pith. sign in

arxiv: 2607.00127 · v1 · pith:RYQJPTJQnew · submitted 2026-06-30 · 💻 cs.LG

A Filtered Mixture-of-Generators for Fully Synthetic Survival Training

Niccol\`o Maria Rizzi , Eugenio Lomurno , Alberto Archetti , Matteo Matteucci This is my paper

Pith reviewed 2026-07-02 19:53 UTC · model grok-4.3

classification 💻 cs.LG
keywords survival analysissynthetic datagenerative modelssample filteringclinical dataC-indexintegrated Brier scoreprivacy
0
0 comments X

The pith

Filtering samples from a mixture of tabular generators using scores from a real-data survival ensemble produces synthetic training sets that match or exceed real clinical data for downstream survival models.

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

The paper introduces FoGS to address scarce and privacy-restricted clinical data for survival analysis by reframing synthetic data construction as sample selection instead of direct generation. Candidates are drawn from four distinct tabular generators and scored for plausibility by an ensemble of seven survival models trained on real data, using proper scoring rules. A two-level pipeline then optimizes generator quotas, scorer weights, random complement, and event-time balancing in an outer loop against held-out performance of an inner-loop XGBoost-Cox model. On 16 public datasets evaluated train-on-synthetic test-on-real, the filtered output improves mean C-index by 2.17 and IBS by 0.67, beating unfiltered synthetic data on 9 datasets for both metrics and at least one metric on 13, while matching or exceeding real-data training on most cohorts with unchanged nearest-neighbour privacy margins.

Core claim

FoGS reframes synthetic-data construction as sample selection rather than generation. A candidate pool is drawn from four architecturally distinct tabular generators, and each sample is scored by an ensemble of seven survival models trained on real data, using proper scoring rules as a per-sample plausibility proxy. A two-level pipeline optimizes, in its outer loop, a selection policy -- generator quotas, scorer weights, a random complement, and stratified balancing on event time and censoring -- against held-out downstream performance, while an inner loop tunes the downstream model (XGBoost-Cox). On 16 public datasets under train-on-synthetic, test-on-real (C-index and IBS, 0--100 scale), F

What carries the argument

The two-level optimization pipeline that selects and balances samples from a heterogeneous generator pool using per-sample plausibility scores supplied by a real-data survival-model ensemble.

Load-bearing premise

An ensemble of survival models trained on the real data supplies an unbiased per-sample plausibility proxy that generalizes to held-out test distributions without introducing selection bias that favors the downstream model architecture.

What would settle it

A new clinical cohort where models trained on the filtered synthetic data show statistically lower C-index or higher IBS than models trained on the real data, or where nearest-neighbour privacy distance drops below the unfiltered baseline.

Figures

Figures reproduced from arXiv: 2607.00127 by Alberto Archetti, Eugenio Lomurno, Matteo Matteucci, Niccol\`o Maria Rizzi.

Figure 1
Figure 1. Figure 1: Overview of the FoGS pipeline. Stage 1 (left) draws a candidate pool from the four generators and scores every sample with the survival-model ensemble trained on the real cohort. Stage 2 (right) filters the pool under the selection policy and trains the downstream model, with the two-level optimization driven by feedback on the real validation split. Dr is held out from the validation and test splits, used… view at source ↗
Figure 2
Figure 2. Figure 2: Per-dataset change of FoGS over the real-data base￾line, ∆C (horizontal) against ∆IBS (vertical), one point per dataset. The shaded upper-right quadrant contains the datasets that improve on both metrics; breast_cancer lies on the ∆C = 0 axis. 4.3 Structure of the Selection Policy The best-trial policies expose a consistent structure ( [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Best-trial selection policy for each of the 16 datasets. Top row: the four generator quotas pg. Bottom row: the random complement ρ and the two stratification flags balt and balc. Binary panels are offset at zero for visibility. side for every dataset. The second ablation isolates the se￾lection policy: a random draw of equal size from the same pool falls below the real-data baseline (∆C = −0.50), whereas … view at source ↗
Figure 4
Figure 4. Figure 4: Best-trial scorer-loss weights for each dataset over the 14 combinations (columns). Colour encodes the weight; each row sums to one. 11 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: UMAP projections of the real and synthetic cohorts (part 1 of 2): datasets gbsg, metabric, whas, aids, breast_cancer, cost, d_oropha_rec, and hepatocellular (rows). Columns, left to right: the real training data, the FoGS-selected set D ∗ s , and the over-generated candidate pools Dg of ARF, Bayesian Network, TabDDPM, and SurvivalCTGAN. For each dataset a single UMAP embedding is fitted on the real event-b… view at source ↗
Figure 6
Figure 6. Figure 6: UMAP projections of the real and synthetic cohorts (part 2 of 2): datasets melanoma, mgus, nki70, pbc, pbc3, prostate, stagec, and uis (rows). Columns, embedding, and colour scale are as in [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
read the original abstract

