AI in Healthcare in Nepal
The order of operations
Every serious AI application in health rests on the same foundations that interoperability rests on:
- Identity — without it you cannot assemble a longitudinal record, and without a longitudinal record most clinically interesting predictions are not computable.
- Structure — free text without terminology binding is not a feature set.
- Terminology — codes are what make data comparable across sources.
- Governance — who may use which data, for what secondary purpose, under what basis.
This is why the AI conversation in Nepal keeps returning to the same place as the interoperability conversation. They are the same problem viewed from different ends.
Where the realistic value is
Language. Nepal's clinical and community documentation is substantially in Nepali, and much of it is unstructured. Language technology that can handle Nepali text — summarising, extracting structured facts, translating between clinical and lay registers — addresses a problem that global vendors have little incentive to solve well. This is the clearest case of local advantage.
Coding support. Suggesting ICD and SNOMED CT codes from clinical narrative, with a human confirming. This improves data quality at the point where quality is actually created, which makes it doubly valuable.
Documentation burden. Reducing double entry and transcription. Time returned to clinicians is the most reliably realised benefit in health IT.
Data quality and surveillance. Anomaly detection over routine reporting — implausible values, sudden reporting-pattern changes, missing units — is a well-matched problem for the data Nepal already has.
Access to specialist knowledge. Screening support where specialist density is low, with the caveats below. See imaging AI for how that path actually works.
What to be careful about
Regulatory position. Software intended for diagnosis, prevention, monitoring, prediction, prognosis or treatment is generally a medical device. That brings risk classification, a quality management system, clinical evaluation and post-market surveillance — and retraining a model is a change to a regulated product. Establish the position first; a pilot that cannot be lawfully deployed is an expensive prototype. Detail: clinical AI.
Representativeness. A model trained on data from urban tertiary hospitals does not describe rural primary care. Deployed there, it fails hardest on the people the health system already serves least well. See AI ethics.
Automation bias. Clinicians accept confident-looking suggestions. Labelling output as "advice" does not transfer accountability if the interface makes disagreement difficult.
Data leaving the country. Sending identifiable health data to an external inference API is a governance decision with legal and sovereignty implications, not an implementation detail. It should be made deliberately, and it is a strong argument for local or on-premise inference where feasible.
Silent decay. Models degrade as practice, populations and upstream data change. Without monitoring, degradation is invisible until it is harmful.
My take
The pressure to announce AI is currently much stronger than the readiness to operate it, and that mismatch has a predictable cost: attention and budget move from the boring layer to the visible one.
If I had one recommendation for Nepal's health sector on AI, it would be this: spend the next two years making the data usable, and use AI where it does not touch clinical decisions. Coding support, documentation, translation, surveillance anomaly detection, administrative triage. Those deliver real value, carry manageable risk, and — critically — they improve the data foundation that any later clinical application will need.
The alternative path, in which a diagnostic model is piloted before there is a client registry or a terminology service, produces a demonstration that cannot scale, cannot be evaluated properly against a real baseline, and cannot be lawfully deployed. I have not seen that path end well anywhere.
There is one more thing worth saying plainly: an AI system that recommends care in a language its users cannot read, on a device they do not have, over connectivity they do not reliably get, is not a health intervention. Nepal's constraints are physical before they are computational, and technology strategy that forgets this produces impressive pilots in Kathmandu and nothing in Karnali.
Where to go next
- Digital Health in Nepal — the wider picture
- Healthcare interoperability in Nepal — the data foundation AI depends on
- Background: machine learning, clinical AI, AI ethics
Sources
- WHO. Ethics and governance of artificial intelligence for health. Geneva: World Health Organization; 2021. who.int
- WHO. Global strategy on digital health 2020–2025. Geneva: World Health Organization; 2021. who.int
- Paudel S, Paudel D, Boucher F. Digital Health in Nepal: A Perspective on Overcoming Challenges and Leveraging Opportunities. Kathmandu University Medical Journal. 2025; 91(3): 386–91. PDF