AINPI 2026-05-22 update — H40 published, one confirmed strict-post-exclusion case, three SAM-NPI false positives caught
For: AINPI subscribers Date: 2026-05-22 Methodology version: 0.7.0-draft Source data: CMS Medicare Physician & Other Practitioners by Provider AND Service (RY2025, CY 2023 service year)
What's new
Two weeks ago we shipped the claims-side cross-audit (H29–H36). This week we sharpened the most actionable finding in that batch — H30a (federally-excluded NPIs billing Medicare Part B) — into the unit-of-work that State Medicaid PI offices actually write recoupment letters against: per-claim, per-HCPCS, per-place-of-service.
- H40 — granular
(NPI, HCPCS, POS)cross-audit against the CMS "by Provider AND Service" file (3.06 GB, 9.66M rows). Per-state CSVs at/api/v1/states/<state>/h40-excluded-partb-by-hcpcs.csv. - H42 — telehealth-dominant filter on H40 (CMS Telehealth Services List HCPCS codes). Null result, honestly framed.
- Confirmed case (lead finding). One federally-excluded physician — primary-source verified across LEIE + SAM.gov + NPPES — billed Medicare Part B for approximately $880,000 in CY 2023, eight years post-exclusion.
- Cohort-builder data-quality finding. Three other candidate strict-post-exclusion NPIs turned out to be SAM-NPI-join false positives on primary-source verification. The methodology is working — the primary-source URLs on every row are doing their job — but the cohort needs a name-match validation step.
The methodology version bumps to 0.7.0-draft to reflect both the new findings and the cohort-builder caveat.
H40 headline numbers
| Metric | Value |
|---|---|
| Source rows scanned | 9,660,647 |
| Distinct NPIs matched (full-window) | 194 of 6,840 |
| Distinct NPIs candidate strict-post-exclusion | 4 |
| Distinct NPIs confirmed strict-post-exclusion | 1 |
| Match rows full-window | 1,686 |
| Match rows strict-post-exclusion candidate | 57 |
| Estimated paid full-window | $27,384,846 |
| Estimated paid strict-post-exclusion candidate | $1,045,336 |
| States with ≥1 match | 41 |
The strict-post-exclusion candidate count (4) is what the automated cross-audit produces. The confirmed count (1) is what stands after primary-source verification of each candidate's LEIE and SAM records by name + NPI. We publish the candidate number with the verification status next to every row — auditors should treat AINPI as a triage tool, not a fraud determination.
H40 vs H30a: −12 NPIs (194 vs 206) is the expected delta from CMS's <11-beneficiary cell-suppression in the granular by-Service file. The aggregate by-Provider file rolls those low-volume cells up; the granular file hides them. H40 is therefore a lower bound on per-claim activity — H30a remains the more inclusive denominator-for-comparison.
The one confirmed case
NPI 1285673012 — Eduardo Siria Miranda, MD (NPPES)
| Source | Status | Detail |
|---|---|---|
| OIG LEIE | ✅ Active exclusion since 2015-06-18, never reinstated | Section 1128(a)(1) — mandatory exclusion for conviction of program-related crimes |
| SAM.gov | ✅ Active reciprocal HHS exclusion, indefinite termination | Created 2015-06-23, active date 2015-06-18 |
| NPPES | ✅ Active NPI (not deactivated), MD, Internal Medicine / Hematology & Oncology | Last updated 2025-06-24 |
| CMS Medicare Part B (CY 2023) | ✅ Billed ~$880,000 across 35 HCPCS codes | Dominated by oncology administration |
Top HCPCS codes billed in CY 2023 (eight years after federal exclusion took effect):
| HCPCS | Description | Services | Est. paid |
|---|---|---|---|
| J9271 | Pembrolizumab (Keytruda), 1 mg | 14,205 | $610,117 |
| J0897 | Denosumab (Prolia), 1 mg | 5,521 | $100,703 |
| 99214 | Established patient office visit, 30-39 min | 475 | $44,372 |
| 96413 | Chemotherapy administration, IV infusion ≤1 hr | 318 | $31,456 |
| 99213 | Established patient office visit, 20-29 min | 509 | $30,407 |
Section 1128(a)(1) of the Social Security Act mandates exclusion from all Federal health care programs upon conviction of a Medicare/Medicaid-related crime. That exclusion bars payment, billing, and ordering of services across Medicare, Medicaid, and all other Federal health care programs (42 USC § 1320a-7(a)(1); 42 CFR § 1001.1901). The published claim data is on data.cms.gov; the exclusion is on the OIG's public list; the NPI is active in NPPES.
This is a single case. It is publicly verifiable in roughly three minutes by any subscriber clicking the three URLs above.
