The Challenge

Medicare pays over $1 trillion annually. Roughly $70 billion of that, about 7 cents on every dollar, is estimated to be improper payments. The GAO has repeatedly criticized CMS for lacking quantifiable fraud detection metrics. The existing model is reactive: pay first, chase later, often years after the money is gone.

The ACT-IAC 2026 AI Hackathon posed a direct question: Can AI shift CMS from reactive to proactive program integrity?

We had 13 days.

The Constraint

The team had no protected health information, data use agreements, or claims-level data. Everything had to work on publicly available CMS datasets, the same data anyone can download from data.cms.gov.

The public data contains annual totals, not individual claims. You can see that a provider billed $2.3M last year. You cannot tell whether they billed the same patient twice on the same day. Most academic fraud detection work assumes claims-level access. Our team had to work without it.

Can AI shift CMS from reactive to proactive program integrity? We had 13 days.

What the evidence tells us

We separate what we built from what we tested and what remains unproven. Each step toward operational use needs more evidence before the Government can accept it.

  1. Rung 1 · Bounded challenge

    A working public-data prototype under a 13-day constraint

    Four people built the system without protected health information, data use agreements, or claims-level records. That demonstrates delivery speed and feasibility on public annual aggregates. Readiness for operational casework remains unproven.

  2. Rung 2 · Historical backtest

    The model concentrated known validation labels near the top

    42%
    Backtest precision@100
    890x
    Backtest lift over random
    0.965
    Validation AUC
    6.9yr
    Retrospective lead before OIG exclusion

    The team trained the model on 2013 to 2019 data and validated it on 2020 to 2023 labels. At the study's 0.047% base rate, 42 of the top 100 matched the validation label. The lead-time calculation is retrospective; it does not establish realized or guaranteed savings.

  3. Rung 3 · Public-record corroboration

    High-ranked providers matched later public enforcement records

    Researchers checked every non-LEIE provider in the top 250 against DOJ releases, OIG actions, court records, and state medical boards. Thirty-seven matched documented cases involving more than $500 million in alleged or resolved fraud. These matches support further testing of the rankings. They do not establish causation or independent fraud findings.

    Examples of high-ranked providers matched to public enforcement records
    RankPublic-record match
    #2DOJ settlement of $1.2 million
    #755,000 false claims involving deceased patients
    #8$14.3 million wound-care fraud case
    #22$250 million opioid conspiracy
    #52$6.5 million false-claims case independently investigated by ProPublica
  4. Rung 4 · Production acceptance

    The Government would define and verify the operational threshold

    Before production, test current authorized data, sample and label quality, ranking thresholds, subgroup performance, drift, and traceability. Also test security, privacy, Section 508 accessibility, investigator workflow, and how to handle false positives and false negatives. Government acceptance needs agreed evidence and human review. A hackathon award and one model metric cannot answer all these questions.

What We Built

A single Go binary that ingests 22 CMS datasets, scores 1.47 million providers, and serves an interactive investigation dashboard.

Indicium runs as a single Go binary on one server. It ingests 22 CMS datasets (548 million rows), scores 1.47 million Medicare providers across 10 years of history, and serves an interactive investigation dashboard.

You can look behind every score

Investigators can inspect the factors, SQL, weights, peer groups, and history behind a score. Authorized users can adjust weights to explore priorities. A score does not become an automatic finding.

Agree on what good looks like before production work

Before production work, we would agree with the Government on ranking quality, traceability, response time, accessibility, security, and usefulness to investigators. Acceptance would depend on evidence that the system meets those standards.

Model mechanicsHow public data, 27 features, and three scoring methods produced reviewable risk signals

The Data Pipeline

The system ingests data from CMS, OIG, DEA, Census, and other public sources directly into an embedded DuckDB database using CSV-to-columnar ingestion with SHA-256 deduplication. It requires no external database server or ETL orchestration platform.

