Can Synthetic Data Replace Real Data? Evaluating the Efficacy, Limits, and Label Trustworthiness of Generative Data in Supervised Learning

The scarcity of high-quality, human-annotated training data has made synthetic data generation an increasingly important component of supervised machine learning pipelines, yet the conditions under which synthetic data can safely substitute for real data remain poorly understood, particularly in tabular healthcare domains where privacy constraints restrict data access. This study presents an empirical and theoretical investigation into the efficacy and limits of CTGAN-generated synthetic data using a real-world medical appointment dataset of 110,527 records with a 4:1 class imbalance. Applying the Train-Synthetic-Test-Real (TSTR) protocol across five supervised classifiers and six synthetic-to-real mixing ratios, we find that AUC-ROC degrades consistently as synthetic proportion increases, with all classifiers entering a critical degradation zone beyond 60% synthetic data. Tree-based architectures exhibit average AUC-ROC reductions of 8.72%, while parametric models show 10.55% reduction at 100% synthetic, both measured against real-data baselines. Beyond distributional mismatch, we identify and formally define synthetic label corruption-a distinct failure mode in which automatically assigned synthetic labels carry an independent error distribution not attributable to generative distributional drift-and demonstrate that 96.6% of these corrupted labels reflect a systematic minority-class overgeneration bias in CTGAN's conditional sampling mechanism. To address this, we propose the Synthetic Label Trustworthiness Estimation (SLTE) framework, a pre-training pipeline comprising Label Confidence Score computation via a real-data anchor model, topology-aware k-nearest neighbour label correction, and uncertainty-stratified sample routing. SLTE reduces tree-based AUC degradation from 8.72% to 3.84% and parametric degradation from 10.55% to 7.94%, recovering 66.6% of Gradient Boosting's lost discriminative performance at full synthetic substitution, offering a principled foundation for label-aware data augmentation in privacy-critical supervised learning environments.