cs.LGAug 23, 2026

KMGen: A Skill-based Approach for Synthetic Individual Patient Data Generation

Authors: Jalen Jiang, Chufan Gao, Ethan Rasmussen, Stephen Z. Xie, Jimeng Sun

Organizations: Siebel School of Computing and Data Science University of Illinois Urbana-Champaign · Mayo Clinic Alix School of Medicine

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Individual patient data (IPD) from clinical trials is the substrate for survival modeling, meta-analysis, and safety research, yet IPD is rarely released. Prior work has addressed only half of this gap: reconstructing Kaplan-Meier (KM) curves from published plots -- typically requiring manual digitization or human-in-the-loop correction -- while offering no mechanism for generating the adverse-event (AE) streams that constitute the other half of a patient record. We introduce KMGen, the first end-to-end framework that (i) fully automates KM curve extraction at accuracy competitive with human-guided tools, and (ii) generates synthetic per-patient AE trajectories from public trial registry records. The extraction stage is a fully automated agentic pipeline -- an agent generates code to extract each step in the KM curve -- achieving a mean Integrated Absolute Error (IAE) of 0.0151 on a 32-plot benchmark spanning clean, edge-case, and adversarial conditions. The IPD generation stage decouples patient archetype extraction from statistical sampling: an LLM distills the trial record into arm-specific statistics, adverse events, patient demographics, and risk multipliers. A mechanistic sampler generates patient events via clinical archetypes, bootstrap rank-correlation coupling to the empirical KM curve (preserving the marginal survival distribution exactly), and cycle-based AE scheduling with an induction/maintenance split. Across three held-out oncology trials spanning an order of magnitude in cohort size and 30 independent regenerations per trial, KMGen achieves mean integrated KM absolute difference ΔKM ≤ 0.051Δ_{\text{KM}}\,{\leq}\,0.051, sex/ECOG JSD ≤ 0.013{\leq}\,0.013 on 5 of 6 demographic slots, and recovers ≥ 71%{\geq}\,71\% of the top-15 AEs by exact MedDRA term under a single fixed parameter set. The pipeline is released as open source at https://github.com/chufangao/kmgen.

Figures & tables

Figure 1: KMGen pipeline. (a) Fully automated KM extraction (b) Individual Patient Data generation from extracted KM plot, ClinicalTrial.gov data, publication, and statistical analysis plan.
CategorynnMean IAEMedian IAEMinMaxStd
Synthetic benchmark (ground truth: rendered step function)
Standard50.00510.00510.00330.00680.0015
Edge70.00890.00700.00120.02370.0080
Single stress120.01400.00870.00510.05400.0136
Combo stress80.02830.02340.00120.07440.0231
Synthetic (all)320.01510.00800.00120.07440.0164
Table 1: Extraction accuracy by plot category. The synthetic benchmark (top block, 32 plots) measures IAE against programmatically-rendered ground truth. The real-trial block compares the extracted published KM curve against the empirical KM curve reconstructed from sponsor-delivered IPD for three oncology trials (one arm per trial). IAE is the Integrated Absolute Error in [0,1][0,1] ; lower is better.
Figure 2: Real (solid) vs. synthetic (dashed) KM curves. Mean ΔKM≤0.051\Delta_{\text{KM}}\leq 0.051 on all trials; Mann–Whitney UU fails to reject the null on both smaller trials (mean p≥0.6p\geq 0.6 ) but rejects on NCT00844649 in 29 of 30 seeds, driven by a 1.3-month median shift at nr=430n_{r}{=}430 .
 NCT03041311NCT02499770NCT00844649
 nr=53, ns=107n_{r}{=}53,\,n_{s}{=}107nr=37, ns=75n_{r}{=}37,\,n_{s}{=}75nr=430, ns=861n_{r}{=}430,\,n_{s}{=}861
Demographics (Stage B, mean ±\pm std over 30 regens)
Sex JSD ↓\downarrow0.0054 ±\pm 0.00640.0033 ±\pm 0.00600.0005 ±\pm 0.0006
ECOG JSD ↓\downarrow0.272 ±\pm 0.0480.013 ±\pm 0.0120.003 ±\pm 0.001
Weight KS ↓\downarrow0.201 ±\pm 0.0290.202 ±\pm 0.0500.129 ±\pm 0.013
Height KS ↓\downarrow0.217 ±\pm 0.0430.166 ±\pm 0.0410.102 ±\pm 0.014
Table 2: End-to-end fidelity and extraction quality. Synthetic cohorts at 2×2{\times} real enrollment. ↓\downarrow = lower is better, ↑\uparrow = higher is better. JSD ∈[0,log⁡2]\in[0,\log 2] , KS ∈[0,1]\in[0,1] . Stage B values are reported as mean ±\pm std across 30 independent regenerations of the synthetic cohort (seeds 0,…,290,\dots,29 ); Stage A metrics are deterministic functions of the extracted configs and have no resampling variability.
Figure 3: Top-12 shared AE terms per trial, ranked by real-IPD prevalence. Top-15 overlap is 7171 – 78%78\% on the small trials and 80%80\% on NCT00844649 (means over 30 regenerations).
Figure 4: AE onset density. Synthetic onsets show cycle-length striping (a direct consequence of cycle-based sampling); real onsets are smoother due to continuous-time recording. Mean onset error is ≤0.71{\leq}0.71 months.
Figure 5: Top-10 SOC distributions. The dominant real SOC is the top synthetic SOC in every trial, confirming correct organ-system targeting by the archetype risk multipliers.
Appendix figures & tables30 assets

