Use the standardized SynthEHRella framework: 80/20 split of source by HASTA_ID, train syntha on the 80%, generate same-size synthetic, train HTN-risk model on each, score on the held-out 20% real. Report ROC-AUC, Brier, calibration plot. Goal: TSTR within 0.02 ROC-AUC of TRTR. Output goes to benchmarks/results/.json regenerated each release.
Use the standardized SynthEHRella framework: 80/20 split of source by HASTA_ID, train syntha on the 80%, generate same-size synthetic, train HTN-risk model on each, score on the held-out 20% real. Report ROC-AUC, Brier, calibration plot. Goal: TSTR within 0.02 ROC-AUC of TRTR. Output goes to benchmarks/results/.json regenerated each release.