Survival analysis models time-to-event data, but in clinical settings training data are costly and scarce: events accrue over years of follow-up, cohorts are small, and privacy regulations restrict sharing across institutions. Tabular generative models promise augmentation and privacy-preserving cohort sharing, yet are themselves data-hungry -- on the small cohorts typical of survival analysis, a single generator rarely characterizes the population well enough for downstream models trained on its output to match real-data performance. FoGS (Filtered Mixture-of-Generators for Survival analysis) reframes synthetic-data construction as sample selection rather than generation. A candidate pool is drawn from four architecturally distinct tabular generators, and each sample is scored by an ensemble of seven survival models trained on real data, using proper scoring rules as a per-sample plausibility proxy. A two-level pipeline optimizes, in its outer loop, a selection policy -- generator quotas, scorer weights, a random complement, and stratified balancing on event time and censoring -- against held-out downstream performance, while an inner loop tunes the downstream model (XGBoost-Cox). On 16 public datasets under train-on-synthetic, test-on-real (C-index and IBS, $0$--$100$ scale), FoGS yields mean improvements of $+2.17$ in C-index and $+0.67$ in IBS, improving both metrics on 9 of 16 datasets and at least one on 13 (one-sided Wilcoxon $p=0.039$ and $p=0.035$). It matches or exceeds real-data training on most cohorts, with no significant change in nearest-neighbour privacy margin relative to unfiltered sampling. Sample filtering over a heterogeneous generator pool is thus a viable substitute for real-data training in privacy-restricted clinical settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents FoGS, which constructs synthetic survival data by sampling from four tabular generators, scoring candidates with an ensemble of seven survival models via proper scoring rules, and optimizing (outer loop) generator quotas, scorer weights, random complement, and stratified balancing on event/censoring to maximize held-out XGBoost-Cox performance, while an inner loop tunes the downstream model. On 16 public datasets under train-on-synthetic/test-on-real (C-index and IBS scaled 0-100), it reports mean gains of +2.17 C-index and +0.67 IBS, with both metrics improved on 9/16 datasets and at least one on 13/16 (one-sided Wilcoxon p=0.039, 0.035), matching or exceeding real-data training on most cohorts and no significant privacy degradation.

Significance. If the outer-loop optimization can be shown to produce a selection policy that is not overfit to XGBoost-Cox biases, the multi-generator filtering approach would offer a practical route to privacy-preserving synthetic cohorts that rival real-data performance in survival tasks. The use of proper scoring rules for per-sample plausibility and the scale of the empirical evaluation across 16 datasets are strengths that would support broader adoption in clinical settings if the central methodological concern is addressed.

major comments (2)
  1. [§4 (outer-loop optimization)] §4 (outer-loop optimization): The selection policy is tuned explicitly against held-out XGBoost-Cox performance. This creates a risk that the reported gains (+2.17 C-index, +0.67 IBS) reflect selection tuned to that model's inductive biases rather than an architecture-agnostic plausibility filter. The manuscript must specify the exact nested CV structure ensuring optimization folds are disjoint from final test folds and should report results with at least one additional downstream model (e.g., CoxPH or RSF) to test generalizability.
  2. [Results section (Table 2 or equivalent)] Results section (Table 2 or equivalent): The headline Wilcoxon tests and 'matches or exceeds real-data training' claim are computed after policy optimization for XGBoost-Cox. Without an ablation showing that the ensemble-derived scores alone (without outer-loop tuning) produce comparable gains, the central claim that FoGS is a 'viable substitute' for real data rests on a potentially circular selection process.
minor comments (2)
  1. [Abstract] Abstract: C-index and IBS are reported on a 0-100 scale; standard definitions use 0-1, so the scaling factor should be stated explicitly for readers.
  2. [§3.1 (generators and ensemble)] §3.1 (generators and ensemble): The choice of the four generators and seven scorers, along with any hyperparameter settings, should be listed in a table for full reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important considerations for the robustness of our evaluation. We address each major point below and commit to revisions that strengthen the manuscript.

