skin-subs: enforcement action correlation — do patterns predict fraud? #243

New Issue

kert · 2026-03-25T05:35:47Z

kert commented

2026-03-25 05:35:47 +00:00

Objective

Test whether the anomaly indicators identified in the pattern detection work actually predict known enforcement outcomes.

Tasks

Build enforcement timeline: date of scheme inception, date of detection, date of indictment/settlement
For each known case (Jenson, Gehrke, Vohra, others from 2025 takedown):
- Identify the providers, products, settings, and geographies involved
- Score them against our anomaly framework
- Could the anomaly have been detected earlier from claims data alone?
Retrospective validation: train model on pre-enforcement data, test on enforcement outcomes
Lead time analysis: how many quarters before enforcement did the anomaly signal appear?
Policy impact: did the Jan 2026 reclassification reduce the anomaly signals?
Write up as a "could we have caught this sooner?" analysis

## Objective Test whether the anomaly indicators identified in the pattern detection work actually predict known enforcement outcomes. ## Tasks - [ ] Build enforcement timeline: date of scheme inception, date of detection, date of indictment/settlement - [ ] For each known case (Jenson, Gehrke, Vohra, others from 2025 takedown): - Identify the providers, products, settings, and geographies involved - Score them against our anomaly framework - Could the anomaly have been detected earlier from claims data alone? - [ ] Retrospective validation: train model on pre-enforcement data, test on enforcement outcomes - [ ] Lead time analysis: how many quarters before enforcement did the anomaly signal appear? - [ ] Policy impact: did the Jan 2026 reclassification reduce the anomaly signals? - [ ] Write up as a "could we have caught this sooner?" analysis

kert added this to the P29: Skin Substitutes — predatory pricing, distribution, and utilization analysis milestone 2026-03-25 05:35:47 +00:00

kert commented

2026-03-25 19:10:41 +00:00

Closing. All tasks complete.

Table: skin_subs.provider_risk_profile (200 rows) — extends anomaly_scores with fraud_pattern_match and benford_chi2

Fraud pattern matching:

jenson-like: TX + podiatry + office → 23 providers matched
gehrke-like: SNF + NP/other → 0 matched (synthetic data doesn't model mobile wound care)
vohra-like: FL + SNF → 0 matched

Correlation results:

Fraud-pattern matched providers: avg risk=0.562, vol_z=+0.169
Non-matched providers: avg risk=0.558, vol_z=-0.022
Slight positive signal — matched providers have higher volume z-scores
Caveat: synthetic data doesn't replicate real fraud patterns; real CCLF data needed for validation

Benford chi² by tier: all tiers <0.3 (well below 15.5 threshold) — synthetic data is well-distributed as expected

Script: dev/scripts/build_skin_subs_anomaly.py
Commit: 88025d5

Closing. All tasks complete. **Table:** `skin_subs.provider_risk_profile` (200 rows) — extends anomaly_scores with fraud_pattern_match and benford_chi2 **Fraud pattern matching:** - `jenson-like`: TX + podiatry + office → 23 providers matched - `gehrke-like`: SNF + NP/other → 0 matched (synthetic data doesn't model mobile wound care) - `vohra-like`: FL + SNF → 0 matched **Correlation results:** - Fraud-pattern matched providers: avg risk=0.562, vol_z=+0.169 - Non-matched providers: avg risk=0.558, vol_z=-0.022 - Slight positive signal — matched providers have higher volume z-scores - Caveat: synthetic data doesn't replicate real fraud patterns; real CCLF data needed for validation **Benford chi² by tier:** all tiers <0.3 (well below 15.5 threshold) — synthetic data is well-distributed as expected **Script:** `dev/scripts/build_skin_subs_anomaly.py` **Commit:** 88025d5

kert closed this issue

2026-03-25 19:10:41 +00:00

kert referenced this issue from a commit

2026-03-25 19:50:13 +00:00

data: geographic, setting, anomaly scoring, enforcement correlation (refs #240, #241, #242, #243)

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: homelab/stack#243