DocNexus.ai · case study
In/out-of-network from payer adjudication
Stop inferring the network. Read what the payer said about it.
- 40.67%
- unknown lines recovered
- 6.54%
- UNKNOWN after recovery
- 84.5%
- naive OON that was non-covered
- 23.40%
- structural ceiling
Classifies every remittance line as in-network, out-of-network, excluded or unknown from the payer’s own adjustment codes, then resolves status at claim level.
Methods and relative results only. Client figures, internal names and scale are not published.
What was rejected
The original design derived a synthetic network: providers above a volume threshold are in-network, everything else is leakage. That is circular. Defining the network defines the leakage, and a payer would rightly reject it. The prior heuristic also mislabelled ~46% of lines.
What was built
Remittances carry claim-adjustment reason codes: the payer’s stated reason for how each line was paid. A contractual write-down (CO+45) cannot exist without an active contract; PR+242 says “not provided by a preferred network provider” in as many words. Network status becomes observed, not inferred.
A 9-rank evidence ladder runs from explicit codes (242/243) through contractual write-downs, pricing mandates, benefit gates, cost-share and coordination-of-benefits down to a math fallback and no signal. The rank maps to a verdict at query time, so a policy change never needs a rebuild.
The consequential ruling
Non-covered is not out-of-network. Non-covered and benefit-gated lines are 84.5% of what a naive reading would call out-of-network; folding them in would have inflated leakage roughly ~5×. Explicitly out-of-network lines are 0.33%.
Recovery by evidence, not imputation
Network status is a property of a contract, so it is uniform within a claim: a claim cannot be partly in-network. Uncoded lines inherit from coded siblings on the same claim, recovering 40.67% of unknown lines and cutting UNKNOWN from 11.02% to 6.54%. A single-year pilot had measured 45.65%; the full-history figure is lower and is the one reported.
Provider-level statistical imputation was sized and deliberately not shipped. With no ground truth to validate it against, it would have traded the product’s only credibility mechanism for coverage it could not defend. Genuinely contradictory claims (0.08%) surface as CONFLICT for review rather than being forced to a verdict.
Decisions
- Contract evidence is checked before coordination-of-benefits exclusions, a choice kept against reviewer advice after measuring the overlap.
- Header-level codes were not copied onto lines with no adjustments: most of those lines were paid in full, so a header code would invent an adjustment.
- Two quoted figures in the audit trail were corrected by measurement: a reversal share claimed at 3.2% is 0.015%, and a match rate was off by a factor of 20×.
Where it falls short
- A remittance exists only for adjudicated claims, so 23.40% of Medicare Advantage claims can ever reach the classifier. Of those, 85.3% get a verdict; net coverage is about ~20%. Quoting the yield as coverage would describe a much bigger product than the real one.