DISCO-TAB: A Hierarchical Reinforcement Learning Framework for Privacy-Preserving Synthesis of Complex Clinical Data
Authors: Arshia Ilaty, Hossein Shirazi, Amir Rahmani, Hajar Homayouni
Organizations: Computational Science Research Center, San Diego State University, San Diego, CA 92182 USA. · School of Nursing and Department of Computer Science, University of California, Irvine, CA 92697 USA. · Fowler College of Business, San Diego State University, San Diego, CA 92182 USA. · Department of Computer Science, San Diego State University, San Diego, CA 92182 USA.
The development of robust clinical decision support systems is frequently impeded by the scarcity of high-fidelity, privacy-preserving biomedical data. While Generative Large Language Models (LLMs) offer a promising avenue for synthetic data generation, they often struggle to capture the complex, non-linear dependencies and severe class imbalances inherent in Electronic Health Records (EHR), leading to statistically plausible but clinically invalid records. To bridge this gap, we introduce DISCO-TAB (DIScriminator-guided COntrol for TABular synthesis), a novel framework that orchestrates a fine-tuned LLM with a multi-objective discriminator system optimized via Reinforcement Learning. Unlike prior methods relying on scalar feedback, DISCO-TAB evaluates synthesis at four granularities, token, sentence, feature, and row, while integrating Automated Constraint Discovery and Inverse-Frequency Reward Shaping to autonomously preserve latent medical logic and resolve minority-class collapse. We rigorously validate our framework across diverse benchmarks, including high-dimensional, small-sample medical datasets (e.g., Heart Failure, Parkinson's). Our results demonstrate that hierarchical feedback yields state-of-the-art performance, achieving up to 38.2% improvement in downstream clinical classifier utility compared to GAN and Diffusion baselines, while ensuring exceptional statistical fidelity (JSD < 0.01) and robust resistance to membership inference attacks. This work establishes a new standard for generating trustworthy, utility-preserving synthetic tabular data for sensitive healthcare applications.
Figures & tables
Fig. 1: The DISCO-TAB framework for tabular data synthesis. (A) Automated Constraint Discovery: The system analyzes the real dataset ( Dreal ) to compute a global correlation matrix and identify statistically significant feature dependencies (e.g., Age vs. Comorbidities), which are used to initialize the feature-level discriminator ( Dfeat ). (B) Hierarchical Discriminators: A multi-objective discriminator ensemble evaluates generated records ( S ) at four semantic granularities: token-level (syntactic validity), sentence-level (local semantic plausibility), feature-level (dependency consistency), and row-level (global coherence). (C) Inverse Frequency Reward Shaping (IFRS): During reinforcement learning, the PPO reward is scaled based on the rarity of the target class associated with the generated record, mitigating mode collapse in imbalanced datasets.
Dataset
CTGAN
TAEGAN
TabDDPM
TabSyn
RealTabFormer
MostlyAI
TAGAL
DISCO-TAB
Δ
Heart Failure
0.778
0.655
0.665
0.877
0.548
0.598
0.839
1.000
+0.123
Breast Cancer
0.549
0.693
0.689
0.934
0.495
0.498
0.990
0.994
+0.004
Liver Disorders
0.496
0.576
0.133
0.510
0.528
0.574
0.608
0.990
+0.382
Parkinsons
0.594
0.595
0.637
0.610
0.637
0.637
0.886
0.966
+0.080
Obesity
0.105
0.215
0.577
0.902
0.250
0.584
0.790
0.929
+0.027
Diabetes
0.881
0.875
0.840
0.950
0.968
0.969
0.882
0.865
-0.104
TABLE I: Downstream ML utility (F1-score) under the Train-on-Synthetic, Test-on-Real (TSTR) protocol. Δ denotes the absolute percentage point differences between DISCO-TAB and the strongest baseline per dataset (positive indicates DISCO-TAB superiority; negative indicates a stronger baseline). The first six datasets are medical benchmarks, while the last three (German Credit, Bank Marketing, Census Income) are non-medical datasets included to evaluate cross-domain generalizability.
