We introduce Pattern Intelligence, a production‑grade system that automatically uncovers systematic underpayment and denial patterns in healthcare revenue cycle management without any supervision. The key idea is to cluster claims based on their adjudication behavior rather than on claim identity. Each claim is encoded as a 130‑dimensional vector comprising:
- Expected‑vs‑actual payment ratios per line
- Claim Adjustment Reason Codes (CARC)
- Denial and payment signals
- Service composition descriptors
Payer and contract identifiers are deliberately omitted from the distance metric because they dominate the geometry and collapse clusters into trivial payer‑wise groups. After clustering, these identifiers are re‑introduced only as a post‑hoc segmentation (by contract and visit type) to keep every surfaced pattern coherent and actionable.
Clustering is performed with HDBSCAN; the noise set undergoes a second‑pass sub‑clustering to recover hidden structure, and low‑confidence memberships are gated out. A large language model (LLM) then translates each cluster centroid into an analyst‑ready natural‑language description of dominant procedures, adjustment reasons, and failure modes.
A recurring observation is that roughly one‑third to one‑half of high‑value denials form an irreducible long tail that no density‑clustering configuration or representation we tried can group—an intrinsic domain property rather than a modeling flaw. We validate cluster quality via coherence analysis and quantify impact with a recoverable‑opportunity ($) metric on two production tenants (≈10K and 157K claims) within a six‑tenant deployment.
The contribution is not a new algorithm but a problem formulation, a representation‑design insight, and empirical findings from real‑world production.
Review: Pattern Intelligence demonstrates that a well‑crafted claim embedding combined with unsupervised clustering and LLM‑based interpretation can scale to hospital claim volumes, automatically surface actionable financial anomalies, and deliver measurable revenue recovery.