Data pipeline scale and coverage
MetricValue
Rows ingested548M+
Providers scored1.47M
Coverage2013 to 2023
Peer groups1,046 (specialty x region x year)
Graph nodes5.8M
Graph edges82.6M

27 Features, 8 Categories

Indicium scores every provider on 27 features in eight categories:

  • Billing Volume: billing far more than specialty peers, sudden spikes, solo practitioners billing like groups.
  • Procedure Mix: upcoding E&M visits, concentrating in expensive codes, deviating from specialty norms.
  • Drug Prescribing: opioid prescribing intensity, high-risk medications for elderly, rare drug combinations.
  • Payment Network: Open Payments manufacturer relationships correlated with prescribing patterns.
  • Sanctions: LEIE exclusions, Corporate Integrity Agreements, SAM exclusions.
  • Geographic: billing deviation from county baselines, cross-payer risk signals.
  • Graph Analysis: six custom network algorithms: prescribing isolation, opioid network contagion, manufacturer concentration, exclusion proximity, geographic deviation, co-prescribing anomalies.
  • Enrichment: career stage anomalies, controlled substance supply chain deviations from DEA ARCOS data.

Three-Way Ensemble Scoring

1. Weighted Composite

27 features normalized to 0-100, combined with tunable weights. Deterministic, reproducible, transparent.

2. ML Corroboration

Isolation Forest, K-Means, Local Outlier Factor. Corroboration boost up to +30 points.

3. Gradient Boosted Trees with Focal Loss

Trained on LEIE exclusion labels. Validation AUC: 0.965.

The Dashboard

The dashboard includes a national heatmap, state and county drill-downs, and detailed provider profiles. Investigators can review peer comparison radar charts, top contributing factors, 10-year billing trajectories, and network visualizations.

Investigators can adjust feature weights in real time and see scores re-rank instantly.

Every score shows exactly how it was computed, including the SQL queries, weights, and peer groups available for inspection.

Every score exposes its SQL queries, weights, peer groups, and calculation for inspection.

A closer look at the implementation

The architecture shows how we made this prototype work and kept it manageable. Those details matter, but the Government still needs to decide whether the system meets its requirements.

Architecture detailArchitecture: one Go binary, one server, and an in-process analytical stack

DuckDB, GraphWizard, the machine-learning libraries, and the HTMX, Alpine, and ECharts interface all ran in process. The prototype therefore needed no external database or graph server at runtime.

Indicium prototype architecture choices and rationale
ComponentChoiceWhy
LanguageGoSingle binary, embedded assets, goroutines for pipeline work
DatabaseDuckDB (embedded)Columnar analytics across 548 million rows without a separate server
Graph engineGraphWizard (in process)Replaced a 100GB-RAM Memgraph dependency
FrontendHTMX + Alpine + EChartsEmbedded in the binary with no production build step
Machine learningGo librariesIsolation Forest, K-Means, LOF, and GBT in process
Deployment shapeSingle VPSBounded prototype footprint; production requirements remain to be defined

Production caveat: This footprint does not establish production authorization, a security baseline, an availability design, a recovery plan, or an independent cost estimate. Those requirements depend on the Government’s data, controls, service levels, and continuity needs.

Open-source implementationGraphWizard: 82.6 million-edge analysis in about 90 seconds with a public Go library

GraphWizard was built during and after the hackathon to replace Memgraph. Its core library provides more than 40 algorithms behind standard gonum interfaces, reports 97.3% test coverage, and is available under the MIT license. The current repository has since added disk-backed graph and loading tools that use DuckDB and SQLite.

GraphWizard implementation and benchmark metrics
MetricValue
Algorithms40+
Direct module dependenciesgonum, DuckDB, and SQLite in the current repository
Test coverage97.3%
LicenseMIT
Graph size handled82.6 million edges and 5.8 million nodes
Analysis timeAbout 90 seconds in the prototype environment
RAM dependency replaced100GB Memgraph deployment replaced with an in-process library