Forecasting Medicare fraud exclusions from public billing data
A model that reads one year of a provider's public Medicare billing and ranks who is likely to be excluded by the HHS Inspector General for fraud or abuse in the next three years. It is evaluated strictly forward in time on 1.2M providers, and audited for who it wrongly flags.
Audit-budget explorer
Investigators can only review a small fraction of providers. Pick a budget and see how many of the providers OIG went on to exclude in 2022–24 would have been in the review queue, based only on their 2021 billing.
Who gets wrongly flagged?
At a 1% budget, the raw model's false alarms land as often on providers serving low-income patients. Removing patient demographics (shown below at α = 0) only partly fixes this, because billing patterns act as proxies. Stratified allocation splits the review budget across patient-mix groups. Slide to see what that costs in recall.
Bars show each group's false-positive rate divided by the overall false-positive rate (1.0 = parity). "Many" = top quartile of dual-eligible (low-income) or under-65 (disabled) patient share; rural = RUCA ≥ 4. α = 0 is the raw model ranking; α = 1 reviews every patient-mix stratum at the same rate.
What the model relies on
Test average precision when each feature group is removed (3-seed LightGBM).
Does it hold up over time?
One model per test year, each trained only on exclusions published before that year.
How it works
- Data: CMS Medicare Physician & Other Practitioners (by provider and by provider × service code), Part D Prescribers, 2016–2022: ~4.5M provider-years and ~70M billing lines. Labels come from the OIG List of Excluded Individuals/Entities.
- Features (~110): volume and markup, patient mix, peer-relative z-scores within specialty × year, per-code intensity vs. peers, E/M level distribution (an upcoding signal), opioid and brand-drug prescribing, and a "does this code mix look like the declared specialty?" model.
- Evaluation: train on 2016–17 (+2019 for the final refit), test on 2021 → exclusions 2022–24. No test-period label is ever seen in training.
- Neural comparison: a set-transformer over each provider's raw billing lines, trained end-to-end and as a self-supervised feature extractor for LightGBM.
- Limits: exclusion ≠ fraud. The labels only reflect who was caught. The scores are for prioritizing human review, never for accusing anyone. No per-provider scores are published.