Saturday, November 21, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News Technology and Engineering

Generative Models Replicating Neonatal Tabular Data Face Steep Realism Tradeoffs

November 21, 2026
in Technology and Engineering
Harold Sullivan
By Harold Sullivan Scienmag Editorial Profile - Maternal and Child Health
Reading Time: 6 mins read
0
Generative Models Replicating Neonatal Tabular Data Face Steep Realism Tradeoffs

Generative Models Replicating Neonatal Tabular Data Face Steep Realism Tradeoffs

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

A comprehensive multidimensional benchmark evaluating synthetic health data algorithms has demonstrated that while deep generative and copula architectures can capture physiological distributions from small neonatal datasets, all evaluated models produce synthetic records that remain readily distinguishable from real clinical cases. The investigation examined how seven distinct generative paradigms balance distributional fidelity, predictive utility, adversarial realism, and privacy preservation when trained on complex electronic records. The empirical findings reveal that high statistical fidelity does not translate into authentic clinical realism or robust privacy protection, indicating that current synthetic generation techniques cannot yet replace real records for predictive modeling or clinical deployment.

Neonatal clinical records represent an exceptionally challenging domain for machine learning due to small cohort sizes, severe diagnostic imbalances, and stringent ethical constraints surrounding patient confidentiality. When researchers develop artificial intelligence tools for fragile newborn populations, the limited availability of high-quality training records often restricts model generalizability and external validation. Generating synthetic tabular datasets has emerged as a proposed solution to expand sample sizes while mitigating the direct exposure of identifiable information. However, medical tabular data contains intricate inter-variable dependencies and highly skewed categorical outcomes, and generative models do not inherently protect privacy or ensure downstream utility without comprehensive multidimensional evaluation.

To investigate these dynamics, researchers evaluated seven generative algorithms: Conditional Tabular GAN (CTGAN), a Variational Autoencoder (VAE), a Conditional GAN (CGAN), an unconditional classical GAN, CopulaGAN, a parametric Gaussian Copula, and a purpose-built Transformer-based architecture operating over per-feature token embeddings. The models were evaluated on 252 neonatal clinical records collected from the Medical Records Department of Rumah Sakit Umum Daerah dr. Soedono in East Java, Indonesia. The dataset documented eight continuous physiological variables—birth weight, body length, heart rate, respiratory rate, body temperature, head circumference, chest circumference, and gestational age—alongside a highly skewed target variable comprising 30 distinct nursing diagnoses.

The experimental framework implemented a shared, stratified 80/20 train–test split computed on a consolidated eight-category diagnosis representation, allocating 201 records for model training and setting aside 51 unseen records for evaluation. Each generative architecture was trained independently across five random seeds, with each seed producing 1,000 synthetic records. The researchers assessed fidelity using the Kolmogorov–Smirnov statistic across numerical features, Jensen–Shannon Divergence across the full 30-category diagnosis distribution, and absolute differences in Pearson correlation matrices. Downstream utility was scored across four machine learning classifiers under Train on Synthetic, Test on Real (TSTR), Train on Real, Test on Real (TRTR), and Train on Real, Test on Synthetic (TRTS) schemes.

Distributional fidelity assessments showed that the Transformer-based model and the Gaussian Copula achieved the highest consistency with the training data. The Transformer-based model attained the lowest numerical distribution disparity, recording a mean Kolmogorov–Smirnov statistic of 0.177, alongside the lowest correlation matrix difference at 0.095. The Gaussian Copula achieved the second-lowest correlation difference at 0.108 and a Kolmogorov–Smirnov value of 0.272. In contrast, classical GAN suffered severe categorical mode collapse, generating synthetic cohorts with Jensen–Shannon Divergence reaching 0.921 and failing to capture multi-class diagnostic diversity. Conditional GAN achieved near-perfect categorical alignment with a divergence of 0.0047, but demonstrated poor numerical accuracy with a Kolmogorov–Smirnov statistic of 0.679.

Analysis of individual clinical variables demonstrated that physiological complexity heavily influenced synthesis quality. Respiratory rate proved the most difficult continuous feature to model across five architectures, reflecting its noisy and skewed empirical baseline in clinical settings. Furthermore, missing value imputation introduced structural vulnerabilities for specific model families: when five missing values in chest circumference were filled with a static column mean, both the Gaussian Copula and CopulaGAN exhibited severe distributional distortions, yielding Kolmogorov–Smirnov statistics of 0.995. When re-evaluated using median imputation during sensitivity testing, the Gaussian Copula’s error on chest circumference dropped to 0.104, confirming that deterministic imputation spikes can distort copula-based probability transforms.

