AI
Charles Trinh — Risk Adjustment and HCC Analytics | Las Vegas, NV · Las Vegas, Nevada · Henderson · Clark County, Nevada · North Las Vegas, Nevada

Risk adjustment rewards finding and documenting the conditions a population actually has. Models that find them are useful only if the findings survive audit and reach the chart. I build both halves and measure the gap between them.
What risk adjustment actually asks
Risk adjustment converts a year of clinical documentation into Hierarchical Condition Categories and then a risk score that drives payment. The job sounds like finding conditions that were missed; the actual ask is narrower. It is finding conditions clinically present, documentable in a form that supports coding, and reviewable by a clinician who affirms the code and signs the note in the measurement year. A suspect condition never documented does nothing for RAF and less for the patient.
The failure mode most analytics teams produce is a suspect list sorted by confidence with no link to the workflow that would document the diagnosis. The downstream team receives a list, the clinicians see a list they did not ask for, and capture happens only for the cases a nurse happens to pick up. I measure the gap between the suspect list and the chart and make that gap a first-class metric of the programme, not an explanation at the end of the year.
The other failure mode is a model that over-weights historical claims. A condition coded in a prior year, not documented in the current year, looks high-confidence to a model trained on claims and is clinically uncertain for the current year. I require that the features distinguish recency and documentation status per condition, and I evaluate capture on current-year documentation specifically rather than carry-forward.
Ground truth is documented conditions, not modelled ones
The ground truth for risk adjustment is what a clinician documents after reviewing the chart. That differs from what a model would code from identical evidence, and the gap is the whole measurement. I build the set by having clinicians affirm present conditions and codes per chart, so the model is scored on the reviewer task rather than a proxy. A few hundred charts suffice, sampled to the population the model will see.
Building the ground truth requires careful stratification, closer to what I do in
ambient-scribe-evaluation than to a claims analytics sample. The set should
contain the conditions that are often missed, the conditions often incorrectly coded, and
the conditions that should not be coded. The last class matters because a model that
correctly flags a valid condition that cannot be coded as documented is not helpful, and
scoring it as a failure of detection misses the real failure of coding.
Inter-reviewer disagreement on the ground truth is normal and has to be measured. Two reviewers will disagree on whether a condition should be coded more often than a modelling team expects, especially for borderline documentation, and the model's accuracy metrics should be presented against that disagreement. Where reviewers cannot agree, the model cannot reasonably be evaluated on those cases, and I exclude them with a note rather than pretending the ground truth is clean.
Feature and model design that survives audit
A model used in risk adjustment work should explain every flag it raises. The explanation is not a generic “feature importance” plot; it is the evidence chain — the specific documented note, the specific lab value, the specific medication that makes the diagnosis plausible. I build the model with an evidence layer: every flag carries pointers to the chart evidence that produced it. The downstream reviewer can decide in a few seconds whether the flag is grounded, and the auditor can re-derive the flag from the evidence.
Transparency in the model architecture is necessary but not sufficient. A gradient-boosted or neural model can be as accurate as a rules engine and less explainable, but explainability of that kind often amounts to feature importance that no one can act on. Where the downstream workflow needs to review each flag, I often find a transparent rules or weighted-evidence design outperforms a black-box model in operational value — not in raw AUC, but in the rate at which flags are accepted by reviewers and become documented.
The scoring should be calibrated, not just ranked. A suspect score that says ninety percent likely is actionable for one reviewer workflow and a suspect score that says “upper quartile” is actionable for another. I report calibration by the score band and update it quarterly. Miscalibrated scores are a common cause of reviewer fatigue and of the “we ignore the list” behaviour that kills risk adjustment programmes.
Review routing and the documentation loop
Every flag should reach either a reviewer or a workflow that will document it. A suspect
list that reaches no one is a report nobody reads. I build routing that sends flags into the
existing documentation workflow — pre-visit review, concurrent review, annual wellness
review — rather than creating a separate queue. The routing policy is part of
human-in-the-loop-review: the rule that decides which flags reach which
reviewer is measured on accepted, rejected and unfollowed rates.
