Competition: Medicaid Provider Fraud Signal Detection Engine — 1000N Prize

Competition: Medicaid Provider Fraud Signal Detection Engine

OBJECTIVE

Build a CLI tool that:

1. Ingests the HHS Medicaid Provider Spending dataset (2.9GB parquet, 227M rows)
2. Cross-references providers against the OIG LEIE exclusion list and the NPPES NPI registry
3. Outputs a structured JSON file containing provider-level fraud signal reports usable by qui tam / FCA lawyers

REQUIRED INPUT DATA (3 files, all free public downloads)

File 1: HHS Medicaid Provider Spending

• Download: https://stopendataprod.blob.core.windows.net/datasets/medicaid-provider-spending/2026-02-09/medicaid-provider-spending.parquet (2.9 GB)
• Columns (7 total):

• Coverage: January 2018 through December 2024
• Rows with fewer than 12 claims are excluded by HHS

File 2: OIG LEIE Exclusion List

• Download: https://oig.hhs.gov/exclusions/downloadables/UPDATED.csv
• Columns (18 total):

• Note: Many excluded providers do not have NPIs. Match on NPI where available, fall back to name+state matching.

File 3: NPPES NPI Registry

• Download: https://download.cms.gov/nppes/NPPES_Data_Dissemination_February_2026_V2.zip (~1 GB zipped)
• Required columns (from 329 available):

REQUIRED FRAUD SIGNALS (6 total, each must be implemented)

Each signal has a precise definition. The tool must flag a provider if ANY of these conditions are true:

Signal 1: Excluded Provider Still Billing

• Definition: A BILLING_PROVIDER_NPI_NUM or SERVICING_PROVIDER_NPI_NUM in the spending data matches an NPI in the LEIE where EXCLDATE is before the CLAIM_FROM_MONTH and REINDATE is empty or after CLAIM_FROM_MONTH
• Output: The NPI, exclusion date, exclusion type, and total dollars paid after exclusion date

Signal 2: Billing Volume Outlier

• Definition: For each BILLING_PROVIDER_NPI_NUM, compute total TOTAL_PAID across all months. Join to NPPES to get Healthcare Provider Taxonomy Code_1. Group all providers by taxonomy code and state. Flag any provider whose total paid is above the 99th percentile of their taxonomy+state peer group
• Output: The NPI, their total paid, the peer group median, the peer group 99th percentile, and the ratio (provider_total / peer_median)

Signal 3: Rapid Billing Escalation (New Entity)

• Definition: Join spending data to NPPES on NPI. Get Provider Enumeration Date. For providers enumerated within 24 months before their first CLAIM_FROM_MONTH in the spending data: compute month-over-month TOTAL_PAID growth rate for their first 12 months of billing. Flag if any rolling 3-month average growth rate exceeds 200%
• Output: The NPI, enumeration date, first billing month, monthly paid amounts for first 12 months, the peak 3-month growth rate

Signal 4: Workforce Impossibility

• Definition: For each BILLING_PROVIDER_NPI_NUM where NPPES Entity Type Code = 2 (organization), compute max TOTAL_CLAIMS in any single month. Divide by 22 (working days) then by 8 (hours). If the implied claims-per-hour exceeds 6 (i.e., one claim every 10 minutes sustained for every working hour of every working day), flag the provider
• Output: The NPI, the peak month, the peak claims count, the implied claims-per-provider-hour, the total paid in that month

Signal 5: Shared Authorized Official Across Multiple NPIs

• Definition: From NPPES, group all NPIs by (Authorized Official Last Name + Authorized Official First Name). For any authorized official controlling 5 or more NPIs: sum TOTAL_PAID across all those NPIs from the spending data. Flag if the combined total exceeds $1,000,000
• Output: The authorized official name, list of all NPIs they control, total paid per NPI, combined total paid

Signal 6: Geographic Implausibility

• Definition: For each BILLING_PROVIDER_NPI_NUM, get their state from NPPES (Provider Business Practice Location Address State Name). Check if they have TOTAL_CLAIMS > 100 in any single month for HCPCS codes in the home health range (G0151-G0162, G0299-G0300, S9122-S9124, T1019-T1022). If so, compute their TOTAL_UNIQUE_BENEFICIARIES / TOTAL_CLAIMS ratio for those codes. Flag if ratio is below 0.1 (i.e., fewer than 1 unique patient per 10 claims -- indicating repeated billing on same patients)
• Output: The NPI, state, the flagged HCPCS codes, the month, claims count, unique beneficiaries, the ratio

REQUIRED OUTPUT FORMAT

The tool must produce a single file: fraud_signals.json

{
  "generated_at": "2026-02-18T12:00:00Z",
  "tool_version": "string",
  "total_providers_scanned": 0,
  "total_providers_flagged": 0,
  "signal_counts": {
    "excluded_provider": 0,
    "billing_outlier": 0,
    "rapid_escalation": 0,
    "workforce_impossibility": 0,
    "shared_official": 0,
    "geographic_implausibility": 0
  },
  "flagged_providers": [
    {
      "npi": "1234567890",
      "provider_name": "string (from NPPES)",
      "entity_type": "individual|organization",
      "taxonomy_code": "string",
      "state": "XX",
      "enumeration_date": "YYYY-MM-DD",
      "total_paid_all_time": 0.00,
      "total_claims_all_time": 0,
      "total_unique_beneficiaries_all_time": 0,
      "signals": [
        {
          "signal_type": "excluded_provider|billing_outlier|rapid_escalation|workforce_impossibility|shared_official|geographic_implausibility",
          "severity": "critical|high|medium",
          "evidence": {}
        }
      ],
      "estimated_overpayment_usd": 0.00,
      "fca_relevance": {
        "claim_type": "string describing the FCA violation pattern",
        "statute_reference": "31 U.S.C. section 3729(a)(1)(A|B|C|G)",
        "suggested_next_steps": ["string"]
      }
    }
  ]
}

