Telemedicine and diagnostic support

Where AI in health actually helps

The honest map of where AI helps in Nepali health, in 2026, has three clear zones.

Imaging and image classification. This is the strongest fit. A trained model can read chest X-rays for pulmonary tuberculosis with sensitivity comparable to a junior radiologist — and Nepal screens hundreds of thousands of suspected TB cases a year, more than the country’s radiologists can read carefully. Models exist (qXR, Lunit, CAD4TB) that have been deployed in Nepal under various pilots; the integration into the public TB programme is partial.

The same logic applies to skin photographs for common dermatological conditions, retinal images for diabetic retinopathy screening, and — increasingly — ultrasound for fetal age estimation in antenatal care. The pattern is consistent: model-assisted screening at the frontline, with confirmation by a clinician.

Triage and symptom checking. A symptom-checker chatbot — well-designed, with clear limits — can help a patient decide whether to go to the local health post, the district hospital, or self-manage. The technology is not new; what is new is that LLM-based interfaces handle ambiguous natural-language complaints much better than the rigid dropdown systems of five years ago. The risk is well known and worth naming: a confident-sounding triage that gets a rare presentation wrong causes real harm. Building this well means heavy guardrails, conservative escalation, and a relentless focus on not being the sole decision-maker.

Clinical decision support. For frontline workers, software that pulls together a patient’s history, today’s symptoms, the protocol for their suspected condition, and prompts the worker through the next correct step is a genuine quality-of-care multiplier. The MOTECH and CommCare platforms have done versions of this for over a decade; modern LLM-enhanced clinical decision support is a clear next step.

What does not work yet

Three pitches that, in 2026, you should remain sceptical of:

Fully autonomous diagnosis from a chatbot. A patient describes their symptoms; the model gives a diagnosis and a treatment recommendation; the patient acts on it without seeing a clinician. The combination of LLM hallucination, the long tail of rare presentations, and the absence of a clinician who can examine the patient makes this unsafe at the population scale. It will be normalised somewhere, eventually; Nepal does not need to be where it is tested first.

AI-only mental health support. The market for affordable, accessible mental health care in Nepal is enormous and almost entirely unmet. Several companies are pitching AI-only mental health products into this gap. The harm pattern from similar products elsewhere — wrong escalation of suicidal ideation, sycophantic reinforcement of distorted thinking — is well documented. The right shape here is AI as an intake and triage tool that feeds a human clinician network, not as a standalone therapist.

Predictive diagnosis of complex conditions from limited data. “We will use AI to predict diabetes risk from your phone.” This kind of pitch usually rests on shaky causal assumptions and weak data. Predictive risk scoring for chronic disease in Nepal is a research problem, not a deployment one — at least until much better longitudinal data exists.

What about LLMs in the clinic

The introduction of general-purpose LLMs into Nepali clinical settings is happening informally already. Junior doctors paste symptom descriptions into ChatGPT to brainstorm differentials. Pharmacists check drug-drug interactions. Health workers ask for plain-Nepali explanations of patient handouts.

This is, on balance, useful — when the clinician treats the LLM as a smart colleague to be checked, not an authority. The failure mode is the reverse: a junior worker trusting a confident-sounding LLM in a domain where the cost of being wrong is high. Health systems in Nepal are quietly grappling with how to integrate this without either banning it (unenforceable) or letting it run unchecked (dangerous).

The infrastructure missing

Three pieces of infrastructure would make every AI-in-health project in Nepal noticeably more useful.

A national health data exchange that allowed appropriately authorised AI models to read across the HMIS, hospital records, and lab systems would unlock dozens of high-value projects. The privacy and consent design is non-trivial; the technical work is straightforward. This is one of the highest-value pieces of public-good AI infrastructure the country could build.

A licensed AI-in-medicine framework from the Nepal Medical Council and the Ministry of Health, defining what AI tools clinicians may use, under what supervision, with what data, and what the chain of responsibility is when a model gets something wrong. The absence of this framework is currently the largest legal-uncertainty barrier to careful deployment.

A locally-validated benchmark set for the most common health-AI tasks — chest X-ray for TB, photograph for common skin conditions, symptom triage in Nepali — that gives the country a credible way to evaluate vendor claims rather than relying on benchmark scores from elsewhere.

Check your understanding

Quick check

—

Which deployment pattern for AI in Nepali health has the strongest evidence of helping patients without unacceptable risk?

1. AA fully autonomous chatbot that diagnoses and treats patients without involving a clinician.
2. BAI-assisted screening (e.g., chest X-rays for TB, skin photos, retinal images) where the model flags suspect cases for a clinician to confirm — augmenting frontline capacity rather than replacing the clinician.
3. CAI-only mental health therapy with no human clinician in the loop.
4. DPredictive disease scoring from phone-collected data alone.

Check answer→

Quick check

—

Why is a national health data exchange one of the highest-leverage infrastructure investments for AI-in-health in Nepal?

1. AIt would replace the need for hospitals.
2. BIt would unlock a wide range of high-value AI projects by making longitudinal, cross-facility patient and population data available — under appropriate consent and privacy controls — that no single hospital or HMIS silo can currently provide.
3. CIt would automate clinical decisions.
4. DIt would let patients chat with their data.

Check answer→

What comes next

Health is one thread of public-good AI. Identity and public records are another — citizenship, vital registration, land, court records. The next section is about what AI can usefully do for the public-records side of the state, and the privacy frame that has to come with it.