Research map / Label-efficient and robust learning / Recognition and long-tail
Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Profile2Gen, a data-centric pipeline that profiles tabular medical data into easy/hard/ambiguous subsets (via a regression-adapted DataIQ), generates synthetic data per subset with seven generative models, and removes hard synthetic samples to improve downstream regression on six OpenML medical datasets.
Problem
Synthetic tabular data for medical regression often has poor representativeness and includes hard-to-learn samples that degrade predictive performance; prior data-centric synthetic data work targets classification only.
Approach
Adapts the classification-focused profiling approach of Hansen et al. (2023) to regression by generalizing DataIQ (DataIQReg), generating synthetic data separately from easy, hard, and ambiguous subsets, then removing hard synthetic samples in a postprocessing step.
Key results
- Augmentation with Profile2Gen reduces RMSE by 4.8% relative to Traditional synthetic augmentation (Figure 5).
- In TSTR, some Profile2Gen-trained predictors reach up to 1.2% lower RMSE than real-data training.
- Profile2Gen vs. Traditional shows statistically significant differences in 79% of augmentation comparisons; Traditional failed to improve predictors in 80% of cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Profile2Gen instead of Traditional for synthetic-data augmentation | Regression predictors on medical datasets · Medical datasets, augmentation | RMSE: reducing RMSE by 4.8% | Helped |
| Train on Profile2Gen synthetic data rather than real data | Regression predictors · TSTR, some cases | RMSE: up to 1.2% lower RMSE | Helped |
| Use Profile2Gen rather than Traditional synthetic-data generation | Regression predictors across medical datasets · Augmentation | RMSE | Helped |
| Use Profile2Gen rather than PreProcess | Regression predictors across datasets · Augmentation | RMSE | Mixed |
| Add post-generation profiling and removal of hard synthetic samples to preprocessing-only generation (ablation) | Profile2Gen · Medical datasets | RMSE | Helped |
| Remove hard synthetic samples after generation (ablation) | Profile2Gen · Multiple datasets and models, augmentation | Performance variability | Helped |
| Increase the synthetic-data proportion to 68.2% (ablation) | Augmented regression predictors · Augmentation | RMSE: 68.2% | Hurt |
| Increase the synthetic-data proportion to 90.9% (ablation) | Augmented regression predictors · Augmentation | RMSE: 90.9% | Hurt |
| Use 22.7% synthetic data for augmentation (ablation) | Regression predictors · Augmentation | Model accuracy: 22.7% | No clear effect |
| Use 45.4% synthetic data for augmentation (ablation) | Regression predictors · Augmentation | RMSE: 45.4% | No clear effect |
| Change the outlier-detection threshold from ±3 to ±4 when examining removed hard samples (ablation) | Profile2Gen hard-sample removal · Cholesterol | Statistical outlier identification: ±3; ±4 | No clear effect |
Surprising findings
- PreProcess sometimes achieves lower RMSE than Profile2Gen, and in the Diabetes dataset PreProcess outperformed Profile2Gen.
- Adding synthetic data beyond about 68% of training size causes noticeable performance degradation in many models.
Limitations
- Evaluation limited to RMSE and Wasserstein distance; fairness, privacy, and clinical utility not assessed.
- Results are inconsistent across datasets and models (e.g., CTGAN variability, PreProcess outperforming in Diabetes).
- Profile2Gen produces synthetic datasets about 10% smaller due to hard-sample removal.
Experimental setup
Datasets: Parkinson Urinary Cholesterol Body Fat (fat) Plasma Levels (plasma) Diabetes
Benchmarks: Parkinson Urinary Cholesterol Body Fat Plasma Levels Diabetes
Metrics: RMSE Wasserstein distance Wilcoxon test p-value Kruskal-Wallis test
Baselines: Traditional synthetic generation PreProcess Real-data-only training
Builds on: DataIQ Hansen et al. (2023) data-centric synthetic generation CleanLab CTGAN TVAE
Closest papers by meaning
- Improving the Generation and Evaluation of Synthetic Data for Downstream Medical Causal Inference (NeurIPS 2025 · similarity 0.89)
- Debiasing Synthetic Data Generated by Deep Generative Models (NeurIPS 2024 · similarity 0.86)
- Improving the Scaling Laws of Synthetic Data with Deliberate Practice (ICML 2025 · similarity 0.86)
- Position: All Current Generative Fidelity and Diversity Metrics are Flawed (ICML 2025 · similarity 0.85)
- Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory (ICLR 2025 · similarity 0.85)
- Collapse or Thrive: Perils and Promises of Synthetic Data in a Self-Generating World (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.