self-contained definitions applied symmetrically to original and synthetic data; no parameter is fitted to a subset and then used to 'predict' a closely related quantity, and no prediction is forced by construction. The central empirical examples use external MIMIC-III data and two external generators (HALO and Health Gym GAN), so the demonstrations that marginal resemblance can coexist with covariance and individual-level failures are data-driven. The only self-citation, reference [1], is used to locate previous metrics and methods, not to justify the proposed metrics or to forbid alternatives; it is not load-bearing. The main caveat is that the construct of 'temporal preservation' is operationalized by the proposed summaries in Definition 3, making the statement that these metrics measure temporal preservation partly definitional. The manually chosen bandwidth h=6 and the schedule-sensitivity concern are correctness/robustness issues, not circularity. Overall, the derivation chain is self-contained; no step reduces to its own input or to a self-citation chain, so the circularity is minor and non-load-bearing.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claims rest on methodological assumptions about comparability of kernel-smoothed summaries, alignment of measurement matrices, and the treatment of missingness. No new physical or latent entities are postulated; the new quantities are summary statistics. Several tuning choices (bandwidth, grid step, quantile set, epsilon, subsampling) are user decisions rather than quantities fitted to data.

free parameters (5)
  • Bandwidth h = 6
    Manually selected in Remark 1 and used for all experiments (Section 5); no sensitivity analysis is provided.
  • Time-grid step Δ_j = unit of time (1 hour after HALO rescaling)
    Section 2.2: the common grid T_j^* is defined with a fixed step Δ_j; the choice affects where kernel estimates are evaluated.
  • Quantile level set Q for Metric 6 = not specified
    Metric 6 depends on a user-chosen set Q of quantile levels; the exact levels used in figures are not stated.
  • KL smoothing constant ε (Metric 10) = not specified
    A small positive constant is added to empirical dropout probabilities before normalization; its value is not reported.
  • Subsample size and iterations for Metrics 9–10 = 2,000 subjects, 100 iterations
    Section 5: Hungarian-alignment metrics are averaged over random subsamples of 2,000 subjects and 100 iterations; the choice is computational, not data-driven.
axioms (5)
  • domain assumption Kernel-smoothed weighted summaries of original and synthetic data are directly comparable despite different measurement schedules.
    Section 2.2: normalized kernel weights depend on each dataset's own observation times, so estimates on the common grid may be biased differently when schedules differ.
  • domain assumption Hungarian assignment of measurement indicator rows gives a meaningful alignment between subjects that have no natural correspondence.
    Metric 9, Section 4.4: original and synthetic subjects are not linked; the Euclidean distance between indicator rows is used as the alignment cost for the Hungarian algorithm.
  • domain assumption Missingness and structural unbalancedness need not be distinguished for temporal-preservation evaluation.
    Section 2.1: the authors explicitly state they do not attempt to differentiate missingness from unbalancedness.
  • standard math Under stationarity, the variogram relative to total variance separates measurement error, serial correlation, and between-subject variability.
    Metric 5 and Figure 2 rely on standard longitudinal-data variogram theory (Diggle et al.; Diggle and Verbyla).
  • domain assumption Univariate evaluation is a sufficient first-order check; multivariate temporal structure adds limited additional insight.
    Section 6.1: the authors explicitly scope the work to univariate metrics and argue multivariate metrics are difficult to interpret.

pith-pipeline@v1.3.0-alltime-deepseek · 23241 in / 10847 out tokens · 100152 ms · 2026-08-03T00:59:38.919911+00:00 · methodology

0 comments
read the original abstract

This study introduces a set of metrics for evaluating temporal preservation in synthetic longitudinal patient data, defined as artificially generated data that mimic real patients' repeated measurements over time. The proposed metrics assess how synthetic data reproduces key temporal characteristics, categorized into marginal, covariance, individual-level and measurement structures. We show that strong marginal-level resemblance may conceal distortions in covariance and disruptions in individual-level trajectories. Temporal preservation is influenced by factors such as original data quality, measurement frequency, and preprocessing strategies, including binning, variable encoding and precision. Variables with sparse or highly irregular measurement times provide limited information for learning temporal dependencies, resulting in reduced resemblance between the synthetic and original data. No single metric adequately captures temporal preservation; instead, a multidimensional evaluation across all characteristics provides a more comprehensive assessment of synthetic data quality. Overall, the proposed metrics clarify how and why temporal structures are preserved or degraded, enabling more reliable evaluation and improvement of generative models and supporting the creation of temporally realistic synthetic longitudinal patient data.