read point-by-point responses

  1. Referee: §4 (outer-loop optimization): The selection policy is tuned explicitly against held-out XGBoost-Cox performance. This creates a risk that the reported gains (+2.17 C-index, +0.67 IBS) reflect selection tuned to that model's inductive biases rather than an architecture-agnostic plausibility filter. The manuscript must specify the exact nested CV structure ensuring optimization folds are disjoint from final test folds and should report results with at least one additional downstream model (e.g., CoxPH or RSF) to test generalizability.

    Authors: We agree that explicit description of the nested CV is necessary to confirm disjoint folds. The outer optimization loop operates on folds separate from the final test set, with the inner loop performing downstream model tuning only within the optimization folds; we will add a precise description of this structure (including fold counts and disjointness) to §4. To address generalizability beyond XGBoost-Cox, we will also report results using CoxPH as an additional downstream model in the revised results section. revision: yes

  2. Referee: Results section (Table 2 or equivalent): The headline Wilcoxon tests and 'matches or exceeds real-data training' claim are computed after policy optimization for XGBoost-Cox. Without an ablation showing that the ensemble-derived scores alone (without outer-loop tuning) produce comparable gains, the central claim that FoGS is a 'viable substitute' for real data rests on a potentially circular selection process.

    Authors: The outer loop is designed to optimize the selection policy parameters using the ensemble scores as the core plausibility signal. We recognize the benefit of an ablation to isolate components. In the revision we will add results comparing (i) unfiltered baselines, (ii) ensemble-score filtering with fixed (non-optimized) policy, and (iii) the full optimized FoGS, allowing readers to assess the contribution of the outer-loop tuning separately from the ensemble filtering itself. revision: yes

Circularity Check

0 steps flagged

No reduction of reported metrics to fitted selection parameters by construction

full rationale

The central claims consist of empirical C-index and IBS improvements measured on held-out real test data after training downstream models on the filtered synthetic samples. The outer-loop optimization tunes generator quotas, scorer weights and balancing against held-out performance, yet the published test-set numbers are not algebraically or statistically defined in terms of those fitted parameters; they remain independent external benchmarks. No self-definitional equations, fitted-input-called-prediction steps, or load-bearing self-citations appear in the abstract or described pipeline. This places the work in the normal non-circular range (0-2).

Axiom & Free-Parameter Ledger

3 free parameters · 1 axioms · 0 invented entities

The method rests on the empirical claim that proper scoring rules from real-data survival models can rank synthetic samples usefully; no new mathematical axioms or invented physical entities are introduced. Free parameters include generator quotas, scorer weights, random complement fraction, and stratification bins for event time and censoring.

free parameters (3)
  • generator quotas
    Proportion of samples kept from each of the four generators; chosen by outer-loop optimization against downstream performance.
  • scorer weights
    Weights on the seven survival models in the ensemble; tuned in the outer loop.
  • stratification bins
    Number and boundaries of bins for event time and censoring status used in balancing.
axioms (1)
  • domain assumption Proper scoring rules applied to survival models trained on real data yield a monotonic proxy for sample plausibility that transfers to held-out distributions.
    Invoked when the ensemble is used to filter the candidate pool before downstream training.

pith-pipeline@v0.9.1-grok · 5858 in / 1618 out tokens · 21197 ms · 2026-07-02T19:53:20.600908+00:00 · methodology

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

27 extracted references · 27 canonical work pages · 1 internal anchor

  1. [1]

    Optuna: A next-generation hyperparameter optimization frame- work

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization frame- work. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623–2631, 2019

    work page 2019

  2. [2]

    Robust variational autoencoder for tabular data with beta divergence.arXiv preprint arXiv:2006.08204, 2020