The documentation loop matters more than the model. If a condition is flagged and the clinician cannot easily document it, the flag is wasted. I work with the documentation template and the EHR to make documentation a short action — a structured field or an existing code selection — rather than a long note-writing task. The friction of documentation is a design problem and it is the reason good analytics often fails to become good RAF.
The loop also needs a feedback channel. When a reviewer rejects a flag, the reason is captured and fed back into the model as a labelled example. The rejection rate is a measure of precision that is more valuable than an offline one because it uses the reviewers, not an abstract benchmark. Reviewers become a ground-truth generator once the workflow supports it, and I build for that.
Audit, compliance and the downcoding risk
Risk adjustment is heavily audited and program-integrity review reads the documentation behind any submitted code. That constrains the model hard: it must not push unsupportable codes and must not pressure clinicians toward conditions not clinically present. I avoid incentivising language in the interface; the flag reads as a request to review the documentation, never as an instruction to add a code.
The audit trail behind every captured code should be as thorough as the trail behind any other clinical documentation. The flag, the evidence chain, the reviewer's decision, the final documentation and the coding confirmation live together and are searchable. That is the artefact a program-integrity reviewer will ask for, and on a programme I have seen it reduce the number of charts pulled during an audit because the reviewers could verify the process rather than re-derive each case.
There is a further constraint I take seriously: a model must not be trained on data whose use for this purpose is restricted. Claims and chart data have specific permitted uses, and some uses for model training are allowed and some are not under the compliance framework the organisation operates under. The data lineage supports the assertion that the training data use is permitted; I do not build without that answer given by the compliance team.
Measurement and the annual cycle
Risk adjustment runs on an annual cycle with an inter-year review, and the analytics programme should measure its effect against that cycle, not just against the model's offline performance. The measurements I track are the capture rate per condition against the ground-truth baseline, the acceptance rate of flags by reviewers, the turnaround from flag to documentation, and the year-over-year change in the conditions that are known to be under-coded in the population. These four measures connect the model to the payment outcome without pretending to sample it from a single year.
The same measures are stratified where the programme is scored by population. Conditions documented per payer, per demographic group and per provider carry different evidence. A programme that improves the mean capture rate while leaving one group unmeasured has not improved the programme, and the audit will say so. I stratify the reporting and require that small groups are reported with a confidence interval or a note about insufficient sample, rather than a missing group.
The measurement harness I build at the start of the engagement is the same one used at the end and in the next cycle. If it changes, the comparison from one year to the next is invalid. That is why the harness belongs in your repository from the start and why I prefer small, additive changes to the scoring procedure over redesigning it when new evidence arrives.
Where this connects
The engineering connects to revenue-cycle, where the payment side of risk
adjustment lives, and to clinical-documentation-ai, where the documentation
that supports any captures has to be produced faithfully. Where a programme is new, the
first deliverable is often the ground truth corpus and the workflow map, not a model, and
that is correct sequencing because a model without either is untestable.
The deliverable is a suspect generation pipeline, a review routing workflow, an evidence chain per flag and a measurement harness. Each one runs in your repository and is runnable without me. Where the vendor-supplied suspect tool exists, my first job is to evaluate it against your ground truth and report the calibration and acceptance rate, so the decision to build or buy rests on the actual numbers rather than a demo.
My work here is engineering and analytics. I do not code charts, do not sign off on clinical documentation and do not answer compliance questions about permitted data use. Those answers come from the organisation's coders, clinicians and compliance team. I build the system that makes their work measurable, and I say plainly which parts of the programme I am accountable for and which parts I am not.
Proof
- ONC (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — public suite URL
- 67 public hostnames, 51 serving, one AWS account
- 104 clearinghouse enrolments across 53 jurisdictions
- USCDI v7 proposal — Patient Digital Identity Credential — listed at isp.healthit.gov
- SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564
- AWS Business Associate Addendum active since 18 August 2026
Where this work is delivered
Engagements run from Las Vegas and reach across the valley: clients in Las Vegas, NV; Las Vegas, Nevada; Henderson; Clark County, Nevada; and North Las Vegas, Nevada — plus remote delivery nationwide.