Figures

Figures reproduced from arXiv: 2602.10643 by Antti Airola, Joni Virta, Kari Auranen, Katariina Perkonoja, Parisa Movahedi.

Figure 1
Figure 1. Figure 1: Evaluation metrics proposed in this work for assessing univariate temporal preservation [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic (theoretical) variograms with exponential and Gaussian correlation under [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Kernel-smoothed mean (Metric 1) and quantile profiles (Metric 2) of systolic blood [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Kernel-smoothed variance profile (Metric 4, blue) and variogram (Metric 5, orange) of [PITH_FULL_IMAGE:figures/full_fig_p023_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Boxplot (left) and density plot (right) showing the kernel-smoothed rank-order variability [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Subject-specific systolic blood pressure trajectories by baseline group for the original [PITH_FULL_IMAGE:figures/full_fig_p024_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Measurement density (Metric 8) of systolic blood pressure for the original data (blue) [PITH_FULL_IMAGE:figures/full_fig_p025_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Panel (a) shows the kernel-smoothed mean and quantile profiles, panel (b) the variance [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Subject-specific weight (kg) trajectories by baseline group for the original (top) and [PITH_FULL_IMAGE:figures/full_fig_p027_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Panel (a) shows the kernel-smoothed mean and quantile profiles, panel (b) the variance [PITH_FULL_IMAGE:figures/full_fig_p028_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Panel (a) shows subject-specific respiratory rate trajectories by baseline group for the [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Kernel-smoothed mean and quantile profiles of aspartate aminotransferase for the orig [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Panel (a) shows the variance profile and variogram (in squared units), and panel (b) the [PITH_FULL_IMAGE:figures/full_fig_p031_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Panel (a) shows subject-specific aspartate aminotrasferase trajectories by baseline group [PITH_FULL_IMAGE:figures/full_fig_p032_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Kernel-smoothed class profiles (Metric 3) of Glasgow Coma Scale score for the original [PITH_FULL_IMAGE:figures/full_fig_p033_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Panel (a) displays the kernel-smoothed transition probabilities (Metric 7) of the Glasgow [PITH_FULL_IMAGE:figures/full_fig_p034_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Kernel-smoothed class profiles of (a) the Glasgow Coma Scale Eye Opening assessment [PITH_FULL_IMAGE:figures/full_fig_p035_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 20 canonical work pages · 1 internal anchor

  1. [1]

    Perkonoja, K

    K. Perkonoja, K. Auranen, J. Virta, Methods for Generating and Evaluating Synthetic Longitudinal Patient Data: A Systematic Review, Journal of Health- care Informatics Research (2025).doi:10.1007/s41666-025-00223-7

    work page doi:10.1007/s41666-025-00223-7 2025

  2. [2]

    G. O. Ghosheh, J. Li, T. Zhu, A Survey of Generative Adversarial Networks for Synthesizing Structured Electronic Health Records, ACM Computing Surveys 56 (6) (2024) 1–34.doi:10.1145/3636424

    work page doi:10.1145/3636424 2024

  3. [3]

    Hernandez, G

    M. Hernandez, G. Epelde, A. Alberdi, R. Cilla, D. Rankin, Synthetic data generation for tabular health records: A systematic review, Neurocomputing 493 (2022) 28–45.doi:10.1016/j.neucom.2022.04.053

    doi:10.1016/j.neucom.2022.04.053 2022

  4. [4]