Supplementary material from the paper’s appendix.

Appendix

FactorKey SOCs affectedDefaultEvidence
Age ≥\geq 65 (“older”)
 Blood & lymphatic1.521 ; 14
 Cardiac1.825
 Infections1.617
 Nervous system1.31
 Renal & urinary1.418
Appendix
Table 3: Default AE risk multipliers by patient factor and organ system. Per-SOC values are clinical estimates informed by the overall effect sizes in the cited studies (Appendix B ). All multipliers are relative to a baseline of 1.0.
Figure 6: Induction fraction sweep. The onset KS statistic is minimized at α=0.85\alpha{=}0.85 (KS  = 0.151\,{=}\,0.151 ), with sharp degradation at low α\alpha , validating the induction–maintenance phase split.
Figure 7: Onset KS as a function of induction fraction α\alpha , showing a clear minimum near α=0.85\alpha{=}0.85 .
Figure 8: Patient propensity sweep. Burden KS decreases monotonically, confirming the lognormal frailty is necessary. The default ση=0.4\sigma_{\eta}{=}0.4 is conservative; higher values further reduce the burden gap.
Figure 9: Burden KS as a function of patient propensity ση\sigma_{\eta} . KS decreases monotonically; the default ση=0.4\sigma_{\eta}{=}0.4 is conservative — higher values further reduce the gap.
Figure 10: Archetype ablation. Removing risk multipliers preserves marginal survival but degrades AE timing and burden realism.
Figure 11: Metric sensitivity to synthetic cohort size. AE frequency metrics improve with sample size; KM fidelity and onset timing are stable, exhibiting bootstrap-level variance.
  ΔKM\Delta_{\text{KM}} ↓\downarrowOS KS ↓\downarrow 
Trial (arm, nrn_{r} )Extraction IAEExtractedOracleExtractedOraclePropagated Δ\Delta
NCT03041311 (Placebo, 53)0.03090.055 ±\pm 0.0180.040 ±\pm 0.0220.238 ±\pm 0.0650.114 ±\pm 0.0350.015
NCT02499770 (E/P, 37)0.03450.056 ±\pm 0.0180.049 ±\pm 0.0180.156 ±\pm 0.0350.147 ±\pm 0.0350.007
NCT00844649 (Gemcitabine, 430)0.01330.018 ±\pm 0.0050.012 ±\pm 0.0060.063 ±\pm 0.0140.043 ±\pm 0.0140.006
Appendix
Table 4: Extraction-error propagation (EXP1). Extracted regenerates from the curve traced off the published figure; oracle regenerates from the KM reconstructed from held-out real IPD. Propagated Δ\Delta is the difference in mean ΔKM\Delta_{\text{KM}} . Values are mean ±\pm std over 30 regenerations. Unlike Table 2 , which pools all arms of a trial, these numbers are for the single matched arm, so the extracted-input column is not identical to the headline ΔKM\Delta_{\text{KM}} .
TrialMetricση=0.4\sigma_{\eta}{=}0.4ση=0.6\sigma_{\eta}{=}0.6ση=0.8\sigma_{\eta}{=}0.8ση=1.0\sigma_{\eta}{=}1.0
NCT03041311AEs/patient (real: 15.1)10.1 ±\pm 0.511.3 ±\pm 0.813.0 ±\pm 1.517.2 ±\pm 2.3
Burden KS ↓\downarrow0.277 ±\pm 0.0290.218 ±\pm 0.0470.190 ±\pm 0.0450.165 ±\pm 0.041
ΔKM\Delta_{\text{KM}} ↓\downarrow0.0490.0490.0490.049
AE cosine ↑\uparrow0.8520.8500.8520.861
NCT02499770AEs/patient (real: 12.8)6.9 ±\pm 0.67.7 ±\pm 0.69.5 ±\pm 1.210.5 ±\pm 1.2
Burden KS ↓\downarrow0.457 ±\pm 0.0570.407 ±\pm 0.0440.344 ±\pm 0.0570.352 ±\pm 0.057
Appendix