    Haleh Akrami, Sergul Aydore, Richard M Leahy, and Anand A Joshi. Robust variational autoencoder for tabular data with beta divergence.arXiv preprint arXiv:2006.08204, 2020

    work page arXiv 2006

  3. [3]

    pgmpy: Apython toolkit for bayesian networks.Journal of Machine Learning Research, 25(265):1–8, 2024

    AnkurAnkanandJohannesTextor. pgmpy: Apython toolkit for bayesian networks.Journal of Machine Learning Research, 25(265):1–8, 2024

    work page 2024

  4. [4]

    Fpboost: Fully parametric gradient boosting for survival analysis.arXiv preprint arXiv:2409.13363, 2024

    Alberto Archetti, Eugenio Lomurno, Diego Piccinotti, and Matteo Matteucci. Fpboost: Fully parametric gradient boosting for survival analysis.arXiv preprint arXiv:2409.13363, 2024

    work page arXiv 2024

  5. [5]

    Xgboost: A scal- able tree boosting system

    Tianqi Chen and Carlos Guestrin. Xgboost: A scal- able tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowl- edge discovery and data mining, pages 785–794, 2016

    work page 2016

  6. [6]

    D. R. Cox. Regression models and life-tables.Jour- nal of the Royal Statistical Society. Series B (Method- ological), 34(2):187–220, 1972. ISSN 00359246

    work page 1972

  7. [7]

    Neural spline flows.Ad- vances in neural information processing systems, 32, 2019

    Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios. Neural spline flows.Ad- vances in neural information processing systems, 32, 2019

    work page 2019

  8. [8]

    Verification of forecasts ex- pressed in terms of probability.Monthly weather review, 78(1):1–3, 1950

    W Brier Glenn et al. Verification of forecasts ex- pressed in terms of probability.Monthly weather review, 78(1):1–3, 1950

    work page 1950

  9. [9]

    Strictly properscoringrules,prediction,andestimation.Jour- nal of the American statistical Association, 102(477): 359–378, 2007

    Tilmann Gneiting and Adrian E Raftery. Strictly properscoringrules,prediction,andestimation.Jour- nal of the American statistical Association, 102(477): 359–378, 2007

    work page 2007

  10. [10]

    Multivariable prognostic models: issues in develop- ing models, evaluating assumptions and adequacy, and measuring and reducing errors.Statistics in medicine, 15(4):361–387, 1996

    Frank E Harrell Jr, Kerry L Lee, and Daniel B Mark. Multivariable prognostic models: issues in develop- ing models, evaluating assumptions and adequacy, and measuring and reducing errors.Statistics in medicine, 15(4):361–387, 1996

    work page 1996

  11. [11]

    Kogalur, Eugene H

    Hemant Ishwaran, Udaya B. Kogalur, Eugene H. Blackstone, and Michael S. Lauer. Random survival forests.The Annals of Applied Statistics, 2(3):841 – 860, 2008. doi:10.1214/08-AOAS169

    work page doi:10.1214/08-aoas169 2008

  12. [12]

    Nonparametric estimation from incomplete observations.Journal of the American statistical association, 53(282):457– 481, 1958

    Edward L Kaplan and Paul Meier. Nonparametric estimation from incomplete observations.Journal of the American statistical association, 53(282):457– 481, 1958

    work page 1958

  13. [13]

    Deepsurv: personalized treatment recommender systemusingacoxproportionalhazardsdeepneural network.BMC medical research methodology, 18: 1–12, 2018

    Jared L Katzman, Uri Shaham, Alexander Cloninger, Jonathan Bates, Tingting Jiang, and Yuval Kluger. Deepsurv: personalized treatment recommender systemusingacoxproportionalhazardsdeepneural network.BMC medical research methodology, 18: 1–12, 2018

    work page 2018

  14. [14]

    Auto- encoding variational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto- encoding variational bayes, 2013

    work page 2013

  15. [15]

    Springer, 2003

    John P Klein, Melvin L Moeschberger, et al.Sur- vivalanalysis: techniquesforcensoredandtruncated data, volume 1230. Springer, 2003

    work page 2003

  16. [16]

