Pattern Intelligence: Unsupervised, LLM-Interpreted Discovery of Systematic Underpayment and Denial Patterns in Healthcare Revenue Cycle Management

AI summary

Overview

Research area: Applied machine learning for healthcare revenue cycle management (RCM), combining unsupervised density-based clustering with large language model (LLM) interpretation of cluster centroids.

Technical level: Intermediate. The paper assumes familiarity with clustering concepts (HDBSCAN, noise labels, feature embeddings) and claims-processing vocabulary (CARC codes, contractual adjustments, denial versus underpayment), but the core ideas are explained in operational rather than mathematical terms.

Scope: A deployed, multi-tenant system called Pattern Intelligence that groups healthcare claims by adjudication behaviour to surface systematic underpayment and denial patterns, evaluated on two production tenants of 157,184 and 10,318 claims.

What This Paper Is About

Hospitals lose revenue when payers deny claims, apply contractual adjustments, or pay below the contracted rate, and many of these shortfalls recur systematically rather than randomly. Today, finding which shortfalls are systematic is done by analysts reviewing claims largely one at a time, which does not scale to the hundreds of thousands of claims a hospital adjudicates. The paper describes a production system that clusters claims by how they were paid rather than who paid them, then uses an LLM to turn each cluster into a plain-language, analyst-ready pattern description.

Key Contributions

1. A problem formulation for unsupervised discovery of systematic underpayment and denial patterns in RCM claims, explicitly distinguished from clinical prediction and from supervised denial-propensity models.

2. A representation-design insight: clustering on adjudication behaviour while deliberately excluding payer and contract identity from the distance metric, because those fields dominate the geometry and collapse clusters into trivial per-payer groups. Identity is reintroduced only as post-hoc segmentation.

3. A coherence mechanism — contract-and-visit segmentation plus second-pass sub-clustering of the noise set plus confidence gating — that keeps every surfaced pattern actionable as a single contract-and-visit unit.

4. An LLM interpretation layer that converts cluster centroids into coded, CARC-referenced, plain-language pattern names and failure-mode summaries.

5. Empirical findings from a real deployment, including an honest negative result: an irreducible long tail of high-dollar, idiosyncratic denials that resisted every density-clustering configuration and representation the authors tried.

Main Findings

• Behavioral clusters are already largely coherent, but not fully: before segmentation, per-pattern homogeneity — the claim-weighted fraction of clustered claims matching their pattern's dominant (contract, visit type) pair — was 92.7% on Tenant B and 66.3% on the larger, more heterogeneous Tenant A.

• Segmentation guarantees actionability at the cost of coverage: after (contract, visit) segmentation, homogeneity is 100% by construction on both tenants, but the unclustered fraction rises from 22.9% to 45.4% on Tenant A and from 30.4% to 32.5% on Tenant B.

• An irreducible long tail exists: between roughly one-third and one-half of high-dollar denials form an idiosyncratic residue that no configuration grouped. On anomalous claims specifically, 34.8%–54.5% fall outside any pattern after segmentation.

• The tail is a property of the data, not a tuning artifact: varying the HDBSCAN minimum cluster size left the noise floor in a narrow 26.2%–32.8% band on Tenant B across candidates from 20 to 62, and essentially flat (22.8%–22.9%) on Tenant A across candidates 100 and 200.

• Continuous features are load-bearing: dropping the continuous financial and per-line behaviour blocks in favour of a categorical-only vector on Tenant B raised noise from 30.4% to 32.8% and coarsened the solution from 92 to 73 clusters.

• Recoverable opportunity is large but unevenly captured: Tenant A carries $9.63M in recoverable opportunity across 55,648 anomalous claims; Tenant B carries $1.20M across 5,912. Behavioral clustering alone captures 25.3% on Tenant A and 41.7% on Tenant B.

• Segmentation reduces captured dollars: after (contract, visit) segmentation, capture is 13.8% ($1.33M) on Tenant A and 39.0% ($469K) on Tenant B, reflecting the same dispersion of high-dollar denials documented in the coherence analysis.

• Dollars recovered are not reported: the system measures opportunity surfaced, not dollars actually recovered; downstream appeal and recovery outcomes are not tracked within the system.

Methodology in Plain English

Each claim is converted into a 130-number fingerprint describing how it was adjudicated: the ratio of expected to actual payment, whether the variance was financially meaningful, which adjustment reason codes were attached, per-line signals such as procedure and revenue code, and what mix of services the claim covered. Crucially, the payer's name and the contract are kept out of this fingerprint, because when included they dominate every distance calculation and the algorithm simply splits claims by payer instead of by problem.

