How we evaluate clinical AI hallucination risk

Notat’s proof story is not a magic accuracy percentage. It is a transparent methodology: extract clinical facts first, generate notes from those facts, show the raw context to the clinician, and review every output against the evidence.

Evaluation method

1. Build the reference facts

A reviewer reads the encounter material and creates a reference set of clinical facts: symptoms, negatives, medication changes, assessment, plan, safety-netting, and coding-relevant details.

2. Compare transcript-direct output

We generate a note from transcript-style input and mark unsupported statements, missing high-salience facts, incorrect attribution, and invented certainty.

3. Compare FactsContext output

We generate documentation from extracted facts and review whether each sentence is supported by a visible fact. The clinician-facing fact list is evaluated as part of the output, not hidden.

4. Record limits, not magic numbers

We do not publish fake accuracy percentages. Each evaluation note includes dataset scope, review date, known limitations, and examples of what the system still requires clinicians to verify.

What reviewers mark

Unsupported clinical assertion

Wrong medication, dose, frequency, or route

Missing red-flag negative or safety-net advice

Wrong diagnosis certainty: possible vs established

Wrong speaker attribution

Unsupported ICD-10 suggestion

Clinician-visible evidence for each key statement

Facts reusable for notes, codes, referrals, and patient instructions

Dated internal note

Current evaluation note, 2026-07-01: this page documents the method and qualitative examples used to evaluate the FactsContext architecture. It does not claim external validation or universal accuracy. The next proof milestone should be a blinded, specialty-stratified review with dataset size, reviewer agreement, and unsupported-statement rate reported openly.

Fact extraction vs transcript-direct generation

Medication change discussed twice

Transcript-direct risk

“Increase amlodipine to 10 mg and stop lisinopril.” The transcript contained a correction: the clinician first considered stopping lisinopril, then decided to continue it after reviewing renal function.

FactsContext output

Facts: amlodipine increased to 10 mg daily; lisinopril continued; renal function normal; review in 6 weeks. Note generated from those facts only.

Why it matters

Transcript-direct generation can smooth over corrections. FactsContext preserves the final decision as a discrete fact before writing.

Negative finding matters

Transcript-direct risk

“No neurological symptoms.” The actual encounter only documented no saddle anesthesia and no bladder symptoms; leg radiation was present.

FactsContext output

Facts: left leg radiation to calf; SLR positive left; no saddle anesthesia; no bladder or bowel symptoms. Note keeps the negatives specific.

Why it matters

Broad negative statements are risky. The fact list keeps the clinical context granular and reviewable.

Code suggestion requires evidence

Transcript-direct risk

Suggested J44.1 for COPD exacerbation without showing the symptom or treatment evidence.

FactsContext output

Facts: increased breathlessness, purulent sputum, prednisolone burst, antibiotics started. Suggested J44.1 with those facts as evidence.

Why it matters

The code is easier to verify because the reason for the suggestion is visible, not buried in prose.

Known limits

Clinician review remains mandatory. Notat drafts; clinicians verify and sign.

The method reduces unsupported statements by architecture, but no clinical AI should claim zero hallucinations.

Small internal evaluations are useful for engineering direction, not a substitute for external clinical validation.

Specialty, language, audio quality, speaker overlap, and local coding rules can change performance.

Read the proof, then inspect your own facts.

The evaluation method is simple because the product is designed to be inspectable: facts first, note second, clinician review always.

Try Notat free