Predictive utility testing across tuned Random Forest, Logistic Regression, XGBoost, and LightGBM classifiers revealed that synthetic training sets lagged behind authentic clinical data. Under the primary TSTR scheme, the Transformer-based model attained the highest mean macro-averaged F1 score of 0.177 with Random Forest, reaching approximately 82 percent of the 0.215 score achieved by the real-data TRTR baseline. While the Transformer architecture maintained numerically superior F1 scores across all four classifiers, statistical comparisons applying Holm and Bonferroni adjustments confirmed that its predictive margins over CTGAN, Gaussian Copula, VAE, and CGAN were not statistically significant, separating only from the failing classical GAN and CopulaGAN architectures.

Realism assessments using cross-validated adversarial detection revealed an acute vulnerability shared by every tested model. When a Random Forest discriminator was trained to distinguish synthetic cases from real neonatal records, the detection Area Under the ROC Curve ranged between 0.986 for CTGAN and 1.000 for CGAN, classical GAN, CopulaGAN, and Gaussian Copula. Because every synthetic dataset remained nearly perfectly distinguishable from real clinical cases, none of the generative algorithms produced statistically convincing neonatal records. Even though automated physiological screening confirmed that over 99.4 percent of synthetic records fell within valid clinical ranges, the adversarial classifier readily separated artificial records based on multidimensional feature patterns.

Privacy evaluations highlighted distinct disclosure profiles across the architectures, demonstrating that high fidelity can coincide with partial memorization. Distance to Closest Record analysis revealed that five of the seven models produced synthetic cases systematically closer to their training data than to held-out records. The Variational Autoencoder presented the highest memorization footprint, with 4.60 percent of its synthetic output matching training records within a normalized distance threshold of 0.01. Conversely, Gaussian Copula produced zero near-duplicates and exhibited a privacy gap ratio above 1.0.

The researchers emphasized several structural constraints that contextualize these experimental findings. The single-center sample of 252 records and reliance on five random seeds restricted statistical power, rendering observed performance differences indicative rather than conclusive. Furthermore, the synthetic cohorts underwent neither external multi-institutional validation nor individual review by practicing neonatal clinicians to confirm that physiological combinations matched assigned nursing diagnoses. Consequently, the authors conclude that generative modeling of neonatal tabular data presents unavoidable multidimensional trade-offs, and synthetic clinical records are not currently suitable for real-world predictive substitution or public data sharing.

Data preparation revealed notable distribution characteristics and required rigorous cleaning prior to model training. Missing values in the raw dataset were confined strictly to three physiological variables: gestational age affected 9.92 percent of records (25 cases), while head circumference and chest circumference each lacked data in 1.98 percent of cases (5 records). Text processing resolved spelling, spacing, and word-order irregularities across 3.97 percent of the clinical entries, translating Indonesian hospital records into standardized English aligned with NANDA International terminology and condensing 31 raw entries into 30 distinct diagnoses. Extreme diagnostic disparity characterized the cohort: the predominant category, Risk for Hypothermia, represented 52.4 percent of all infants with 132 cases, whereas 17 conditions appeared only once.

Architectural design choices further illuminated differences in how models processed tabular clinical features. The custom Transformer architecture converted each physiological scalar and the diagnostic category into individual feature tokens, processing the resulting nine-token sequence through a two-layer, four-head encoder with learned slot embeddings before mapping it to a 16-dimensional latent Gaussian space with a Kullback–Leibler divergence weight of 0.1. In comparison, CTGAN implemented a Wasserstein critic regularized by a gradient penalty parameter of ten and used conditional vector sampling. Downstream utility results demonstrated that predictive retention varied sharply across classifier families. Relative to the authentic real-data baseline, synthetic data generated by the Transformer attained 104 percent of baseline performance under Logistic Regression, 82 percent under Random Forest, 67 percent under LightGBM, and 47 percent under XGBoost. The Transformer architecture also produced the lowest multiclass Brier score at 0.730, confirming superior probability calibration compared to classical GAN, which registered an error of 1.592.

Supplementary testing explored whether synthetic cohorts could serve as substitute test sets or benefit from architectural adjustments. When real-trained classifiers were evaluated against synthetic records under the Train on Real, Test on Synthetic scheme, CTGAN registered the highest macro-averaged F1 score at 0.249. In contrast, the Variational Autoencoder yielded the highest overall classification accuracy at 0.791 but achieved a low macro F1 score of 0.157. This divergence revealed that the autoencoder heavily over-represented the majority hypothermia category, illustrating how unweighted accuracy metrics can mask severe synthetic class collapse. Sensitivity tests examining generated sample volumes across 500, 1,000, and 2,000 records confirmed that distributional fidelity and predictive utility remained largely invariant to cohort size, with the Transformer maintaining the top macro F1 score at each volume (0.159, 0.177, and 0.209). Furthermore, retraining CopulaGAN with larger networks and extended epochs reduced its Kolmogorov–Smirnov metric from 0.724 to 0.513 without improving predictive utility, demonstrating that expanded capacity alone cannot overcome tabular structural limitations.

