Medical Scribing AI for Oman's Health Ministry Workflows

A consultant in a Muscat referral hospital sees thirty patients on a busy clinic day. Each encounter generates a structured note, an Arabic narrative for the family, an English summary for the file, an order set, and a follow-up plan. The Ministry of Health does not need its clinicians spending another two hours each evening typing what they already said out loud. It needs an Arabic-first ambient scribe sitting on ministry hardware, drafting the note while the clinician examines the patient, and never sending a byte of patient audio to a foreign cloud.

The ministry-class workflow

Oman's public health system, organised by the Ministry of Health, runs three encounter shapes that a scribe must serve cleanly:

• Referral hospitals (Royal Hospital, Sultan Qaboos University Hospital, Khoula, Nizwa, Sohar): high-acuity multispecialty, long ward rounds, multidisciplinary case conferences, mixed Arabic family conversation with English clinical reasoning.
• Polyclinics and health centres: high-volume short visits, primary care in Omani Arabic, repeat prescriptions, paediatric and antenatal templates, narrow vocabulary but heavy throughput.
• Specialty clinics (oncology, cardiology, endocrinology, mental health): structured assessments, longer dictations, sensitive disclosures, dense terminology in both languages.

A useful scribe is not a transcription gadget bolted onto an EHR. It is a workflow that reads the encounter, drafts a structured SOAP note, surfaces orders and follow-ups for confirmation, posts to the existing electronic health record through a documented integration, and lets the clinician sign off in under sixty seconds.

Arabic-first ambient capture

The scribe stack is two models, not one. The first listens; the second writes. Both must be evaluated on actual Omani clinical audio before any commitment.

• Speech-to-text. Whisper Large V3 handles Arabic and English in the same utterance, which matches how a Muscat consultant actually speaks. Light fine-tuning on Omani medical audio (consented and de-identified) lifts accuracy on local drug names, hospital names, and dialect markers.
• Note generation. Qwen 3.6 in the 30B to 70B range covers bilingual reasoning, structured output, and template adherence. Clinics that prefer a Gulf-trained foundation can substitute Falcon Arabic from TII, which targets MSA and several Arabic dialects.
• Diarisation and redaction. Speaker separation distinguishes clinician, patient, and family member. A redaction layer strips numeric identifiers (civil ID, file number) before storage of the working transcript.

The pattern is the same one the global market calls ambient clinical intelligence ACI, popularised by Nuance DAX. The substantive difference for a sovereign deployment is Arabic medical NLP that is not an afterthought, and weights that never leave the building.

Hardware sizing per facility tier

Sizing tracks concurrent encounters, not headcount. A useful rule of thumb for a Hosn-class deployment:

1. Polyclinic, twenty clinicians. A single 2U appliance with two enterprise GPUs (RTX 6000 Ada or H100 PCIe), 256GB RAM, and a small NVMe pool. Handles ambient capture, on-device transcription, and note drafting at peak clinic load.
2. Regional hospital, two hundred concurrent encounters. A 4U rack node with four to eight GPUs, plus a separate transcription pool to keep ASR latency under two seconds. Local object storage for short-retention audio, encrypted at rest, with a documented purge cycle once the note is signed.
3. National referral centre. A small enclave of two to four nodes with redundancy at the appliance level, integrated with the hospital information system through HL7 FHIR or the existing SDK, and replicated to a warm standby in a second ministry data hall.

Power, cooling, and rack space slot inside facilities the ministry already operates. There is no public cloud subscription, no vendor lock-in on inference per token, and no surprise egress bill.

PDPL and clinician supervision posture

Health data is the most sensitive category named under the Omani Personal Data Protection Law (Royal Decree 6/2022). A scribe that reaches the wards has to clear three doctrines on day one.

• Data residency by default. Audio, transcripts, generated notes, and audit logs all live on ministry hardware. Cross-border processing is permitted only with a documented basis under PDPL, which a sovereign deployment removes from the design space entirely.
• Clinician supervision on every output. The model produces a draft. The clinician reads, edits, and signs. The signed note is the medical-legal record, not the model output. This pattern matches the WHO guidance on the ethics and governance of AI for health, large multi-modal models, which insists on human accountability for every clinical decision.
• Audit trail in writing. Every prompt, every retrieved chunk, every model edit, and every clinician override is logged with timestamp, user, and case identifier. A senior officer can answer the question "who saw what, when" without forensics.

For the broader treatment of this pattern across specialties and validation regimes, see our pillar on the Arabic medical scribe Oman playbook.

Phased rollout from one specialty to ministry-wide

Ministry-scale rollouts succeed when they start narrow, prove value, and expand on evidence. A defensible sequence:

1. Phase 1, single specialty pilot. Pick one specialty (often endocrinology or cardiology) in one referral hospital. Six to ten consultants, eight weeks, instrumented with edit-distance metrics and time-to-sign. The pilot question is simple: does the consultant save time without losing clinical accuracy?
2. Phase 2, multi-specialty in the same hospital. Expand to four specialties, refine prompts and templates per discipline, add an oncology safety review for any oncological dictation. Validate against a sample of double-scribed notes.
3. Phase 3, governorate cluster. Roll out to all polyclinics in one governorate plus the regional hospital. Train local champions, set up a tier-2 support rota, and start retiring the manual end-of-day typing burden.
4. Phase 4, ministry-wide. Replicate the appliance pattern across hospitals and health centres. Centralise model governance and version control; keep inference local. Continue clinical validation on a rotating sample.

Each phase has an exit criterion, a measurable saving, and a written sign-off from the medical director. The rollout never outpaces the evidence.

To discuss a Hosn-class medical scribe deployment for a single specialty pilot or a multi-hospital programme, email [email protected] for a one-hour briefing. We will walk through hardware sizing, model selection, validation design, and procurement framing in line with PDPL and ministry expectations.