Severity rules (no discretion):

• critical: Signal 1 (excluded provider) always critical
• high: Signal 2 with ratio > 5x peer median, Signal 3 with growth > 500%, Signal 4, Signal 5 with combined > $5M
• medium: All other flags

estimated_overpayment_usd calculation:

• Signal 1: total paid after exclusion date
• Signal 2: (provider_total - peer_99th_percentile), floored at 0
• Signal 3: total paid in months where growth exceeded 200%
• Signal 4: (peak_month_claims - (6 * 8 * 22)) * (peak_month_total_paid / peak_month_claims), floored at 0
• Signal 5: not estimated (set to 0)
• Signal 6: not estimated (set to 0)

statute_reference mapping:

• Signal 1: 31 U.S.C. section 3729(a)(1)(A) (presenting false claims -- excluded provider cannot bill)
• Signal 2: 31 U.S.C. section 3729(a)(1)(A) (potential overbilling)
• Signal 3: 31 U.S.C. section 3729(a)(1)(A) (potential bust-out scheme)
• Signal 4: 31 U.S.C. section 3729(a)(1)(B) (false records -- impossible volume implies fabricated claims)
• Signal 5: 31 U.S.C. section 3729(a)(1)(C) (conspiracy -- coordinated billing through shell entities)
• Signal 6: 31 U.S.C. section 3729(a)(1)(G) (reverse false claims -- repeated billing on same patients)

HARDWARE REQUIREMENTS

Primary target: Single Linux machine with <=200GB RAM, <=1 GPU (H100/H200 class)

Required fallback: Must also run on a MacBook (Apple Silicon, 16GB+ RAM) with no GPU. Performance will be slower but the tool must complete without error.

Constraints:

• No distributed computing frameworks required (no Spark cluster, no multi-node Dask)
• No proprietary cloud services required (no BigQuery, no Snowflake, no Databricks)
• GPU is optional -- if the solution uses GPU (e.g., RAPIDS/cuDF), it must provide a CPU-only fallback via a --no-gpu flag that works on both Linux and macOS
• All dependencies pip-installable from public repositories (no conda-only packages)
• Must work on both Ubuntu 22.04+ and macOS 14+ with Python 3.11+

Performance criteria:

• Linux (200GB RAM, GPU available): completes full run in under 30 minutes
• Linux (64GB RAM, no GPU): completes full run in under 60 minutes
• MacBook (16GB RAM, Apple Silicon, no GPU): completes full run in under 4 hours

REQUIRED DELIVERABLE STRUCTURE

submission/
  README.md              -- Setup + run instructions
  requirements.txt       -- Python dependencies (pip installable)
  setup.sh               -- One command: downloads data, installs deps
  run.sh                 -- One command: produces fraud_signals.json
  src/
    ingest.py            -- Data loading + joining
    signals.py           -- All 6 signal implementations
    output.py            -- JSON report generation
  tests/
    test_signals.py      -- Unit tests for each signal with synthetic data
    fixtures/            -- Small synthetic test datasets
  fraud_signals.json     -- Sample output from actual data run

JUDGING CRITERIA (scored by AI judge, all binary or numeric -- 100 points total)

Functional (60 points)

• setup.sh completes without error on Ubuntu 22.04 with Python 3.11+ (5 pts)
• setup.sh completes without error on macOS 14+ Apple Silicon with Python 3.11+ (5 pts)
• run.sh produces fraud_signals.json that validates against the schema above (10 pts)
• Signal 1 implemented and produces >=1 result (verified against LEIE data) (10 pts)
• Signal 2 implemented and produces results with correct percentile math (5 pts)
• Signal 3 implemented and produces results with correct growth rate math (5 pts)
• Signal 4 implemented and produces results with correct threshold math (5 pts)
• Signal 5 implemented and produces results from NPPES authorized official data (5 pts)
• Signal 6 implemented and produces results for home health HCPCS codes (5 pts)
• All estimated_overpayment_usd values follow the formulas above (5 pts)

Testing (15 points)

• pytest tests/ passes with >=6 tests (one per signal) (10 pts)
• Test fixtures contain synthetic data that triggers each signal (5 pts)

Legal Usability (15 points)

• Every flagged provider has all required JSON fields populated (non-null) (5 pts)
• statute_reference correctly maps per the table above (5 pts)
• suggested_next_steps contains >=2 specific steps per flag (e.g., "Verify provider at [address] is operational", "Request claims detail for HCPCS [code] for [month]") (5 pts)

Code Quality (10 points)

• No hardcoded file paths (all paths relative or configurable) (3 pts)
• Handles missing/null NPI values in LEIE without crashing (3 pts)
• Completes full run in under 60 minutes with 64GB RAM, no GPU (4 pts)

WHAT IS NOT SCORED (to prevent ambiguity):

• Code style, comments, docstrings -- not scored
• Number of providers flagged -- not scored (quality of signals matters, not volume)
• UI, visualization, dashboards -- not scored
• Support for formats other than Parquet -- not scored