Coding gaps, with the evidence attached
A coding-gap list gets worked only when every line shows where it came from.
- Service
- Risk adjustment
- Status
- in production
- Built
- September and October 2026
357 → 2,154
conditions confirmed as captured for one health plan, once the bot read the practice's own claims
Source: The bot's run logs and build notes, Oct 2026
37,077
claims read in the history backfill
Source: The bot's run logs, Oct 2026
29.7 s → 3 ms
provider summary, after its results were prepared ahead of time
Source: Build notes, Oct 2026
What the list looks like
Every row carries its evidence: a plan claim, our own billing or the problem list. The rows here are invented sample data; no screen on this site is a real system.
Coding gaps · this week
Sample data
M-0412 · Type 2 diabetes with CKD
OpenPlan claims · Nephrology claim, Mar 2026
M-0097 · Chronic systolic heart failure
CapturedOur billing · Visit claim, Jun 2026
M-1183 · Major depressive disorder
OpenProblem list · Active since 2024, not billed this year
M-0655 · COPD
Needs reviewPlan claims · Reported once, nothing else on file
M-0730 · Morbid obesity
CapturedOur billing · Visit claim, Feb 2026
M-0921 · Atrial fibrillation
OpenProblem list · Cardiology note, Jan 2026
| Member | Condition | Status | Evidence |
|---|---|---|---|
| M-0412 | Type 2 diabetes with CKD | Open | Plan claimsNephrology claim, Mar 2026 |
| M-0097 | Chronic systolic heart failure | Captured | Our billingVisit claim, Jun 2026 |
| M-1183 | Major depressive disorder | Open | Problem listActive since 2024, not billed this year |
| M-0655 | COPD | Needs review | Plan claimsReported once, nothing else on file |
| M-0730 | Morbid obesity | Captured | Our billingVisit claim, Feb 2026 |
| M-0921 | Atrial fibrillation | Open | Problem listCardiology note, Jan 2026 |
3 open · 2 captured · 1 for a person to decide
Illustrative, not a real system
The problem
Which risk-adjusting conditions haven't been billed this year, and why? The health plans' claims files answer part of it, but they arrive weeks to months late and barely show the practice's own office visits. The first version of the tool matched diagnoses by approximate names, and most of its problem-list suggestions were wrong. A list like that gets ignored.
What was built
A coding-gap engine maps diagnoses from three places, the plans' claims, the practice's own billing and the EHR problem list, to the official CMS-HCC V28 categories, including the rule that a more serious condition outranks a milder related one. Each patient's categories are marked captured or open; open gaps are ranked, for example "dropped since last year" or "billed only by a specialist". Every suggestion cites its evidence.
To stop waiting on the plans, an RPA bot opens the claim forms the practice sent from the EHR every 30 minutes during work hours and records the diagnosis and procedure codes, with no names or birth dates.
Each provider's page shows, in plain words, which of their own patients have a code worth a second look, and the whole report prints in landscape for a one-on-one.
What was hard
The first fix was exact codes instead of names. The second was the practice manager checking six flagged patients by hand and finding false alarms: a more serious code already coded, a cancer that was history only, a patient who had switched plans. Each one became a rule, and anything staff rule out stays off the list for the rest of the year instead of coming back every day.
Removing finished rows had to happen on the server, not only on screen, or the headline counts and the lists would disagree.
Result
For one health plan, conditions confirmed as captured went from 357 to 2,154 once the bot read the practice's own claims. The history backfill read 37,077 claims. Preparing the provider summary ahead of time took it from 29.7 seconds to 3 milliseconds, and the all-providers list went from about 3.9 seconds to about 0.6.
What would show the business result is the share of open gaps closed by year end, validated against the plans' accepted claims. That is being tracked and is not yet claimed here.
Where the numbers come from
Build notes, the bot's run logs and database counts from September and October 2026.