Memisis: Orchestrating and Evaluating Synthetic Data for Tabular Health Datasets
Researchers have developed a method to distill knowledge from large, computationally expensive tabular foundation models (TFMs) into smaller, faster models for structured health data. This technique, tested across 19 healthcare datasets, allows distilled models to retain over 90% of the original model's predictive accuracy while operating significantly faster and maintaining crucial calibration and fairness properties. The study also found that averaging predictions from multiple teachers did not consistently outperform the best single teacher, suggesting a more streamlined approach to deploying TFM-quality insights in resource-constrained health settings. Separately, a new tool called Memisis has been introduced to orchestrate and evaluate synthetic data generation for tabular health datasets, aiming to balance privacy, utility, and fairness. AI
IMPACT Distillation techniques offer a path to deploy high-performing models in resource-constrained healthcare environments, while synthetic data tools aim to improve data availability and privacy.