    A. D. Lautrup, T. Hyrup, A. Zimek, P. Schneider-Kamp, Systematic Review of Generative Modelling Tools and Utility Metrics for Fully Synthetic Tabular Data, ACM Computing Surveys (2024).doi:10.1145/3704437

    work page doi:10.1145/3704437 2024

  5. [5]

    Murtaza, M

    H. Murtaza, M. Ahmed, N. F. Khan, G. Murtaza, S. Zafar, A. Bano, Synthetic data generation: State of the art in health care domain, Computer Science Review 48 (2023) 100546.doi:10.1016/j.cosrev.2023.100546

    arXiv 2023

  6. [6]

    Ibrahim, Y

    M. Ibrahim, Y. A. Khalil, S. Amirrajab, C. Sun, M. Breeuwer, J. Pluim, B. Elen, G. Ertaylan, M. Dumontier, Generative AI for synthetic data across multiple medical modalities: A systematic review of recent developments and challenges, Computers in Biology and Medicine 189 (2025) 109834.doi: 10.1016/j.compbiomed.2025.109834

    arXiv 2025

  7. [7]

    Towards European Health Data Space (TEHDAS), Draft guideline on data min- imisation, pseudonymisation, anonymisation and synthetic data, Tech. rep. (9 2025)

    2025

  8. [8]

    A. M. Alaa, B. Van Breugel, E. Saveliev, M. Van Der Schaar, How Faithful is your Synthetic Data? Sample-level Metrics for Evaluating and Auditing Gen- erative Models, in: K. Chaudhuri, S. Jegelka, L. Song, C. Szepesvari, G. Niu, S. Sabato (Eds.), Proceedings of the 39th International Conference on Machine Learning, PMLR, Baltimore, Maryland, 2022, pp. 290–306

    2022

  9. [9]

    Bhanot, M

    K. Bhanot, M. Qi, J. S. Erickson, I. Guyon, K. P. Bennett, The Problem of Fairness in Synthetic Healthcare Data, Entropy 23 (9) (2021) 1165.doi:10. 3390/e23091165. 45

    2021

  10. [10]

    Pereira, M

    M. Pereira, M. Kshirsagar, S. Mukherjee, R. Dodhia, J. Lavista Ferres, R. de Sousa, Assessment of differentially private synthetic data for utility and fairness in end-to-end machine learning pipelines for tabular data, PLOS ONE 19 (2) (2024) e0297271.doi:10.1371/journal.pone.0297271

    doi:10.1371/journal.pone.0297271 2024

  11. [11]

    P. A. Osorio-Marulanda, G. Epelde, M. Hernandez, I. Isasa, N. M. Reyes, A. B. Iraola, Privacy Mechanisms and Evaluation Metrics for Synthetic Data Gen- eration: A Systematic Review, IEEE Access 12 (2024) 88048–88074.doi: 10.1109/ACCESS.2024.3417608

    arXiv 2024

  12. [13]

    Hernadez, G

    M. Hernadez, G. Epelde, A. Alberdi, R. Cilla, D. Rankin, Synthetic Tabular Data Evaluation in the Health Domain Covering Resemblance, Utility, and Pri- vacy Dimensions, Methods of Information in Medicine 62 (S 01) (2023) e19–e38. doi:10.1055/s-0042-1760247

    doi:10.1055/s-0042-1760247 2023

  13. [14]

    F. K. Dankar, M. K. Ibrahim, L. Ismail, A Multi-Dimensional Evaluation of Synthetic Data Generators, IEEE Access 10 (2022) 11147–11158.doi:10.1109/ ACCESS.2022.3144765

    arXiv 2022

  14. [15]

    C. Yan, Y. Yan, Z. Wan, Z. Zhang, L. Omberg, J. Guinney, S. D. Mooney, B. A. Malin, A Multifaceted benchmarking of synthetic electronic health record generation models, Nature Communications 13 (1) (2022) 7609.doi:10.1038/ s41467-022-35295-1

    2022

  15. [16]

    V. B. Vallevik, A. Babic, S. E. Marshall, S. Elvatun, H. M. Brøgger, S. Ala- garatnam, B. Edwin, N. R. Veeraragavan, A. K. Befring, J. F. Nygård, Can I trust my fake data – A comprehensive quality assessment framework for syn- thetic tabular data in healthcare, International Journal of Medical Informatics 185 (2024) 105413.doi:10.1016/j.ijmedinf.2024.105413

    arXiv 2024

  16. [17]

    Belgodere, P

    B. Belgodere, P. Dognin, A. Ivankay, I. Melnyk, Y. Mroueh, A. Mojsilovic, J. Navratil, A. Nitsure, I. Padhi, M. Rigotti, J. Ross, Y. Schiff, R. Vedpathak, R. A. Young, Auditing and Generating Synthetic Data with Controllable Trust Trade-offs, IEEE Journal on Emerging and Selected Topics in Circuits and Sys- tems (2024).doi:10.1109/JETCAS.2024.3477976. 46

    arXiv 2024

  17. [18]

    Johnson, T

    A. Johnson, T. Pollard, R. Mark, MIMIC-III clinical database (version 1.4), PhysioNet (2016).doi:10.13026/C2XW26

    doi:10.13026/c2xw26 2016

  18. [19]

    A. E. Johnson, T. J. Pollard, L. Shen, L. W. H. Lehman, M. Feng, M. Ghassemi, B. Moody, P. Szolovits, L. Anthony Celi, R. G. Mark, MIMIC-III, a freely accessible critical care database, Scientific Data 3 (2016).doi:10.1038/sdata. 2016.35

    doi:10.1038/sdata 2016

  19. [20]

    Theodorou, C

    B. Theodorou, C. Xiao, J. Sun, Synthesize high-dimensional longitudinal elec- tronic health records via hierarchical autoregressive language model, Nature Communications 14 (1) (2023) 5305.doi:10.1038/s41467-023-41093-0

    work page doi:10.1038/s41467-023-41093-0 2023

  20. [21]

    N. I.-H. Kuo, M. N. Polizzotto, S. Finfer, F. Garcia, A. Sönnerborg, M. Zazzi, M. Böhm, R. Kaiser, L. Jorm, S. Barbieri, The Health Gym: synthetic health- related datasets for the development of reinforcement learning algorithms, Sci- entific Data 9 (1) (2022) 693.doi:10.1038/s41597-022-01784-7

    work page doi:10.1038/s41597-022-01784-7 2022

  21. [22]

    M. P. Wand, M. C. Jones, Kernel Smoothing, 1995.doi:10.1007/ 978-1-4899-4493-1

    1995

  22. [23]

    Zhang, C

    Z. Zhang, C. Yan, T. A. Lasko, J. Sun, B. A. Malin, SynTEG: a framework for temporal structured electronic health data simulation, Journal of the American Medical Informatics Association 28 (3) (2021) 596–604.doi:10.1093/jamia/ ocaa262

    doi:10.1093/jamia/ 2021

  23. [24]

    J. Li, B. J. Cairns, J. Li, T. Zhu, Generating synthetic mixed-type longitudi- nal electronic health records for artificial intelligent applications, npj Digital Medicine 6 (1) (2023) 98.doi:10.1038/s41746-023-00834-7

    work page doi:10.1038/s41746-023-00834-7 2023

  24. [25]

    M. S. Haleem, A. Ekuban, A. Antonini, S. Pagliara, L. Pecchia, C. Allocca, Deep-Learning-Driven Techniques for Real-Time Multimodal Health and Phys- ical Data Synthesis, Electronics (Switzerland) 12 (9) (2023).doi:10.3390/ electronics12091989

    2023

  25. [26]

    J. L. Achterberg, M. R. Haas, M. R. Spruit, On the evaluation of synthetic lon- gitudinal electronic health records, BMC Medical Research Methodology 24 (1) (2024).doi:10.1186/s12874-024-02304-4

    work page doi:10.1186/s12874-024-02304-4 2024

  26. [27]

    Wang, T.-W

    W. Wang, T.-W. Pai, Enhancing Small Tabular Clinical Trial Dataset through Hybrid Data Augmentation: Combining SMOTE and WCGAN-GP, Data 8 (9) (2023) 135.doi:10.3390/data8090135. 47

    work page doi:10.3390/data8090135 2023

  27. [28]

