DocNexus.ai · case study
Full-history claims cost estimator
Costing every medical claim line in a decade of US claims, and discovering that half of them never needed estimating.
- −44.4%
- line bias before
- −0.5%
- line bias after
- +0.1%
- DRG bias
- 49.5%
- dollars that are actuals
Assigns paid, patient-responsibility and allowed amounts, each with low/high bands, a source tag and a confidence, to every medical claim line in the canonical backbone from 2017 to 2026: all payers, all claim types, not cohort-scoped.
Methods and relative results only. Client figures, internal names and scale are not published.
The insight that reframed the build
The closed feed already discloses actual line-level paid and allowed amounts. Those lines had been going through the estimator. They now ship as actuals, and the estimator was re-fit to serve only the genuine remainder, then validated against the disclosed truth.
Before this, the line path carried a −44.4% bias that nothing in the build could have surfaced, because there was no independent actual to compare against. It is now −0.5% on the line path and +0.1% on the DRG path. The absence of a validation surface was the defect; the bias was its symptom.
Three structural changes
Disclosed closed actuals became their own scoring tier; the share of disclosed lines that landed in it went from 24.4% to 93.1%.
The claim key moved off an unstable surrogate onto a rebuilt, lineage-consistent backbone.
Every model surface is keyed by source instead of pooled, because closed and open populations price differently. Pooling lands the estimate in the middle: wrong for both.
A self-correction worth keeping
The first measurement of disclosure came from an incremental update slice and read almost complete. It was recency bias: the slice held only a sliver of closed lines. Re-measured on the full view, disclosure is 68–74% per year and flat across years, which means the remainder estimator carries a real share of dollars in every era, not a rounding error.
An independent cross-check: the disclosed-actuals tier covers 33.6% of the earliest year’s lines; a separate measurement of closed share times disclosure rate gives ≈0.34. Two measurements, no shared computation path, agreeing closely.
Decisions
- Inpatient cost is a median admission amount per payer and DRG, empirical-Bayes shrunk by n/(n+12), era-keyed and support-floored, because charge and paid barely correlate for admissions.
- Zero-charge lines use the probability of a non-zero payment times the expected payment given one; a plain median degenerates to zero.
- Rates are built from remittance aggregates, not line-to-line joins, which are unreliable at this grain.
- The deployed table diverged from its source file more than once, so the build manifest now records parameters measured from the table, not declared in code.
- A predicted fix that explained almost none of its target lines was reported as failed, not tuned until it passed.
Where it falls short
- A bottom-end charge gradient: very cheap lines need a paid-to-charge ratio near ~11.8 but share a band with lines that need ~5.2. The residual is 0.035% of the line path: a band-granularity limit, not clipping.
- The line lattice is era-blind while the DRG path is era-keyed; the open estimated path has no era correction and no truth to measure against.
- One aggregate column is correct in total and meaningless per line, and the data dictionary says so.
- The current year is provisional: immature adjudication, not a defect.