Subject of Research: Multidimensional benchmarking of generative algorithms synthesizing neonatal clinical tabular data

Article Title: Benchmarking fidelity, utility, realism, and privacy tradeoffs among generative models for synthetic neonatal clinical tabular data

Article References: Munandar, T. A., Lestari, T. S., Noeman, A., Alimuddin, & Arafiyah, R. (2026). Benchmarking fidelity, utility, realism, and privacy tradeoffs among generative models for synthetic neonatal clinical tabular data. Discover Artificial Intelligence, 6(1), Article 1937. https://doi.org/10.1007/s44163-026-02764-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02764-1

Keywords: synthetic health data, neonatal clinical data, tabular generative models, generative adversarial networks, variational autoencoders, Gaussian copula, CTGAN, CopulaGAN, transformer models, data fidelity, predictive utility, adversarial detection, membership inference attack, distance to closest record, class imbalance, clinical machine learning

Cite Scienmag News

Harold Sullivan. (November 21, 2026). Generative Models Replicating Neonatal Tabular Data Face Steep Realism Tradeoffs. Scienmag. https://scienmag.com/generative-models-replicating-neonatal-tabular-data-face-steep-realism-tradeoffs/

Harold Sullivan. "Generative Models Replicating Neonatal Tabular Data Face Steep Realism Tradeoffs." Scienmag, 21 November 2026, https://scienmag.com/generative-models-replicating-neonatal-tabular-data-face-steep-realism-tradeoffs/. Accessed 21 November 2026.

Harold Sullivan. "Generative Models Replicating Neonatal Tabular Data Face Steep Realism Tradeoffs." Scienmag. November 21, 2026. https://scienmag.com/generative-models-replicating-neonatal-tabular-data-face-steep-realism-tradeoffs/

Tags: adversarial detectionclass imbalanceclinical machine learningCopulaGANCTGANdata fidelitydistance to closest recordGaussian copulagenerative adversarial networksmembership inference attackneonatal clinical datapredictive utilitysynthetic health datatabular generative modelstransformer modelsVariational autoencoders
Share26Tweet16
Previous Post

Picroside II Inhibits Temozolomide Resistant Glioblastoma Growth by Upregulating DHRS2

Next Post

Child Psychiatry Fellows Report Positive Experiences with New Psychoanalysis Curriculum

Related Posts

Discrete Dislocation Dynamics Model Simulates Physical Dislocation Transmission Across Grain Boundaries
Technology and Engineering

Discrete Dislocation Dynamics Model Simulates Physical Dislocation Transmission Across Grain Boundaries

November 21, 2026
Simulated Carbon Nanotube Interlayers Boost Lead-Free Perovskite Solar Efficiency to 29 Percent
Technology and Engineering

Simulated Carbon Nanotube Interlayers Boost Lead-Free Perovskite Solar Efficiency to 29 Percent

November 21, 2026
Decoupled Drone Vision Architecture Counts Grazing Sheep in Oblique Aerial Imagery
Agriculture

Decoupled Drone Vision Architecture Counts Grazing Sheep in Oblique Aerial Imagery

November 21, 2026
Optimized Quarry Blasting Cuts Limestone Fragment Size and Processing Costs in Ethiopia
Technology and Engineering

Optimized Quarry Blasting Cuts Limestone Fragment Size and Processing Costs in Ethiopia

November 21, 2026
Parallel Multi-Transformer Ensemble Accurately Estimates Agile Software Development Effort
Technology and Engineering

Parallel Multi-Transformer Ensemble Accurately Estimates Agile Software Development Effort

November 21, 2026
Clay-Supported Calcium-Magnesium Catalyst Converts Jatropha Oil to High-Yield Biodiesel
Technology and Engineering

Clay-Supported Calcium-Magnesium Catalyst Converts Jatropha Oil to High-Yield Biodiesel

November 21, 2026
Next Post
Child Psychiatry Fellows Report Positive Experiences with New Psychoanalysis Curriculum

Child Psychiatry Fellows Report Positive Experiences with New Psychoanalysis Curriculum

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Slightly Alkaline Growth Conditions Maximize Magnetite Crystal Nucleation in Magnetic Bacteria
  • Nonlocal Multilayer Metastructure Demonstrates Far-Field Superresolution Imaging via Momentum Space Superoscillation
  • Hop Compound Xanthohumol Restores Colistin Susceptibility in Resistant Escherichia Coli
  • Discrete Dislocation Dynamics Model Simulates Physical Dislocation Transmission Across Grain Boundaries

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading