Stanford researchers report that augmenting healthcare AI training with synthetic electronic health records (EHRs) reduced demographic bias by 47% while keeping diagnostic accuracy high. The result positions synthetic data as a practical lever for fairness work without expanding access to regulated patient records.
Stanford: synthetic EHR augmentation reduces demographic bias 47% without sacrificing accuracy
Researchers at Stanford University published results showing that synthetic electronic health records can materially reduce demographic bias in healthcare diagnostic AI models. In the study, adding synthetic EHRs to training data cut measured demographic bias by 47% while maintaining high accuracy across diagnostic tasks.
The work frames synthetic data as a way to improve representation for underserved populations that are often under-sampled in real-world clinical datasets—without increasing exposure to sensitive patient information. For teams constrained by privacy, consent, or access limitations, the study argues that synthetic augmentation can improve fairness outcomes without requiring broader access to raw EHRs.
- Fairness work gets a concrete knob to turn: A quantified bias reduction (47%) gives ML leads something testable to replicate—synthetic augmentation becomes a measurable intervention, not just a governance talking point.
- Lower privacy and compliance surface area: If similar results hold in production settings, teams may be able to iterate on model development with less handling of regulated patient data, reducing exposure risk while still improving model behavior across demographics.
- Better coverage of underserved groups: Synthetic generation can be used to target gaps in representation (e.g., demographics missing from training data), potentially improving performance parity without waiting for new clinical data collection.
- Procurement and validation implications: Buyers should ask vendors how synthetic EHRs are generated and validated, and whether bias metrics and accuracy trade-offs are reported consistently across subgroups.