← All work

DocNexus.ai · case study

PayerLens

A Medicare Advantage panel and leakage worklist that says “here is our reasoning, check it” instead of “trust our score”.

ShippedJul – Aug 2026Data & Analytics, Full stack

A member panel, network status, out-of-network leakage, denial friction and a recruitable-provider worklist for Medicare Advantage payers, delivered as a Next.js dashboard over pre-aggregated marts.

Methods and relative results only. Client figures, internal names and scale are not published.

The asymmetry that shapes it

FinSignal has ground truth; PayerLens does not. Nobody publishes which claims were truly out-of-network or what a plan’s risk score should have been. So credibility cannot come from an accuracy score. It comes from evidence and falsifiability. That is a weaker claim, made honestly rather than dressed up as a model score.

Definitions a payer can argue with

Claims never state an enrolment span, so membership is activity-defined, disclosed on every downstream number, and shipped with a threshold-sensitivity table so a payer can argue with the definition instead of taking the panel on trust.

Plan switchers are 1.34% of the panel but about ~41% of the special-needs base, so the assignment rule swings that cohort by 68%. The shipped rule (special-needs precedence with a ≥2 active-month floor) was chosen by measurement: overlapping members average 6.40 special-needs months against 3.52 standard months, and only 11.1% have a single special-needs month.

Leading with concentration, not the rate

Out-of-network leakage is a small share of spend, and no payer buys a small-percentage finding. The defensible claim is a worklist: the providers who account for half the out-of-network spend, each traceable to its claims.

A recruitment worklist collapses if the providers on it are emergency physicians, anesthesiologists, pathologists or radiologists: EMTALA compels treatment regardless of network and the No Surprises Act bars balance billing. Measured, that risk does not materialise at the top of the list (3.3% ancillary in the top hundred) but rises steeply further down, so the worklist filters on ancillary share by default.

Decisions

  • Two corrections were made against our own numbers, both of which made the product look worse: a concentration query that grouped by provider without excluding nulls, and a provider denominator that was off by more than an order of magnitude. The corrections shipped.
  • Money comes from the remittance for every cohort on equal footing, cross-validated by a paid-to-charge ratio of 16.6% against a documented chargemaster ratio of about ~17%.
  • One month is elevated, and the natural explanation (annual contract-load lag) was tested and rejected because the same month the next year is flat. It is labelled unexplained and excluded from trend baselines. Labelling it with a cause the data contradicts would be worse than saying we do not know.
  • Risk-adjustment outputs are suspecting only: evidenced, labelled, human-reviewed, with no submission path in any code path.