DISCO-TAB Metrics
Baseline KS Comparison (Lower is Better)
Dataset
JS ↓
KL ↓
Hell. ↓
KS ↓
Pearson ↑
MostlyAI
CTGAN
TAEGAN
RealTab
TabDDPM
TabSyn
TAGAL
Heart Failure
0.003
0.017
0.044
0.022
0.968
0.085
0.147
0.212
0.125
0.098
0.045
0.039
Breast Cancer
0.011
0.090
0.095
0.025
0.992
0.038
0.339
0.268
0.150
0.210
0.032
0.045
Liver Disorders
0.017
0.147
0.094
0.017
0.978
0.063
0.337
0.416
0.220
0.350
0.085
0.095
Parkinsons
0.005
0.028
0.067
0.034
0.999
0.066
0.315
0.374
0.110
0.095
0.052
0.062
Obesity
0.008
0.044
0.060
0.023
0.986
0.184
0.301
0.274
0.250
0.045
0.035
0.080
TABLE II: Unified statistical fidelity comparison. Columns under DISCO-TAB report our performance across five metrics. The right side compares the KS Statistic (lower is better) against baselines. DISCO-TAB achieves consistently low KS divergence across all medical datasets, consistently outperforming Diffusion models (TabDDPM) on complex data. Pearson correlation is reported exclusively for continuous feature pairs. In addition, distributional similarity metrics between real and synthetic data. KL divergence ( DKL(P∥Q) ) is computed with additive smoothing ( ϵ=10−8 ) to ensure numerical stability. Jensen–Shannon Divergence (JSD) is additionally reported as a symmetric and bounded alternative.
Dataset
Metric
TabDDPM
MostlyAI
DISCO-TAB
Heart Failure
Utility (F1)
0.665
0.598
1.000
Privacy (DCR)
4830.6
895.0
1.27
Obesity
Utility (F1)
0.577
0.584
0.929
Privacy (DCR)
6.8
1.9
1.5
German Credit
Utility (F1)
0.592
0.626
0.925
Privacy (DCR)
3.3
2.1
1.9
TABLE III: Privacy-Utility Trade-off across diverse scales. We report Utility (F1-Score) and Privacy (DCR). TabDDPM consistently achieves extreme DCR values due to diffusion noise, but often at the cost of utility (e.g., on Obesity). DISCO-TAB maintains a pragmatic safe zone ( DCR>0 ) while strictly prioritizing the clinical fidelity required for high F1 scores.
Dataset
CTGAN
TAEGAN
TabDDPM
TabSyn
MostlyAI
Real.
DISCO-Tab
Heart Fail.
2.2
2.5
10.1
8.5
4.2
5.1
4.8
Breast Can.
2.3
2.6
10.4
8.7
4.3
5.2
4.9
Liver Dis.
2.1
2.4
9.8
8.3
4.1
4.9
4.6
Parkinsons
2.1
2.4
9.8
8.4
4.1
5.0
4.7
Obesity
2.4
2.7
10.8
8.9
4.4
5.3
5.0
Diabetes
2.4
2.8
11.4
9.2
4.7
5.5
5.2
TABLE IV: Computational Efficiency: Peak GPU Memory Usage (GB). Datasets are grouped by scale (Small vs. Large). DISCO-TAB remains memory-efficient ( <6 GB), unlike Diffusion models.
CTGAN
TAEGAN
TabDDPM
TabSyn
RealTabFormer
TAGAL
MostlyAI
DISCO-TAB
Num. JSD
0.145
0.130
0.098
0.085
0.092
0.140
0.035
0.038
Cat. JSD
0.210
0.195
0.155
0.120
0.048
0.030
0.042
0.015
TABLE V: Feature-Type Fidelity Analysis (Mean JSD ↓ ). DISCO-TAB bridges the gap between numerical and categorical precision.
Model
Overall FAITH
Factuality
Alignment
Integrity
Tracking
Behavioral Profile
MostlyAI
0.9747
0.9993
0.9026
0.9970
0.9999
Balanced Commercial Baseline
Taegan
0.9572
1.0000
0.8618
1.0000
0.9671
High Constraint / Lower Correlation
Tab-DDPM
0.6773
0.1976
0.8722
1.0000
0.6392
Privacy via Noise (High Integrity, Low Factuality)
DISCO-TAB
0.7703
0.8750
0.9839
0.2303
0.9920
Precision via Manifold Adherence
TABLE VI: FAITH Framework Decomposition. DISCO-TAB dominates in Alignment ( 0.9839 ), proving it captures the physics of clinical data better than any baseline. However, its low Integrity score reveals a fundamental trade-off: unlike diffusion models ( Tab-DDPM ) that achieve high Integrity by generating noisy samples far from the manifold, DISCO-TAB adheres strictly to the clinical distribution, prioritizing precision over diversity.
