AI medical scribes in Arabic: what actually matters

Why Arabic is harder than English here

Four properties of Arabic clinical speech each defeat a system built for English, and they compound rather than add.

Diglossia. The Arabic a clinician writes is not the Arabic a clinician speaks. A model trained on written Modern Standard Arabic — news, documents, books, which is where most Arabic text data comes from — meets a consultation conducted in regional speech and performs far worse than its published benchmarks suggest.

Code-switching. The sentence is Arabic; the drug is English. So is the procedure, frequently the anatomy, and almost always the abbreviation. A system that has to decide on one language per utterance will be wrong several times per consultation, and wrong in exactly the places that matter clinically.

Diacritics are omitted. Written Arabic normally leaves out the short-vowel marks, so words that are distinct in speech are identical on the page. Resolving them takes context, and in a clinical note the context is medical rather than general.

Direction and rendering. The note is right-to-left, the embedded English is left-to-right, and the numerals may be either. A note that is correct as data and mangled as a document is still a note nobody will sign.

What to test before you buy

A demonstration proves that a vendor can produce a good note from a recording they chose. It proves nothing about yours. Six tests, in the order that eliminates fastest:

1. Your own audio, unprepared. Record real consultations in your own rooms, with your own microphones and your own background noise, and have the vendor run those. A scribe evaluated only on clean studio audio is evaluated on the easiest problem it will ever face.

2. Code-switched sentences specifically. Not an Arabic consultation and an English one — a single sentence carrying an Arabic clinical frame and an English drug name. That is the normal case in the Gulf and the one most likely to break.

3. The speakers you actually have. Clinicians in the region come from many countries and speak accordingly, as do patients. Test with the range your clinic really sees rather than one representative speaker.

4. What happens when it is unsure. A system that guesses confidently is more dangerous than one that flags. Ask to see how uncertainty is surfaced to the clinician, because that behaviour is what determines whether review is real or a formality.

5. Structure, not prose. A paragraph is easy; a note organised into the sections your record expects, with the right content in each, is the actual product. Ask how many sections and check they are the ones your specialty uses.

6. Codeable output. See the next section — this is the test most buyers skip and most regret skipping.

The dialect question, answered honestly

Vendors advertise dialect coverage as a list of country names. Treat the list as a claim to be tested rather than a specification, for a simple reason: there is no agreed benchmark behind it. Nobody publishes what "supports the Gulf dialects" means, what error rate it corresponds to, or on what audio it was measured.

The question that produces a usable answer is narrower. Ask for word error rate on YOUR recordings, and separately ask for accuracy on the clinical entities — drugs, doses, diagnoses, negation — because those are where an error changes the meaning of the note rather than its readability. A scribe can score respectably overall and still lose the one word that mattered.

Negation deserves its own test. "No chest pain" and "chest pain" differ by one word, and in a structured note that word decides what is recorded as a finding. Ask specifically how negation is handled in Arabic, where it can be expressed in more than one construction.

Finally, ask what happens to accuracy over time. A model that is not being retrained on the speech it actually meets will drift away from your clinicians rather than towards them.

The note is not the finish line

A scribe that produces a beautiful note and nothing else has solved the smaller half of the problem. In a region moving to classification-based funding, the same encounter has to survive being coded, grouped and billed — and the note is the evidence for all three.

That is a documentation-integrity question, not a transcription one. The note has to carry the specificity a coder needs: the severity, the comorbidity, the laterality, the complication documented as a complication. A note that reads well to a clinician and vaguely to a coder produces a lower-weighted group and a smaller payment, and nobody notices because nothing failed.

So the question to ask a scribe vendor is not only "is the note accurate" but "is the note codeable". Ask which classifications the output is aligned to, and whether anything downstream checks the note against what the coder will need before the claim is built.

This is where an ambient scribe and clinical documentation integrity stop being separate purchases. The scribe writes it; documentation integrity defends it.

Regulation, residency and integration in the Gulf

Three practical constraints decide whether a technically good scribe can actually be deployed here.

Regulatory acceptance. Saudi Arabia regulates health data and AI actively, and a vendor that can point to recognition from the relevant national authority is answering a procurement question that will be asked regardless. Ask what the recognition covers and when it was granted rather than accepting the logo.

Data residency. Ask where the audio goes, where it is processed, where it is stored, for how long, and who can reach it. Ask it about the transcription and the audio separately, because the answers are often different.

Integration without migration. The scribe has to reach the record you already run. A tool that requires replacing the HIS to work is not a scribe deployment, it is a record migration with a scribe attached — and it will be scheduled accordingly. Ask how it attaches to the existing system and what format it writes back in.

How DoctorSense answers this

DoctorSense is NANO Health Suite’s ambient clinical documentation product, and these are the claims it makes on its own page rather than a summary written for this guide.

Language. Medical speech recognition in Arabic and English with full code-switching support — built for consultations that move between the two rather than for one language at a time.

Output. A structured clinical note organised into 33+ clinical sections, following ICD-10, SNOMED-CT and LOINC, exported as HL7, FHIR or JSON into Epic, Cerner, Meditech and local systems.

Recognition and scale. Accredited by SDAIA and in the hands of more than 30,000 healthcare professionals, and compliant with Saudi eHealth regulations.

Deployment. A one-click Chrome extension over the HIS or EHR already in use, with go-live inside 24 hours and no migration.

The coding loop. DoctorSense writes the note at the source and NANO AI CDI 360 validates, codes and strengthens it downstream — which is the answer to the "is the note codeable" test above, and the part a standalone scribe leaves to you.