Sizing Bundled-Payment Opportunity Across 3,000 Hospitals from a 5% Claims Sample

Challenge

The company needed to identify and win hospital partners by sizing each hospital’s financial opportunity across five surgical episode categories, then convert that analysis into a self-serve sales engine its own team could run. The inputs made this hard: Medicare LDS claims provide only a 5% Part B sample, CMS does not publish hospital-specific target prices, and the payment-standardization rules are intricate. The sizing had to be defensible enough to put in front of a hospital CFO and consistent enough to rank thousands of facilities on the same basis.

Approach

We built a multi-stage, deterministic analytics pipeline over roughly 7 million Medicare episodes in DuckDB, moving from raw claims to a ranked opportunity per hospital. AI was confined to build time (reading CMS methodology documents, payment-standardization specifications, and program rules, and turning them into precise, testable specifications). The decisions themselves run as deterministic, versioned code: standardization arithmetic, RVU lookups, percentile benchmarks, and the opportunity score are tables and rules, not model judgment, so any number can be traced back to its inputs.

• Implemented CMS bilateral payment standardization to the published ResDAC v.14 methodology, so episode dollars are comparable across geographies.
• Estimated Part B spend from the 5% sample using RVU-based scaling rather than treating the partial sample as complete, and labeled those figures as estimates.
• Benchmarked target prices against peer-cohort percentiles, giving each hospital a defensible reference point in the absence of published CMS targets.
• Scored opportunity through an internal care-pathway cost model layered on the standardized episode economics.
• Validated the approach against external anchors: ACO double-dip attribution was checked against CMS MSSP public-use files.
• Extended the same deterministic foundation from the 30-day TEAMs window to a 90-day CJR-X window, including two-sided risk caps, without rebuilding the core.

Impact

• Phase 1 opportunity-ranking shipped and was signed off by the client.
• Produced a rank-ordered list covering about 3,000 hospitals, of which roughly 335 fall under the mandatory program, giving the sales team a prioritized target set.
• Delivered a production cloud analytics platform running in the client’s own AWS account (load balancer to Fargate to Aurora Serverless, with CI/CD), so the client owns and operates the asset.
• Built a JSON-driven, self-serve prospect-deck generator so the client’s team can produce hospital-specific pitch decks without engineering involvement.
• Extended the data foundation to a second bundled-payment program and to ACO overlap, sizing tens of millions in annual post-acute-care opportunity per flagship system. These opportunity figures are modeled from claims-derived episode economics and peer benchmarks, not realized savings.

Capabilities demonstrated

• CMS claims engineering at scale: episode construction and bilateral payment standardization (ResDAC v.14) over roughly 7M Medicare episodes.
• Sound statistical handling of partial data, including RVU-based Part B estimation from a 5% sample with estimates labeled as such.
• Peer-cohort percentile benchmarking to stand in for unpublished CMS target prices.
• Deterministic, auditable decision logic: standardization, scoring, and risk caps run as versioned code and tables, with AI limited to extraction and specification.
• Client-owned production delivery on AWS (Fargate, Aurora Serverless, CI/CD) plus a self-serve, JSON-driven deck generator.

Anonymized by design: client names stay off the narrative per our reference policy. Figures that are modeled, small-sample, or targets are identified as such above.