Category
Configuration
Relative Perf.
Time (min)
System
Full DISCO-TAB (PPO)
100%
26.3
Variant: DPO
99.2%
24.8
No Hierarchy (Row Only)
86.9%
16.5
Minimal Baseline (No RL)
63.0%
8.7
Discriminators
All Four Levels
100%
26.3
w/o Token Disc
97.2%
24.1
TABLE VII: Comprehensive Ablation Study Results. Aggregate performance is computed as the unweighted mean TSTR F1-score across all datasets. All datasets are weighted equally to avoid dominance by large benchmarks.
The development of medical AI is constrained by limited access to high-quality clinical data due to institutional silos and strict privacy regulations such as HIPAA and GDPR. Synthetic data generation offers a potential solution, but existing methods lack principled mechanisms to explicitly manage the privacy-utility trade-off, often degrading clinically meaningful patterns or risking patient re-identification. We present PSyGenTAB, a privacy-preserving generative framework that formulates synthetic healthcare data generation as a constrained optimization problem solved using the Augmented Lagrangian Method. By embedding configurable privacy constraints directly into model training, PSyGenTAB enforces minimum privacy thresholds while maximizing clinical data utility. Across multiple clinically motivated benchmarks, PSyGenTAB preserves inter-feature clinical relationships and minority-class diagnostic patterns essential for reliable health AI. Downstream evaluation using Train-on-Synthetic, Test-on-Real and Train-on-Real, Test-on-Synthetic protocols shows that models trained on synthetic data achieve performance comparable to those trained on real patient records. Privacy auditing further demonstrates reduced exact record reproduction and strong resilience to membership inference attacks. These results establish PSyGenTAB as a principled framework for balancing privacy protection and clinical utility in synthetic healthcare data, supporting secure cross-institutional AI development.
Arshia Ilaty, Hossein Shirazi, Manasi Chitale +5
San Diego State University, San Diego, CA 92182 USA. · University of California, Irvine, CA 92697 USA.
Deep generative models can help with data scarcity and privacy by producing synthetic training data, but they struggle in low-data, imbalanced tabular settings to fully learn the complex data distribution. We argue that striving for the full joint distribution could be overkill; for greater data efficiency, models should prioritize learning the conditional distribution P(y∣X), as suggested by recent theoretical analysis. Therefore, we overcome this limitation with ReTabSyn, a Reinforced Tabular Synthesis pipeline that provides direct feedback on feature correlation preservation during synthesizer training. This objective encourages the generator to prioritize the most useful predictive signals when training data is limited, thereby strengthening downstream model utility. We empirically fine-tune a language model-based generator using this approach, and across benchmarks with small sample sizes, class imbalance, and distribution shift, ReTabSyn consistently outperforms state-of-the-art baselines. Moreover, our approach can be readily extended to control various aspects of synthetic tabular data, such as applying expert-specified constraints on generated observations.
Xiaofeng Lin, Seungbae Kim, Zhuoya Li +3
University of California, Los Angeles, USA · University of South Florida, USA · Cisco, USA +1
Synthetic data is widely used in healthcare to create datasets that preserve statistical properties of real data without exposing sensitive patient information. Generating and evaluating synthetic data across privacy, utility, and fairness dimensions is crucial for enabling high-quality data availability in downstream prediction tasks and clinical decision making. We present Memisis, a tool that orchestrates and evaluates synthetic data by leveraging existing synthesis libraries, large language models (LLMs), and state-of-the-art evaluation metrics. Our tool creates a unified workflow for data generation, validation, and evaluation. Users can control training size, training epochs, and the number of synthetic rows to sample. Beyond manual configuration, an interactive agent mode allows users to specify data generation goals in natural language, and the tool orchestrates the full pipeline by invoking existing synthesizers while performing the requisite evaluation. For the demo, we use an open-source schizophrenia dataset with protected attributes related to race and gender, evaluate six synthesizers spanning GANs, VAEs, diffusion models, and normalizing flows, and use a local LLM to orchestrate the workflow. The system affords users flexibility and control over the data generation and evaluation process.