685 - Evaluating Generative Augmentation and Synthetic Data Generation to Improve Predictive Modeling of Necrotizing Enterocolitis in Preterm Infants

Publication Number: 1662.685

Shilpa Muralidhar, Texas Children's Hospital, Bellaire, TX, United States; Mohan Pammi, Baylor College of Medicine, Bellaire, TX, United States

Background: Necrotizing enterocolitis (NEC) is a severe gastrointestinal emergency in preterm infants, linked to high mortality and morbidity. Predictive modeling holds promise for early detection of neonatal diseases such as NEC. However, neonatal clinical datasets are often small, imbalanced, and constrained by privacy regulations. Generative models can synthesize realistic data to enhance statistical power, improve prediction of rare outcomes, and protect confidentiality.

Objective: To evaluate whether synthetic neonatal data generated using a Tabular Variational Autoencoder (TVAE) improve model performance for NEC prediction while maintaining resemblance and privacy

Design/Methods: We analyzed Vermont Oxford dataset, encompassing 3,293 preterm infants (NEC 6.4%, 50 features) at Texas Children's Hospital from 2008 to 2022. Approval for this study was granted by the Baylor College of Medicine Institutional Review Board. Class imbalance was handled using SMOTE sampling; TVAE (Synthetic Data Vault) generated additional synthetic samples.
Evaluations:
• Resemblance of synthetic data to real data: univariate and distance-based (Euclidean, Gower) measures.
• Utility of synthetic data: TRTR (Train/Test Real) vs TSTR (Train Synthetic/Test Real).
• Privacy of synthetic data: 1-nearest-neighbor similarity.
Relationships among resemblance, utility, and privacy are shown in Figure 1A-C.

Results: SMOTE sampled data closely resembled real data but showed greater privacy risk. Synthetic data increased privacy (Euclidean 0.02→0.12; Gower 0.03→0.09) with moderate resemblance loss. Recall improved from 0.07 (TRTR) to 0.98 (TSTR); F1score increased from 0.11 to 0.15; precision remained same (≈ 0.17-0.18). Continuous clinical features such as gestational-age distributions of synthetic data differed from real (p < 0.001) but matched SMOTE sampled data (p = 0.678). Gains were modest by F1, higher recall, showing enhanced sensitivity for rare NEC cases.

Conclusion(s): TVAE-based generative augmentation improved recall (real data: 0.07, synthetic: 0.98) and modestly improved F1 (real data: 0.11, synthetic: 0.15), balancing sensitivity and privacy. These preliminary findings support exploring NEC generation methods in addition sampling to optimize the privacy-utility-resemblance trade-off to address lack of disease-specific data and class-imbalance.

Comparative Evaluation of Synthetic Data by Resemblance, Privacy Distance, and Utility Metrics.
Figure 1A. Distribution Resemblance (Gestational Age KDE). Synthetic vs real distributions differ (p < 0.001), while synthetic vs SMOTE are similar (p = 0.678), indicating preserved shape with enhanced privacy.
Figure 1B. Privacy Fingerprint. Radar plot of Euclidean (0.02-0.12) and Gower (0.03-0.10) distances for real, SMOTE, and synthetic (TVAE) datasets. Euclidean shows stronger divergence; Gower yields smaller distances, reflecting conservative privacy estimates.
Figure 1C. Privacy-Utility Trade-off. Recall (0.07→0.98) and F1 (0.11→0.15) increase with distance to real data. Euclidean rises faster, indicating higher privacy but lower resemblance; Gower changes more smoothly, showing a balanced trade-off.