    Nikolentzos, M

    G. Nikolentzos, M. Vazirgiannis, C. Xypolopoulos, M. Lingman, E. G. Brandt, Synthetic electronic health records generated with variational graph autoen- coders, npj Digital Medicine 6 (1) (2023).doi:10.1038/s41746-023-00822-x

    work page doi:10.1038/s41746-023-00822-x 2023

  28. [29]

    Isasa, M

    I. Isasa, M. Hernandez, G. Epelde, F. Londoño, A. Beristain, X. Larrea, A. Al- berdi, P. Bamidis, E. Konstantinidis, Comparative assessment of synthetic time series generation approaches in healthcare: leveraging patient metadata for ac- curate data synthesis, BMC Medical Informatics and Decision Making 24 (1) (2024).doi:10.1186/s12911-024-02427-0

    work page doi:10.1186/s12911-024-02427-0 2024

  29. [30]

    Mosquera, K

    L. Mosquera, K. El Emam, L. Ding, V. Sharma, X. H. Zhang, S. E. Kababji, C. Carvalho, B. Hamilton, D. Palfrey, L. Kong, B. Jiang, D. T. Eurich, A method for generating synthetic longitudinal health data, BMC Medical Re- search Methodology 23 (1) (2023).doi:10.1186/s12874-023-01869-w

    work page doi:10.1186/s12874-023-01869-w 2023

  30. [31]

    H. Sun, H. Lin, R. Yan, Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference, arXiv preprint (2023).doi:10.48550/ arXiv.2312.14646

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2312.14646 2023

  31. [32]

    Theodorou, S

    B. Theodorou, S. Jain, C. Xiao, J. Sun, ConSequence: Synthesizing Logically Constrained Sequences for Electronic Health Record Generation, Proceedings of the AAAI Conference on Artificial Intelligence 38 (14) (2024) 15355–15363. doi:10.1609/aaai.v38i14.29460

    work page doi:10.1609/aaai.v38i14.29460 2024

  32. [33]

    C. Pang, X. Jiang, N. P. Pavinkurve, K. S. Kalluri, E. L. Minto, J. Patterson, L. Zhang, G. Hripcsak, G. Gürsoy, N. Elhadad, K. Natarajan, CEHR-GPT: Generating Electronic Health Records with Chronological Patient Timelines, arXiv preprint (2024).doi:10.48550/arXiv.2402.04400

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.04400 2024

  33. [34]

    Z. Wang, J. Sun, PromptEHR: Conditional Electronic Healthcare Records Generation with Prompt Learning, in: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Association for Com- putational Linguistics, Stroudsburg, PA, USA, 2022, pp. 2873–2885.doi: 10.18653/v1/2022.emnlp-main.185

    work page doi:10.18653/v1/2022.emnlp-main.185 2022

  34. [35]

    A. S. Hashemi, K. Etminani, A. Soliman, O. Hamed, J. Lundström, Time-series Anonymization of Tabular Health Data using Generative Adversarial Network, in: 2023InternationalJointConferenceonNeuralNetworks(IJCNN),Vol.2023- June, IEEE, 2023, pp. 1–8.doi:10.1109/IJCNN54540.2023.10191367. 48

    arXiv 2023

  35. [36]

    N. I.-H. Kuo, F. Garcia, A. Sönnerborg, M. Böhm, R. Kaiser, M. Zazzi, M. Poliz- zotto, L. Jorm, S. Barbieri, Generating synthetic clinical data that capture class imbalanced distributions with generative adversarial networks: Example using antiretroviral therapy for HIV, Journal of Biomedical Informatics 144 (2023) 104436.doi:10.1016/j.jbi.2023.104436

    arXiv 2023

  36. [37]

    Diggle, P

    P. Diggle, P. Heagerty, K.-Y. Liang, S. Zeger, Exploring longitudinal data, in: Analysis of Longitudinal Data, 2nd Edition, Oxford University Press, 2002, Ch. 3, pp. 39–52.doi:10.1093/oso/9780198524847.001.0001

    arXiv 2002

  37. [38]