Table 5: Propensity dispersion sweep (EXP2), 30 regenerations per cell. Burden approaches the real value monotonically in ση\sigma_{\eta} while ΔKM\Delta_{\text{KM}} is invariant (identical to four decimal places, as survival assignment does not depend on ση\sigma_{\eta} ) and AE frequency moves within noise. Best burden KS per trial in bold.
 NCT03041311NCT02499770NCT00844649
 nr=53, ns=107n_{r}{=}53,\,n_{s}{=}107nr=37, ns=75n_{r}{=}37,\,n_{s}{=}75nr=430, ns=861n_{r}{=}430,\,n_{s}{=}861
Demographics
Sex JSD ↓\downarrow0.0054 ±\pm 0.00640.0033 ±\pm 0.00600.0005 ±\pm 0.0006
ECOG JSD ↓\downarrow0.272 ±\pm 0.0480.013 ±\pm 0.0120.003 ±\pm 0.001
Weight KS ↓\downarrow0.204 ±\pm 0.0390.200 ±\pm 0.0560.125 ±\pm 0.017
Height KS ↓\downarrow0.207 ±\pm 0.0290.179 ±\pm 0.0420.103 ±\pm 0.013
Appendix
Table 6: Full Stage-B fidelity at the recommended ση=0.8\sigma_{\eta}{=}0.8 (compare Table 2 , which uses ση=0.4\sigma_{\eta}{=}0.4 ). Mean ±\pm std over 30 independent regenerations, synthetic cohorts at 2×2{\times} real enrollment. Only the burden rows change materially; demographic, survival, and AE-frequency rows are within resampling noise of the ση=0.4\sigma_{\eta}{=}0.4 values.
Figure 12: Weight, height, and BMI distributions (real vs. synthetic) for all three trials. KS statistics are ≤0.24{\leq}0.24 in all cases; the truncated-Gaussian sampling reproduces the overall shape of real anthropometric distributions.
Figure 13: AE burden per patient. Dashed lines show means.
#PlotCategoryIAEMed. AEMed. OS err (mo)Arms
1stress_combo_flat_overlapCombo stress0.00120.00050.0002
2edge_high_survivalEdge0.00120.00070.0002
3edge_near_flatEdge0.00290.00160.0322
4synthetic_002Standard0.00330.00160.0952
5edge_ci_shadingEdge0.00350.00190.1052
6synthetic_003Standard0.00400.00170.2322
Appendix
Table 7: Per-plot attempt-1 extraction metrics, sorted by IAE (best first).
Figure 14: synthetic_002 : Standard 2-arm plot over 36 mo. Clean color separation; steps tracked precisely. synthetic_003 : Standard 2-arm plot over 36 mo with tighter step spacing.
Figure 15: synthetic_001 : Two well-separated arms over 12 mo with clear step structure. Extraction markers match the steps visually with no systematic drift. synthetic_004 : Standard 2-arm, 18 mo. Accurate step tracking.
Figure 16: synthetic_005 : Standard 2-arm, 12 mo with steep drop. Vertical-drop detection picks up the final step.
Figure 17: edge_high_survival : Minimal events over 144 mo follow-up. Red and blue ring markers track both arms cleanly through the entire range, including closely-overlapping segments near 96–120 mo. edge_near_flat : Near-flat mortality profile. Ring density is high and follows the tight step structure closely.
Figure 18: edge_ci_shading : Confidence-interval shading suppressed by the color mask; curves tracked cleanly. edge_four_arms : 4-arm comparison; color separation and continuity tracking both hold.
Figure 19: edge_small_dense : Dense step structure on a 36-month scale; fine sampling required. edge_cumulative_incidence : Cumulative-incidence plot (curves rising). Monotonicity constraint inverted; rising structure recovered.
