AI in Indian healthcare: the research questions that matter

AI for health is mostly built elsewhere and arrives in India as a finished product. Some of it works here; some of it does not; and for much of it, nobody has checked. The questions that decide whether AI helps Indian patients are research questions, and many of them are still open.

These are the ones we think matter most, what is known about each so far, and what is not. We are building our own research work around them, and we would like to hear from anyone working on the same problems.

1. Can a machine understand an Indian consultation?

An Indian consult rarely stays in one language. A doctor may ask about symptoms in Tamil, name the drug in English, and give instructions in a mix of both. For an AI scribe, that is the hardest case there is.

What is known. General speech recognition for Indian languages has improved a great deal, helped by open work from AI4Bharat. Its Vistaar benchmark set brings together 59 benchmarks across Indian languages; its IndicWhisper models had the lowest error rate on 39 of them. Its IndicVoices dataset covers 22 languages from more than 16,000 speakers across 145 districts.

What is not. Clinical speech is much less studied, and the evidence so far is in preprints rather than peer-reviewed papers. A 2025 preprint testing clinical interviews in Kannada, Hindi and Indian English found that some systems did well on Indian English but failed on code-mixed or vernacular speech. A 2026 preprint on Hindi-English primary care consultations reduced word errors from about 27% to 19% by adapting a model to the setting. We found no published clinical dataset for Tamil-English speech.

The open questions. How well do current systems handle drug names, doses and abbreviations inside code-mixed speech? What error rate is acceptable for a note a doctor will review? How can a clinical speech dataset be built with patients’ consent and without exposing their identity?

2. How safe are AI-written clinical notes?

When AI drafts a note or summarises a history, two kinds of error matter: adding something that was not said (a hallucination), and leaving out something that was (an omission).

What is known. A 2024 study in Nature Medicine found that adapted language models could produce clinical summaries that physicians rated as equivalent to or better than experts’ summaries in most cases, while still making errors that could cause harm. A 2025 study in npj Digital Medicine, built on almost 13,000 sentences annotated by clinicians, measured a hallucination rate of 1.47% and an omission rate of 3.45%, and found that a share of each was serious. And a 2026 study in Nature Medicine, MedHELM, argued that medical licensing exam scores say little about how models perform on real clinical tasks.

The open questions. Are error rates higher in notes drawn from code-mixed conversations? Which errors matter most for Indian prescriptions, where brand names, combinations and doses vary? How well does a doctor’s review actually catch errors under time pressure, and how should a review screen be designed to help?

3. Do AI scribes save time in Indian clinics?

What is known. A randomised trial at UCLA, published in NEJM AI in 2025 with 238 physicians, found that one ambient scribe cut time on notes by 9.5% compared with usual care, while another made no significant difference. Both were associated with less burnout, and clinicians reported occasional clinically significant inaccuracies. A 2026 study in JAMA of 8,581 clinicians found that scribe users spent about 16 fewer minutes on documentation per eight scheduled hours, and its authors described the effect as modest.

What is not. These are US studies, in clinics where a consultation is long and the documentation burden is heavy. An Indian out-patient department may see several times the number of patients, in shorter consultations, often with a junior writing the notes. Whether a scribe saves time there, frees the doctor to see more patients, or improves the record itself is unmeasured.

The open questions. What does a scribe change in a high-volume Indian OPD: time per patient, completeness of the record, patient understanding? Which specialties benefit most?

4. Will a model work on our patients?

What is known. Models often perform worse away from where they were built. The best-known example is a sepsis prediction model used across many US hospitals. Checked at Michigan Medicine against 38,455 hospital stays, it missed 67% of sepsis cases while alerting on 18% of all patients, with an AUC of 0.63.

The open questions. How do widely used models perform on Indian patients, whose disease patterns, body sizes, lab reference ranges and care pathways differ? How should a hospital without a data science team validate a tool before using it, and monitor it after? What does “drift” look like in Indian data, as recording habits and patient mix change?

5. What counts as good enough data?

AI is only as good as the records it reads, and the measurement of clinical data quality is well described in the research literature: completeness, correctness, concordance, plausibility and currency. We wrote about this in Clinical data quality: why AI starts with clean records.

The open questions. How good are records in Indian clinics and hospitals today, measured the same way across many sites? Which interventions at the point of entry improve quality most, for the least effort? How should quality be reported alongside any study or model that uses the data?

6. Can everyday care produce trustworthy evidence?

Real-world evidence, drawn from routine records rather than trials, is now used by regulators and drug makers elsewhere to answer questions trials cannot. India has an enormous volume of care, but its records are fragmented, often on paper, and rarely linked.

The open questions. How can records from many independent clinics be combined, with consent, into datasets that hold up to scrutiny? How far can ABHA-linked records help follow a patient across providers? Which methods correct for the gaps that routine Indian records have?

7. Governance that works in a 50-bed hospital

What is known. There is no shortage of principles. ICMR published ethical guidelines for AI in biomedical research and healthcare in 2023. WHO’s 2021 guidance sets out six principles. NITI Aayog’s Responsible AI papers set out seven for India. ISO/IEC 42001:2023 defines a management system for AI.

The open question. Most of these assume an organisation with a governance office. What is the smallest set of practices that lets a small hospital or a single-doctor clinic use AI safely, and how can software make those practices the default rather than extra work?

How we work on these questions

92health builds the EMR and AI tools that Indian clinics and hospitals use every day, which puts us close to these questions. We are building our research work around them, and we want to do it in the open: with clinicians and academic groups, with patients’ consent, under ethics committee oversight where it is required, and with results published.

If you are working on any of these questions, or want to, book a call with our team.

Sources

• Bhogale KS et al. Vistaar: diverse benchmarks and training sets for Indian language ASR. Interspeech 2023. arXiv:2305.15386
• IndicVoices (AI4Bharat). arXiv:2403.01926, 2024.
• ASR under the stethoscope (preprint). arXiv:2512.10967, 2025.
• DASHA (preprint). arXiv:2603.06373, 2026.
• Van Veen D et al. Adapted large language models can outperform medical experts in clinical text summarization. Nature Medicine, 2024;30:1134-1142. doi:10.1038/s41591-024-02855-5
• Asgari E et al. A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine, 2025;8:274. doi:10.1038/s41746-025-01670-7
• Bedi S et al. Holistic evaluation of large language models for medical tasks with MedHELM. Nature Medicine, 2026;32:943-951.
• Lukac PJ et al. Ambient AI scribes in clinical practice: a randomized trial. NEJM AI, 2025. doi:10.1056/AIoa2501000
• Rotenstein LS et al. JAMA, 2026;335(16):1408-1417.
• Wong A et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 2021. doi:10.1001/jamainternmed.2021.2626
• WHO. Ethics and governance of artificial intelligence for health, 2021.