Skip to content

Healthcare / Health Plans · Risk Adjustment Operationshealth plan

Suspect Condition Identification & Pursuit Decisions

EnhancesShifting
Near-term
Proven but early — expect one to three years to mainstream.

Readiness: Now Deployable with established commercial tools today · Near-term Proven but early — expect one to three years to mainstream · Emerging Demonstrated, not yet production-mainstream

Readiness reflects an editorial assessment against a published rubric as of August 2026 — an observation about current tool maturity and adoption, not a prediction about specific products or timelines.

Trajectories describe the observable direction of human effort — not a prediction about specific roles, headcount, or individual careers.

What You Do Today

Every payment year starts empty. A chronic condition documented last year earns nothing this year unless a clinician sees the member again, evaluates it again, and documents it again. So you run two lists that behave differently: conditions coded in a prior year that have not appeared yet this year, and conditions never coded at all that the data hints at — a member filling a diabetes medication with no diabetes on this year's claims, a laboratory result a clinician would have acted on, a referral pattern that only makes sense if something is being treated. Then comes the part that is actually your job: deciding which suspects are worth pursuing. Each one spends something. A line on a gap list handed to a practice that already resents your gap lists. An outreach call to a member who did not ask for one. A chart request, or a home visit. And a bad suspect costs more than nothing, because you are telling a physician their patient may have a condition, and if it gets documented because you asked rather than because they found it, you have manufactured the exact record an auditor will read back to you. All of it runs against a calendar: the encounter has to happen, be documented, be coded, be submitted, and clear editing before the year's submission deadline.

AI Technologies

Roles Involved

Who works on this
Director of OperationsCoding ManagerMedical Coder
DirectorManager/SupervisorIndividual Contributor

How It Works

Models read the structured history the plan already holds — claims across settings, pharmacy fills, laboratory results, prior-year submitted diagnoses, durable medical equipment, utilization patterns — and propose conditions the data implies but the current year's coded record does not carry. Clinical NLP extends the same search into unstructured material the plan can already see: records retrieved in earlier chart review cycles, documentation attached to authorization requests, assessment write-ups, discharge summaries. That is where a condition is usually described in a sentence and never coded. Recapture modeling handles the separate and larger population of conditions coded in a prior year, ranking by whether the member still appears to be under care for it and whether an encounter that could capture it is already on the schedule. Ranking then orders everything by how likely a suspect is to be confirmed and how much work confirming it would take, which is what decides the channel: a short pre-visit summary a clinician will actually read before the appointment, a bulk file a delegated group loads into its own system, a member outreach queue, a retrospective chart request, or nothing at all. The routing decision is the output that matters, because the same suspect is cheap through one channel and expensive through another.

What Changes

Suspects come from the record as well as from rule lists over claims, so conditions described in a note and never coded become findable. Recapture and net-new stop being one undifferentiated queue. Gap lists can get shorter and more specific, because low-confidence suspects can be held back rather than sent to fill the list. And because ranking accounts for effort, the expensive channels get reserved for the suspects that need them.

What Stays the Same

A suspect is a question, not a diagnosis. Only the treating clinician can decide whether the member has the condition, and the record has to show that they evaluated it — a model's confidence is not clinical evidence and never becomes documentation. How the question is asked is the whole ethical line: prompting a physician to consider something is a different act from telling them what to code, and the difference lives in the wording of the gap list, in the incentive attached to it, and in whether the programme measures itself only on suspects confirmed. A programme scored in one direction applies pressure in one direction, and that asymmetry is what turns risk adjustment into a False Claims Act matter. Somebody has to own the fact that the right confirmation rate is well short of everything you sent, and defend that gap when it is read as lost revenue. Where a diagnosis is allowed to come from is regulated, not optional: the encounter has to be face-to-face with an acceptable provider type in an acceptable setting, and a condition surfaced by an assessment that led to no evaluation or treatment is thin support no matter how good the suspect looked. Model bias runs the wrong way here — suspects are found where data exists, so a member with many encounters generates many suspects while a member who avoids care looks healthy, which is the opposite of who needs finding. And the negotiation with a practice over how much of its attention you get to spend this year is a relationship, held by people who have to go back next year.

Evidence & Sources

  • CMS Medicare Advantage risk adjustment program and CMS-HCC model
  • CMS Advance Notice and Rate Announcement for Medicare Advantage capitation rates and risk adjustment methodology
  • ICD-10-CM Official Guidelines for Coding and Reporting
  • AHIMA Standards of Ethical Coding
  • HHS Office of Inspector General audits of Medicare Advantage risk adjustment

Sources listed are directional references, not formal citations. Verify against primary sources before using in business cases or presentations.

Last reviewed: August 2026

What To Do Next

This section won't tell you what your numbers should be. It will show you how to find them yourself. Every instruction below produces a real, verifiable result in your organization. No benchmarks, no projections — just the steps to build your own evidence.

1

Establish Your Baseline

Know where you are before you move

Before adopting AI tools for suspect condition identification & pursuit decisions, document your current state in utilization management.

Map your current process: Document how suspect condition identification & pursuit decisions works today — who does what, how long each step takes, and where the bottlenecks are. Use your risk adjustment platform data to establish a factual baseline.
Identify the judgment calls: A suspected condition is a hypothesis for a clinician to confirm or reject, never a diagnosis, and a programme that only ever adds codes is not a compliance programme. — these are the boundaries AI won't cross. Know them before you start.
Check your data readiness: AI tools for utilization management need clean, accessible data. Check whether your risk adjustment platform has the historical data, integrations, and quality to support ML Predicted LOS tools.

Without a baseline, you can't tell whether AI actually improved suspect condition identification & pursuit decisions or just changed who does it.

2

Define Your Measures

What to track and how to calculate it

patient outcomes

How to calculate

Measure patient outcomes for suspect condition identification & pursuit decisions before and after AI adoption. Pull from your risk adjustment platform.

Why it matters

This is the most direct indicator of whether AI is adding value to utilization management.

clinical documentation quality

How to calculate

Track clinical documentation quality using the same methodology you use today. Don't change how you measure just because you changed how you work.

Why it matters

Speed without quality is just faster mistakes. Measure both together.

When to check: Check after 30 days of consistent use, then quarterly.
The commitment: Give new tools at least 30 days before judging. The first week is always awkward.
What NOT to measure: Don't measure AI adoption rate as a goal. Measure outcomes. If the tool helps with suspect condition identification & pursuit decisions, people will use it.
3

Start These Conversations

Who to talk to and what to ask

CMO or VP Clinical Operations

What's our plan for AI in utilization management? Are we piloting, planning, or waiting?

This tells you whether to experiment quietly or push for formal investment in suspect condition identification & pursuit decisions.

your risk adjustment platform administrator or vendor

What AI capabilities exist in our current EHR system that we're not using? Most platforms are adding AI features faster than teams adopt them.

The cheapest AI adoption is the features already included in your existing license.

a practitioner in utilization management at another organization

Have you deployed AI for suspect condition identification & pursuit decisions? What worked, what didn't, and what would you do differently?

Peer experience is more useful than vendor demos. Find someone who has actually done this.

4

Check Your Prerequisites

Confirm readiness before you invest

Check items as you confirm them.

More in Risk Adjustment Operations