A Japan-calibrated synthetic medical claims benchmark for anti-fraud machine learning

Healthcare anti-fraud machine learning is constrained by restricted claim-level access, sparse confirmed fraud labels, privacy risk, and weak reproducibility across jurisdictions. This study presents a Japan-calibrated synthetic medical claims benchmark for anti-fraud analytics. The framework generates relational claims data from Japanese reference assets, temporal episodes of care, diagnosis-anchored procedures, drugs, Diagnosis Procedure Combination rows, provider-patient graph edges, controlled fraud injection, validation reports, and leakage-screened feature marts. The manuscript-scale run produced 50,000 synthetic patients, 7,640 providers, 150,000 episodes, 250,823 claims, 430,286 diagnosis rows, 320,999 procedure rows, 55,110 drug rows, 12,532 DPC rows, and 10,742 fraud labels across eight realized fraud mechanisms. Validation passed hard structural, billing, clinical, temporal, graph, fraud-reconciliation, privacy, feature-leakage, and manuscript-readiness gates. Machine-learning experiments evaluated linear, tree ensemble, gradient-boosting, and anomaly-detection baselines with top-k, calibration, threshold, ablation, and feature-importance analyses. HistGradientBoosting achieved the highest claim-level PR AUC (0.947), whereas LightGBM achieved the highest recall (0.919). Logistic regression recovered 108.4 million yen of injected fraud amount in the top 1% ranked claims. The benchmark is intended as a reproducible research substrate for synthetic-data fidelity, privacy screening, graph-aware feature engineering, and healthcare anti-fraud model evaluation without using real patient records or real provider fraud labels.