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

Restricted access to claim-level records and confirmed fraud labels limits reproducible healthcare anti-fraud research. We present a relational, longitudinal synthetic benchmark in which “Japan-calibrated” means that selected demographic, utilization, provider, coding, and reimbursement dimensions are constrained by documented Japanese public or aggregate references; it does not imply equivalence to Japanese claim microdata. Clinical trajectories, provider identities, fraud mechanisms, event rates, and claim labels are generated by design. After removing a model-visible target-proxy channel from episode semantics, we generated a new precommitted seed-202611 scenario. Model family and hyperparameters were selected by nested temporal cross-validation on development data using injected amount recovered at a 1% review budget; Platt calibration and the operating threshold were fitted on a separate pre-final interval. The frozen LightGBM pipeline was then evaluated once on 53,609 final-period claims containing 1573 designed positives. Non-interpolated average precision was 0.8403 (95% bootstrap confidence interval 0.8266–0.8545), receiver-operating-characteristic area under the curve was 0.9764, Brier score was 0.01509, and expected calibration error was 0.03375. Reviewing the top 1% selected 537 true positives with no false positives, recall 0.3414, and JPY 128,752,597 in injected amount recovered. Corrected five-seed and designed-scenario results are reported separately as diagnostic evidence because their final partitions had prior access. This is an internal synthetic confirmation, not real-claims validation; the study does not establish Japanese fraud prevalence, claim-level realism, production performance, or deployment readiness.