    Tabddpm: Modelling tabular data with diffusion models

    Akim Kotelnikov, Dmitry Baranchuk, Ivan Rubachev, and Artem Babenko. Tabddpm: Modelling tabular data with diffusion models. InInternational Confer- ence on Machine Learning, pages 17564–17579. PMLR, 2023

    work page 2023

  17. [17]

    Deephit: A deep learning approach to survival analysis with competing risks

    Changhee Lee, William Zame, Jinsung Yoon, and Mihaela Van Der Schaar. Deephit: A deep learning approach to survival analysis with competing risks. InProceedings of the AAAI conference on artificial intelligence, volume 32, page 1, 2018

    work page 2018

  18. [18]

    Inference-Time Refinement Closes the Synthetic-Real Gap in Tabular Diffusion

    Eugenio Lomurno, Filippo Balzarini, Francesco Benelle, Francesca Pia Panaccione, and Matteo Matteucci. Inference-time refinement closes the synthetic-real gap in tabular diffusion.arXiv preprint arXiv:2605.06261, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    Survivalgan: Generating time-to-event data for survival analysis

    Alexander Norcliffe, Bogdan Cebere, Fergus Imrie, Pietro Lio, and Mihaela van der Schaar. Survivalgan: Generating time-to-event data for survival analysis. InInternational Conference on Artificial Intelligence and Statistics, pages 10279–10304. PMLR, 2023

    work page 2023

  20. [20]

    Synthcity: a benchmark framework for diverse use cases of tabular synthetic data.Ad- vances in neural information processing systems, 36: 3173–3188, 2023

    Zhaozhi Qian, Rob Davis, and Mihaela Van Der Schaar. Synthcity: a benchmark framework for diverse use cases of tabular synthetic data.Ad- vances in neural information processing systems, 36: 3173–3188, 2023

    work page 2023

  21. [21]

    Your image gen- erator is your new private dataset.Image and Vision Computing, page 105727, 2025

    Nicolò Francesco Resmini, Eugenio Lomurno, Cris- tian Sbrolli, and Matteo Matteucci. Your image gen- erator is your new private dataset.Image and Vision Computing, page 105727, 2025

    work page 2025

  22. [22]

    Deep unsuper- vised learning using nonequilibrium thermodynam- ics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Mah- eswaranathan, and Surya Ganguli. Deep unsuper- vised learning using nonequilibrium thermodynam- ics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015. 8 A Filtered Mixture-of-Generators for Fully Synthetic Survival Training

    work page 2015

  23. [23]

    The max-minhill-climbing bayesian network structure learning algorithm.Machine learn- ing, 65:31–78, 2006

    Ioannis Tsamardinos, Laura E Brown, and Con- stantinF Aliferis. The max-minhill-climbing bayesian network structure learning algorithm.Machine learn- ing, 65:31–78, 2006

    work page 2006

  24. [24]

    Ma- chine learning for survival analysis: A survey.ACM Computing Surveys (CSUR), 51(6):1–36, 2019

    Ping Wang, Yan Li, and Chandan K Reddy. Ma- chine learning for survival analysis: A survey.ACM Computing Surveys (CSUR), 51(6):1–36, 2019

    work page 2019

  25. [25]

    Adversarial random forests for density estimation and generative modeling

    David S Watson, Kristin Blesch, Jan Kapar, and Mar- vin N Wright. Adversarial random forests for density estimation and generative modeling. InInternational Conference on Artificial Intelligence and Statistics, pages 5357–5375. PMLR, 2023

    work page 2023

  26. [26]

    Modeling tabular data using conditional gan.Advances in neural informa- tion processing systems, 32, 2019

    Lei Xu, Maria Skoularidou, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Modeling tabular data using conditional gan.Advances in neural informa- tion processing systems, 32, 2019

    work page 2019

  27. [27]

    Anonymization through data synthe- sis using generative adversarial networks (ads-gan)

    Jinsung Yoon, Lydia N Drumright, and Mihaela Van Der Schaar. Anonymization through data synthe- sis using generative adversarial networks (ads-gan). IEEE journal of biomedical and health informatics, 24(8):2378–2388, 2020. 9 A Filtered Mixture-of-Generators for Fully Synthetic Survival Training Table 4:Outer-Optuna search space (21 parameters). The 14 sco...

    work page arXiv 2020