How an engagement runs
1. Trace one real case
Follow a single patient or claim end to end with a timestamp at every hop, before any document is written.
2. Written findings
A findings list naming the defect, the system, the owner and the date.
3. Monthly review
Review the list monthly and check whether the dates held.
Frequently asked
What is this service?
This page describes charles trinh — risk adjustment and hcc analytics — the scope, the failure modes it prevents and how the work is measured. It sits in the Interoperability category at trinh.io.
Where is the work delivered?
Engagements run from Las Vegas, Nevada across the valley — Las Vegas, Henderson, North Las Vegas, Spring Valley, Paradise, Winchester, Sunrise Manor, Enterprise, Summerlin, Clark County — and remotely anywhere in the United States. On-site work in the Las Vegas valley, remote delivery elsewhere.
What does it cost?
Standard rate $195 per hour, committed $165 per hour at twenty hours or more per week for three months or more; $1,600 standard day, $1,350 committed day. Scoped in writing before work starts, fixed fee where the deliverable is defined.
Who is Charles Trinh?
A technology architect in Las Vegas, Nevada, working in healthcare identity, consent and interoperability. Not a physician, not a medical practice, and not affiliated with any other Charles Trinh.
How does an engagement start?
With one trace of a real case and a written findings list naming the defect, the system, the owner and the date. The first call is twenty minutes and free: 702-559-3489.
Engagement
Scoped in writing before work starts; fixed fee where the deliverable is defined. First call is 20 minutes and free. 702-559-3489 · charles@trinh.io
Talk to Charles
- Phone: 702-559-3489
- Email: charles@trinh.io
- Based in Las Vegas, Nevada · working remotely across the United States
Record
Profile
Healthcare technology architect working where federal regulation becomes production infrastructure. Built and operates a FHIR R4 clinical platform — SMART App Launch authorization server, patient and provider portals, SQLite — on a hardened single-host AWS estate of 67 public hostnames. The platform returned a self-test PASS on the ONC (g)(10) Standardized API test suite in April 2026, 317 of 317 subtests, publicly verifiable. Designed a consent architecture in which a permissioned ledger records patient grant, refusal and revocation as immutable events while all clinical data stays off-chain and encrypted per patient — no PHI on the ledger by design. Filed a proposed USCDI v7 data element with ONC on verifiable patient identity. Translates the Federal Register, HHS/CMS guidance and HITRUST requirements into architecture, controls and audit evidence.
Selected engineering scale
- Self-managed estate of 67 public hostnames, 51 serving, on one AWS account.
- FHIR R4 API with a SMART App Launch authorization server; patient and provider portals; an AI service layer; Hyperledger Fabric peer, orderer and CA; IPFS; a 337-table SQLite clinical store.
- Two permissioned distributed ledgers in production, one operating a consent registry that records grant, refusal and revocation with no PHI on chain.
- X12 005010 pipelines across eight transaction sets.
- Designed, built, deployed and operated end to end by one person.
Credentials and standards work
- Inferno (g)(10) self-test PASS, 317 of 317 subtests, April 2026 — the ONC Standardized API test suite. Standards validated: US Core 6.1.0, USCDI v3, SMART App Launch 2.0.0, Bulk Data 2.0.0. Public suite URL.
- USCDI v7 public comment submission, April 2026 — proposed Patient Digital Identity Credential under the Provenance data class, listed publicly on the ONC Interoperability Standards Platform.
- SAM.gov UEI MVKQW9DKKGB8 · CAGE 20KM5 · NPI 1194645564. AWS Business Associate Addendum active 18 August 2026.
Technical environment
AWS us-east-1 (compute, storage, messaging, SMS, email, voice) · Linux and systemd hardening · nginx · Fabric and EVM ledgers · FHIR R4 · X12 005010 · HL7 · SQLite · self-hosted mail, drive, search and monitoring · Git-based change control and CI release gates.