AP and Lateral

Medical AI, held to the light.

Radiograph of a lumbar spine, front viewAP
Radiograph of the same lumbar spine, side viewLAT

A weekly brief on medical AI in the Middle East, written for clinicians. Every story is shown in two views - what was announced, and the angle it leaves out.

Issue 2, week to 19 November 2026

This issue in one minute

  1. Saudi health ministry warns the public against using AI to diagnose themselves
  2. Kuwait's health ministry describes a chatbot that explains patients' records and prescriptions
  3. Egypt's health minister and Huma discuss AI for the national screening programmes
  4. Vivan Therapeutics approved to study AI treatment-ranking software in colorectal cancer in Abu Dhabi
  5. Adult chest X-ray AI retrained for newborns across 30 intensive care units, with sites in Doha, Riyadh and Beirut
  6. OpenEvidence gave doctors in six countries an evidence base drawn mostly from the US, small study finds
  7. Model for choosing a biologic in severe asthma was tested on 56 patients
  8. Two open models are badly overconfident on Arabic medical exam questions, preprint finds
Cairo Kuwait City Riyadh Doha Muscat Abu Dhabi
Where this issue's stories come from: UAE 2, Saudi Arabia 1, Kuwait 1, Egypt 1, Qatar 1, Oman 1.

This week 4 stories

Saudi Arabia, public guidance

Saudi health ministry warns the public against using AI to diagnose themselves

AP - what was announced

The Ministry of Health told the public that AI is no doctor and should not be asked to diagnose a condition, the Saudi news site Ajel reported on 17 November, citing the ministry's account on X. People can use it to understand their health, the post said, but should not give it test results and ask for a treatment plan. A second report the next day says the ministry accepts AI for explaining medical terms, giving general information about a disease or a test, and preparing questions for a consultation. It urges people to use the 937 call centre and the Sehhaty app.

Lateral - the other view

The public may find that line hard to see, because the same report describes an AI tool inside the ministry's own Sehhaty app that takes symptoms and history by chat and writes a summary for the doctor before the consultation. Neither report says how that tool was tested, or how often its summary leaves out something a clinician would have asked.

Not disclosedWhat prompted the warning, the date of the ministry's post, and any figures on harm from self-diagnosis with AI.

Sources: Ajel (Arabic), Arabisk London (Arabic)

Kuwait, health ministry

Kuwait's health ministry describes a chatbot that explains patients' records and prescriptions

AP - what was announced

A Kuwaiti health ministry assistant undersecretary set out the ministry's digital tools on 11 November at an Arab League conference in Cairo, the Kuwait News Agency reported. One is an application he called a medical ChatGPT. It lets people explore their health records by text and voice, summarises and translates them and explains prescriptions, and he said it is meant for information and does not replace a doctor or pharmacist. He outlined the use of RapidAI to help diagnose stroke on non-contrast CT, and said AI that detects tissue abnormalities during colonoscopy is in advanced testing and clinical evaluation in Kuwait.

Lateral - the other view

A chatbot that explains a prescription can state a wrong dose as fluently as a right one, and the patient reading it has no way to tell. The report does not say which model is behind the application or whether patients can use it yet. Nothing is said about how the stroke and colonoscopy tools perform on scans and scopes done in Kuwait.

Not disclosedThe model behind the records chatbot, which of these tools are live, how many hospitals use RapidAI, and the size of the colonoscopy evaluation.

Source: Kuwait Times

Egypt, screening

Egypt's health minister and Huma discuss AI for the national screening programmes

AP - what was announced

Egypt's health minister met the founder and chief executive of Huma, a health technology company, to discuss using AI in the 100 Million Healthy Lives initiatives, state media reported on 16 November. The talks covered using the national databases those initiatives built to improve diagnostic accuracy, predict disease earlier and identify high-risk groups, with breast cancer screening and newborn care both mentioned. No agreement was reported. A week earlier, at Cairo ICT, the minister said AI is already part of early cancer detection and MRI analysis in Egypt, and one report of that speech names Roche as the partner in a national digital pathology network.