Figure 20: edge_multi_panel : Multi-panel layout (4 arms across subplots). Per-panel xmaxx_{\text{max}} differs, inflating the aggregate IAE relative to what any single panel would show.
Figure 21: stress_gridlines : Light gridlines scrubbed by the >> 30% row/column coverage rule before tracing. stress_annotation_heavy : Multiple in-plot annotations masked before tracing.
Figure 22: stress_three_similar : Three similarly-hued arms. HSL mask separates them with a narrow margin. stress_multi_panel_clean : Clean multi-panel layout; per-panel xmaxx_{\text{max}} causes minor ( <0.002<0.002 ) mismatch between saved and recomputed IAE.
Figure 23: stress_jpeg_artifact : JPEG compression smears the edge; extraction stays within 1–2 px. stress_legend_overlap : Legend overlaps the plot area; masked as a rectangular exclusion zone before tracing.
Figure 24: stress_bw : Black & white with solid vs dashed distinction. Grayscale separation holds; dashed gaps bridged. stress_tiny : Low pixel density; small mask errors translate to larger absolute errors in S(t)S(t) .
Figure 25: stress_stretched_tall : Stretched vertically; tick calibration resolves the bbox but column density is lower. stress_blurry : Gaussian blur widens the curve edge symmetrically; centroid strategy absorbs it.
Figure 26: stress_stretched_wide : Stretched horizontally; abundant columns, shallow SS -axis resolution. stress_diep_like : Two-panel cumulative-incidence plot with 8 arms total and heavy overlap. Per-arm IAE ranges 0.042–0.072; color attribution is the bottleneck—several arms share hue bands, so the tracer occasionally jumps between arms in dense regions.
Figure 27: stress_combo_flat_overlap : Both arms stay between 0.95–1.00 across 60 mo. Step markers land on the curves throughout; the combination of overlap and flatness defeats itself—any pixel near the curve gives near-zero error. stress_combo_bw_overlap : B&W with overlapping solid and dashed curves; grayscale plus gap bridging resolves both.
Figure 28: stress_combo_4arm_tiny : 4 arms at tiny resolution; color separation holds but per-column sample count is low. stress_combo_tiny_bw : Tiny plus B&W; low pixel count compounds grayscale ambiguity.
Figure 29: stress_combo_tiny_blurry : Tiny plus blurry; edge widening dominates, yielding the highest IAE among these single-style combos. stress_combo_stretched_dense : Stretched aspect with dense step structure; occasional drift in densest regions.
Figure 30: stress_combo_jpeg_blurry_dark : Heavy JPEG artifacts and blur widen the curve edge. Centroid tracking keeps the extraction within a few pixels of truth, but the smeared edge produces noisier step timing. stress_combo_grid_annotation_bw : Adversarial combo: gridlines leak into the color mask, the “Median OS” diagonal leader line pulls the red markers off the true curve near t=14t=14 , and B&W forces grayscale separation of solid vs dashed lines. Three failure modes stack.
More in this topic