    P. J. Diggle, A. P. Verbyla, Nonparametric Estimation of Covariance Structure in Longitudinal Data, Biometrics 54 (2) (1998) 401.doi:10.2307/3109751

    work page doi:10.2307/3109751 1998

  38. [39]

    Gringarten, C

    E. Gringarten, C. V. Deutsch, Teacher’s Aide Variogram Interpretation and Modeling, Mathematical Geology 33 (4) (2001) 507–534.doi:10.1023/A: 1011093014141

    doi:10.1023/a: 2001

  39. [40]

    H. W. Kuhn, The Hungarian method for the assignment problem, Naval Re- search Logistics Quarterly 2 (1-2) (1955).doi:10.1002/nav.3800020109

    doi:10.1002/nav.3800020109 1955

  40. [41]

    Tomizawa, On some techniques useful for solution of transportation network problems, Networks 1 (2) (1971).doi:10.1002/net.3230010206

    N. Tomizawa, On some techniques useful for solution of transportation network problems, Networks 1 (2) (1971).doi:10.1002/net.3230010206

    work page doi:10.1002/net.3230010206 1971

  41. [42]

    Edmonds, R

    J. Edmonds, R. M. Karp, Theoretical Improvements in Algorithmic Efficiency for Network Flow Problems, Journal of the ACM (JACM) 19 (2) (1972).doi: 10.1145/321694.321699

    arXiv 1972

  42. [43]

    Biswal, S

    S. Biswal, S. Ghosh, J. Duke, B. Malin, W. Stewart, J. Sun, EVA: Generating Longitudinal Electronic Health Records Using Conditional Variational Autoen- coders, in: K. Jung, S. Yeung, M. Sendak, M. Sjoding, R. Ranganath (Eds.), Proceedings of the 6th Machine Learning for Healthcare Conference, PMLR, 2021, pp. 260–282

    2021

  43. [44]

    Gottesman, J

    O. Gottesman, J. Futoma, Y. Liu, S. Parbhoo, L. A. Celi, E. Brunskill, F. Doshi- Velez, Interpretable Off-Policy Evaluation in Reinforcement Learning by High- lighting Influential Transitions, in: H. D. III, A. Singh (Eds.), Proceedings of the 37th International Conference on Machine Learning, PMLR, 2020, pp. 3658– 3667.doi:https://dl.acm.org/doi/10.5555/...

    arXiv 2020

  44. [45]

    Aristodimou, A

    A. Aristodimou, A. Diavastos, C. S. Pattichis, A fast supervised density-based discretization algorithm for classification tasks in the medical domain, Health Informatics Journal 28 (1) (2022).doi:10.1177/14604582211065397. 49

    work page doi:10.1177/14604582211065397 2022

  45. [46]

    Y. Wu, W. Wu, D. Yang, C. Xu, Z. Li, Neural response generation with dynamic vocabularies, in: 32nd AAAI Conference on Artificial Intelligence, AAAI 2018, 2018.doi:10.1609/aaai.v32i1.11943

    work page doi:10.1609/aaai.v32i1.11943 2018

  46. [47]

    M. H. Huque, J. B. Carlin, J. A. Simpson, K. J. Lee, A comparison of multiple imputation methods for missing data in longitudinal studies, BMC Medical Research Methodology 18 (1) (2018) 168.doi:10.1186/s12874-018-0615-6

    work page doi:10.1186/s12874-018-0615-6 2018

  47. [48]

    R. J. Little, A test of missing completely at random for multivariate data with missing values, Journal of the American Statistical Association 83 (404) (1988). doi:10.1080/01621459.1988.10478722

    arXiv 1988

  48. [49]

    Li, Little’s test of missing completely at random, Stata Journal 13 (4) (2013)

    C. Li, Little’s test of missing completely at random, Stata Journal 13 (4) (2013). doi:10.1177/1536867x1301300407. 50

    work page doi:10.1177/1536867x1301300407 2013