The three SAM-NPI-join false positives
Three other candidates surfaced strict-post in the automated cross-audit:
| Cohort NPI | Cohort name | Actual LEIE | Actual SAM | Verdict |
|---|---|---|---|---|
| 1982713020 | CULLEN, EDWARD (RI) | Not in LEIE | NPI in SAM points to ORATE/CARLOS/FLORES (3 different individuals, same NPI) | SAM-NPI false positive |
| 1518952506 | KEARNEY, TIMOTHY (PA) | Not in LEIE | NPI in SAM points to BROWN, TERRI CASS | SAM-NPI false positive |
| 1376654624 | DESAI, BAKUL (NJ) | Not in LEIE | NPI in SAM points to FOLTS, JESSICA NICOLE | SAM-NPI false positive |
The SAM.gov Public Extract sometimes carries an NPI field that does not actually belong to the excluded party — clerical errors at SAM data-entry, or NPIs reused across unrelated records. The AINPI cohort builder currently treats any non-empty npi field on an active SAM row as a match without cross-checking the name against NPPES.
Fix path (separate PR): add a NPPES name-match step to the cohort builder's SAM join. If the SAM-row name and the NPPES-row name for that NPI don't share a token, downgrade the row to "needs review" rather than including it in the published critical cohort. We'll publish a corrected high-risk-cohort-export.csv and re-run downstream H29/H30a/H30b/H32/H40 as a separate update.
This finding strengthens the published audit, not weakens it: the primary-source URLs on every cohort row are exactly what caught these false positives. The published cohort is a triage tool, not a fraud determination — the methodology assumes downstream readers will verify against the linked LEIE + SAM + NPPES portals.
H42 — null hypothesis supported
Zero federally-excluded NPIs in CY 2023 show ≥80% of post-exclusion Medicare Part B services billed under telehealth-specific HCPCS codes.
The published headline is honest about two competing readings:
- Federal exclusion screening for telehealth-specific Part B is working — LEIE/SAM screening is catching telehealth-credentialed excluded providers pre-payment.
- The post-exclusion cohort billing Part B is too small for the dominant-share threshold to register at all (4 NPIs / 57 service-rows national, of which 1 is confirmed).
H42 was always intended as a sharpened sub-test, not a headline cohort. H40 remains the actionable file.
Data quality + methodology
Both items below are documented in the provenance doc landing at docs/methodology/runs/2026-05-22-h40-h41-h42-baseline.md:
- CMS cell suppression — any (NPI, HCPCS, POS) cell with <11 beneficiaries is suppressed in the by-Service file but rolled up into the by-Provider aggregate. H40's NPI count is therefore a lower bound vs H30a's. Documented per-finding.
- SAM-NPI false positives in the cohort builder — flagged above. Fix in a follow-up PR.
- LEIE-NPI join quality — every published row carries the LEIE / SAM / NPPES verification URLs so readers can do the same 3-source primary-source check we did for Miranda above. Treat every candidate as a candidate, not a confirmed case.
H41 — running but not yet complete
The third planned finding (NPPES taxonomy vs Medicare-billed-specialty drift) ran for ~4 minutes before its BigQuery NPPES query stalled mid-iteration and the process hung. We killed it and will retry; the script is committed and the methodology entry is pre-registered. Numbers in the next update.
What state Medicaid PI offices should do
Per-state CSVs are at https://ainpi.dev/api/v1/states/<state-abbrev>/h40-excluded-partb-by-hcpcs.csv. Each row is one (NPI, HCPCS, place-of-service) tuple — the unit your team writes a recoupment letter against. Every row carries the primary-source LEIE + SAM + NPPES verification URLs. Treat any single row as a candidate, not a finding, until you've verified the name match against LEIE/SAM.
For the Miranda case specifically: any Medicare Part B activity by NPI 1285673012 between June 18, 2015 and today is on the table per § 1320a-7. Texas Medicaid's PI office can pull the same file and check whether any state activity occurred on the same NPI; we have not run that cross-audit ourselves.
Links
- H40 finding: https://ainpi.dev/findings/excluded-billing-medicare-partb-by-hcpcs
- H42 finding: https://ainpi.dev/findings/excluded-telehealth-dominant-post-exclusion
- Confirmed case (LEIE): https://exclusions.oig.hhs.gov/
- Confirmed case (SAM.gov): https://sam.gov/search/?index=ex
- Confirmed case (NPPES): https://npiregistry.cms.hhs.gov/provider-view/1285673012
- Provenance + reproducibility:
docs/methodology/runs/2026-05-22-h40-h41-h42-baseline.md - Methodology: https://ainpi.dev/methodology
— AINPI, 2026-05-22