Synthetic Clinical Data Generation

See all 22 papers

Explore similar work

CardsList
Mar 6, 2026

From Statistical Fidelity to Clinical Consistency: Scalable Generation and Auditing of Synthetic Patient Trajectories

Access to electronic health records (EHRs) for digital health research is often limited by privacy regulations and institutional barriers. Synthetic EHRs have been proposed as a way to enable safe and sovereign data sharing; however, existing methods may produce records that capture overall statistical properties of real data but present inconsistencies across clinical processes and observations. We developed an integrated pipeline to make synthetic patient trajectories clinically consistent through two synergistic steps: high-fidelity generation and scalable auditing. Using the MIMIC-IV database, we trained a knowledge-grounded generative model that represents nearly 32,000 distinct clinical events, including demographics, laboratory measurements, medications, procedures, and diagnoses, while enforcing structural integrity. To support clinical consistency at scale, we incorporated an automated auditing module leveraging large language models to filter out clinical inconsistencies (e.g., contraindicated medications) that escape probabilistic generation. We generated 18,071 synthetic patient records derived from a source cohort of 180,712 real patients. While synthetic clinical event probabilities demonstrated robust agreement (mean bias effectively 0.00) and high correlation (R2=0.99) with the real counterparts, review of a random sample of synthetic records (N=20) by three clinicians identified inconsistencies in 45-60% of them. Automated auditing reduced the difference between real and synthetic data (Cohen's effect size d between 0.59 and 1.60 before auditing, and between 0.18 and 0.67 after auditing). Downstream models trained on audited data matched or even exceeded real-data performance. We found no evidence of privacy risks, with membership inference performance indistinguishable from random guessing (F1-score=0.51).

Guanglin Zhou, Armin Catic, Motahare Shabestari +4

The University of Queensland, Brisbane, QLD, Australia · The University of New South Wales, Sydney, NSW, Australia · Shahid Sadoughi University of Medical Sciences and Health Services, Yazd, Iran +1

Synthetic Data GenerationSynthetic Clinical Data GenerationSynthetic Data EvaluationLongitudinal EHR Modeling
AskPDF
Jun 30, 2026

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

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, 00--100100 scale), FoGS yields mean improvements of +2.17+2.17 in C-index and +0.67+0.67 in IBS, improving both metrics on 9 of 16 datasets and at least one on 13 (one-sided Wilcoxon p=0.039p=0.039 and p=0.035p=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.

Niccolò Maria Rizzi, Eugenio Lomurno, Alberto Archetti +1

Politecnico di Milano, Milan, Italy

Survival AnalysisSynthetic Data GenerationSynthetic Data EvaluationSynthetic Data SelectionSynthetic Population Generation
AskPDF
Apr 19, 2026

Prior-Fitted Functional Flow: In-Context Generative Models for Pharmacokinetics

We introduce Prior-Fitted Functional Flows, a generative foundation model for pharmacokinetics that enables zero-shot population synthesis and individual forecasting without manual parameter tuning. We learn functional vector fields, explicitly conditioned on the sparse, irregular data of an entire study population. This enables the generation of coherent virtual cohorts as well as forecasting of partially observed patient trajectories with calibrated uncertainty. We construct a new open-access literature corpus to inform our priors, and demonstrate state-of-the-art predictive accuracy on extensive real-world datasets.

César Ojeda, Niklas Hartung, Wilhelm Huisinga +6

University of Potsdam, Germany · Technical University Berlin, Germany · Freie Universität Berlin, Germany +2

Probabilistic Time Series ForecastingHealthcareSynthetic Data GenerationZero-Shot Time Series ForecastingContinuous Normalizing FlowsPrior-Data Fitted Networks+2
AskPDF