The fingerprints are clustered with HDBSCAN, which infers how many groups exist rather than requiring that number in advance and labels claims with no dense neighbourhood as "noise." A second clustering pass runs over that noise set to recover structure the first pass missed. Each resulting group is then split by contract and visit type, so that any pattern shown to an analyst describes every claim inside it. Claims whose cluster membership probability falls below a threshold are pushed back into noise so a pattern's visible membership is exactly its reported membership.

Finally, each group's centre is decoded into its dominant procedure codes, adjustment codes, and line-level behaviour; the system classifies a failure mode (denial, contractual zero-allowable, or underpayment) and prompts an LLM to write a short, specific pattern name and summary grounded in those decoded signals. The paper evaluates the result two ways: internal cluster-coherence analysis, and a dollar-weighted "recoverable opportunity" measure, both on two production tenants.

Why This Matters

Impact on research: The paper contributes a transferable design principle rather than a new algorithm — put behaviour in the distance metric and identity in post-hoc segmentation — which applies to any administrative-data clustering problem where a dominant categorical field would otherwise capture the geometry. It also models how to report a negative result honestly in an applied setting: the long tail is documented, quantified across parameter sweeps, and framed as a domain property rather than hidden or tuned away.

Real-world applications:

• Prioritising which denied or underpaid claims an RCM analyst should appeal first, ranked by dollar-dense coherent patterns rather than by claim volume.
• Informing contract renegotiation by showing which contracts systematically reimburse below the expected rate for specific service lines.
• Identifying coding or billing-process fixes, where a recurring adjustment reason code points to a correctable upstream error.
• Triaging across multiple hospital systems, since the system is described as deployed across six hospital tenants processing roughly 10K–157K claims each.

Industry relevance: The work targets the operational and financial layer of healthcare, which the authors note is comparatively under-studied relative to clinical prediction. It reframes the success criterion from exhaustive coverage to dollar-weighted triage — a framing hospital finance teams can act on directly, and one that sets realistic expectations about what unsupervised discovery can and cannot surface.

Future Directions

• Ground-truth validation. The authors explicitly call for controlled fault-injection recovery tests and blinded analyst adjudication, which their current indirect validity evidence (coherence and dollar-weighted outcomes) does not provide.

• Tracking actual recovery. The system reports opportunity surfaced, not dollars recovered; closing that loop would let the dollar metric be validated against real appeal outcomes.

• Generalisation beyond the current data. Results come from proprietary claims across a limited set of tenants and one vendor pipeline, and generalisation to other providers and payers is described as untested.

• Attacking or accepting the long tail. The paper argues the idiosyncratic residue should be triaged rather than exhaustively clustered, but leaves open whether any complementary approach — for instance, case-level retrieval rather than density clustering — could make high-dollar one-off denials actionable.

Target Audience

Applied machine learning engineers and data scientists working on administrative or operational data, particularly those designing unsupervised systems where a dominant categorical field threatens to overwhelm the feature space. It is also relevant to healthcare RCM analysts, revenue integrity teams, and health-system finance leaders who want to understand what pattern discovery can realistically deliver. Researchers interested in LLM-based cluster interpretation and in honest reporting of negative results in deployed systems will find the evaluation framing useful. A basic grounding in clustering is helpful but not essential.

Authors’ abstract

Healthcare providers lose substantial revenue to systematic payer underpayments and claim denials, yet identifying the recurring, correctable patterns behind them remains largely manual and does not scale to hospital claim volumes. We present Pattern Intelligence, a deployed system that discovers such patterns without supervision by clustering claims on their adjudication behaviour rather than their identity. Each claim is embedded in a 130-dimensional vector combining per-line expected-versus-actual payment ratios, claim adjustment reason (CARC) codes, denial and payment signals, and service composition. Payer and contract identity are deliberately excluded from the distance metric - we show they dominate the geometry and collapse clusters into trivial per-payer groups - and are reintroduced only as a post-hoc segmentation (by contract and visit type) that keeps every surfaced pattern coherent and actionable. We cluster with HDBSCAN, recover structure from the noise set via second-pass sub-clustering, and gate low-confidence memberships. A large language model translates each centroid into an analyst-ready description of dominant procedures, adjustment reasons, and failure mode. We report a recurring finding: roughly one-third to one-half of high-dollar denials, depending on tenant and segmentation, form an irreducible long tail that no density-clustering configuration or representation we tried groups - a domain property, we argue, rather than a modeling deficiency. We evaluate validity through cluster-coherence analysis and quantify impact with a recoverable-opportunity ($) metric on two production tenants (approx. 10K and 157K claims) within a six-tenant deployment. Rather than a new algorithm, the contribution is a problem formulation, a representation-design insight, and empirical findings from production.