Lateral - the other view

Nothing reported says who would hold the screening data once a company's models are trained on it, or whether any tool would be tested against Egyptian radiologists before it reads Egyptian mammograms.

Not disclosedWhich Huma products were discussed, any timeline, and the evidence behind the minister's statement that AI is already used in diagnosis.

Sources: Maspero (Arabic), El Balad on the Cairo ICT speech (Arabic), Cairo 24 (Arabic)

Abu Dhabi, oncology

Vivan Therapeutics approved to study AI treatment-ranking software in colorectal cancer in Abu Dhabi

AP - what was announced

Vivan Therapeutics, a Hub71 startup, has received regulatory approval to begin a clinical registration study of its TuMatch software at Burjeel Cancer Institute, MENA Startup Digest reported on 13 November. TuMatch compares the genomic profile of a patient's tumour with the company's drug-response data and published evidence, then gives the oncologist a ranked list of treatment options. The prospective study will enrol at least 10 patients with colorectal cancer and will also collect health economics data for later talks on reimbursement. A larger multicentre study in pancreatic and other gastrointestinal cancers is planned for Europe and the Gulf.

Lateral - the other view

Ten patients are enough to show the software can be run alongside a tumour board. Whether acting on its ranking changes response or survival needs a comparison group, and the report names neither a comparator nor a primary endpoint.

Not disclosedWhich regulator gave the approval, the trial registration, and who is funding the study.

Source: MENA Startup Digest

Research watch 4 papers

Qatar, neonatal imaging

Adult chest X-ray AI retrained for newborns across 30 intensive care units, with sites in Doha, Riyadh and Beirut

AP - what was published

A study in Nature Biomedical Engineering took a model built to check endotracheal tube position on adult chest X-rays and retrained it for newborns one hospital after another, with no images pooled centrally. Thirty neonatal intensive care units in 20 countries and territories took part, and the authors include staff at Sidra Medicine in Doha, two Riyadh hospitals and the American University of Beirut. Each hospital gave 50 radiographs for retraining, according to the abstract. Error in locating anatomical landmarks was 0.67 cm, against 1.29 cm when a hospital tuned the model on its own images alone and 4.30 cm for the adult model unchanged, and the method did better at 29 of 30 sites. The paper appeared on 13 October and Sidra announced its part on 11 November.

Lateral - the other view

This was a retrospective study, and the abstract reports error in centimetres across all babies when a neonatologist will want it for the smallest ones, where there is least room. Detection of tubes placed too deep improved at 27 of 30 sites, with no sensitivity given.

Sources: PubMed abstract, the paper (DOI), The Peninsula

Oman, evidence tools

OpenEvidence gave doctors in six countries an evidence base drawn mostly from the US, small study finds

AP - what was published

Eleven clinicians in six countries, Oman among them, ran standard searches on 10 common conditions in OpenEvidence in June and July 2025 and scored the 80 outputs. They rated 88% completely or almost entirely accurate. The references came overwhelmingly from the US, none came from the Global South, and the answers did not change with the clinician's location even when the query stated it. A search about a febrile child in Kenya did not raise malaria. A twelfth clinician, in India, could not get access to the tool. The paper has two authors at Sultan Qaboos University, came out in Cureus on 16 October and was added to PubMed this month.

Lateral - the other view

OpenEvidence is on the tools list further down this page, and its answers cite their sources, so you can see where each recommendation was written. The 88% is the raters' own judgement on a rubric. Only one rater was in the Arab world, and the paper cannot tell you how the tool handles a differential that should include brucellosis or MERS.

