HarnessHealthStart a pilot
For clinicians · AI-suggested diagnoses

The AI suggested it. What do the records say?

Paste the diagnoses a risk-adjustment or coding tool suggested, and open the patient's FHIR records. For each one you see what the records hold for it and against it, every piece opening to its record. Then you accept, reject or ask for more, in your own words. Fixed rules, no AI model. Nothing leaves your browser.

In a crisis, call or text 988. In immediate danger, call 911.

1. Load the patient's records

A FHIR R4 Bundle or NDJSON: an EHR export, a patient portal export, or Medicare's Blue Button (its claims are checked too). Read on this device; not uploaded, stored or sent.

2. Paste the suggested diagnoses

What the rules check

For every suggestion: the problem list (active, or marked resolved, refuted or entered in error) and diagnoses on claims, matched by ICD-10 category. For seven families it also looks at active medicines often used for the condition (never proof: most treat more than one thing) and results against a published definition, named on the page: HbA1c at or over 6.5% (ADA); eGFR under 60, or albumin-to-creatinine ratio at or over 30, on two results at least 90 days apart (KDIGO 2024); BNP or NT-proBNP the lab flagged high; post-bronchodilator FEV1/FVC under 0.70 (GOLD 2025); BMI at or over 30, or 40 for class 3 (CDC); PHQ-9 at or over 10 (Kroenke 2001). A family-file self-check never counts as a result. When nothing in favour is dated this year, it says so.

It never suggests a diagnosis and never says a code is right, billable or enough for an audit. What the records hold is not the same as what was examined and addressed at the visit; that is yours to document. For the whole chart before a visit, use the pre-visit brief.

Not medical, coding or legal advice. The check lists what the records say, including their errors and gaps.