Source: full text on PubMed Central

UAE, asthma

Model for choosing a biologic in severe asthma was tested on 56 patients

AP - what was published

Researchers at Dubai Health and the American University of Sharjah, with clinicians in Abu Dhabi, Riyadh, Poland and Canada, trained a random forest on registry data from 801 patients with severe asthma to predict who improves on a biologic. Improvement meant meeting at least two of three criteria covering the Asthma Control Test score, exacerbations and oral steroid use. After the team cut the model down to its most informative features, accuracy on held-out patients was 82.14% and specificity 66.67%. The strongest predictors were the control test score, eosinophil count, and nasal polyposis or sinusitis. The authors report no validation by site or country. Frontiers in Digital Health published the paper on 30 October and it was added to PubMed this month.

Lateral - the other view

The held-out set was 56 patients, 12 of them non-responders, so the rise in specificity from 33% to 67% is four more patients classified correctly, and the 62% chance of improving after a switch from omalizumab to benralizumab quoted in the abstract is a simulation for one illustrative patient that was never checked against a real switch.

Sources: full text on PubMed Central, the paper (DOI)

Arabic language models

Two open models are badly overconfident on Arabic medical exam questions, preprint finds

AP - what was published

A single-author preprint from an engineering college in Hyderabad, posted to arXiv on 12 November, tested two open models, Qwen2.5 and Jais, on English and Arabic medical multiple-choice questions. The Arabic questions came from MedAraBench, the benchmark covered in Issue 1. On Arabic, Qwen answered 43% correctly and Jais 31%, where guessing would score 24%. Both models were far more confident than their accuracy justified, and a standard correction called temperature scaling removed most of the gap. When each model was made to return a set of answers that contains the right one 90% of the time, the set held 78% to 89% of the options on the Arabic questions.

Lateral - the other view

A set that has to hold most of the options before it can be trusted narrows nothing down for a clinician, and that matters for anyone hoping a confidence score will flag which answers to distrust. The author notes that the two question sets are different exams and that no fluent Arabic reader checked the questions.

Source: arXiv 2611.16821

Money and deadlines

Date29 Nov

The World Innovation Summit for Health opens in Doha for two days. One of its five tracks covers AI and human-centred health, and registration is open on the summit's site. WISH

Competition1st

Rology, the Cairo teleradiology company, took first place in the Siemens Healthineers SHIFT Startup Connect programme in Saudi Arabia, where it competed against 16 healthcare ventures, according to AUC Venture Lab. MENA Startup Digest

Tools doctors are using

Being listed here does not mean we recommend a tool, and no one pays to be listed. Before you put patient information into any of them, check where the data is stored and whether your hospital and regulator allow it.

Evidence search, worldwideOpenEvidence

Answers clinical questions from the medical literature and shows its citations.

CheckSign-up has been tied to US clinician verification and may not work from your country.

openevidence.com
AI scribe, worldwideHeidi

Listens to the consultation and drafts the note, letters and referrals for you to edit.

CheckIt records the consultation, so you need the patient's consent and your hospital's approval.

heidihealth.com
Evidence summaries, orthopaedicsOrthoEvidence

Short, appraised summaries of orthopaedic trials, each with a risk of bias rating.

CheckOrthopaedics only, and full access needs a subscription or a free trial.

myorthoevidence.com
Decision support, built in the regionHakeemDx

Clinical guidance in Arabic and English at the point of care, from Hakeem Health.

CheckIt is sold to hospitals, universities and payers, and an individual doctor may not be able to sign up.

Wamda report
Clinical assistant, UK and QatarRhazes AI

Transcribes consultations, drafts documents and suggests diagnoses. Piloted in a hospital in southern Lebanon.

CheckSuggested diagnoses are the part to test hardest against your own cases.

Tech.eu report
Arabic documentation, Saudi ArabiaNuxera AI

Turns Arabic clinical speech into notes and standard codes such as SNOMED and ICD-10.

CheckThe 70% cut in documentation time is the company's own figure.

Entrepreneur Middle East report

Next issue in your inbox

Corrections and suggestions

Earlier issues

Issue 1, week to 12 November 2026