AI Discharge Instructions in Indian Languages: Safe Ways to Do It

AI discharge instructions in a patient's own language are one of the most practical uses of generative AI for Indian hospitals. Most discharge summaries are written in English, in medical shorthand, for other doctors. Many patients and families in Kerala, Tamil Nadu, Uttar Pradesh or anywhere else read them poorly or not at all, and what they miss (a changed dose, a warning sign, a follow-up date) is exactly what brings them back.

Large language models can rewrite a summary in plain language, and translation models can render it in Malayalam, Hindi, Tamil and other Indian languages. Studies from 2024 and 2025 show big gains in readability, alongside omissions, occasional hallucinations and uneven translation quality. India's national language platform, Bhashini, is also being connected to public health platforms.

This guide covers what the evidence shows, what is happening in India, and a review workflow that captures the benefit without putting patients at risk.

The problem with discharge summaries

A typical Indian discharge summary contains diagnoses, procedures, investigation results, a medication list and follow-up advice, written for clinicians. Patients need something different:

• What happened to me, in two or three sentences.
• Which medicines to take, when, and which ones changed or stopped.
• What I should and should not do at home.
• Which symptoms mean I must come back or call.
• When and where my follow-up is.

Our guide on reducing discharge time covers the discharge process; this post focuses on the patient-facing document.

What the evidence shows

Plain language works, but accuracy needs review. A 2024 NYU Langone study in JAMA Network Open used GPT-4 to convert 50 inpatient discharge summaries into patient-friendly versions.

But physician review found problems: in 100 reviews, 54 gave the top accuracy score, 18 raised safety concerns, mainly omissions with some hallucinations, and summaries were rated fully complete in only 56% of reviews. The authors concluded physician review is still required.

Machine translation is uneven by language. A September 2025 study in JAMA Network Open compared AI translation with professional translation for 34 paediatric discharge instructions (148 sections) at Seattle Children's in Spanish, Chinese, Somali and Vietnamese. Overall, AI scored lower than professional translators on fluency, adequacy and meaning; Spanish was roughly comparable, Chinese consistently worse. Earlier UCSF work on emergency discharge instructions found Google Translate 92% accurate for Spanish and 81% for Chinese, with errors linked to jargon and long sentences.

None of these studies covered Indian languages. The pattern is consistent, though: quality is better for languages with more training data and worse for others, and medical jargon is where translation breaks.

What is happening in India

India has invested heavily in language technology:

• Bhashini, the government's language platform under the Digital India Bhashini Division, provides translation and speech services. In January 2026 the National Health Authority signed an MoU with Bhashini to bring multilingual, voice-enabled AI to PM-JAY and ABDM platforms. The Ministry of Ayush signed a similar MoU in May 2026.
• eSanjeevani, the national telemedicine service, crossed 50 crore teleconsultations in 2026 and supports 15 languages.
• Open models. AI4Bharat's IndicTrans2 supports all 22 scheduled Indian languages and is released under a permissive licence, so hospitals and vendors can run it themselves.
• Indian LLMs. Companies such as Sarvam have released models trained for Indian languages.

What we did not find is published evidence on the accuracy of AI-translated discharge instructions in Indian languages. Until it exists, hospitals should generate their own.

Where AI goes wrong in this task

• Omissions. Dropping a medication change or a warning sign while simplifying.
• Hallucinations. Adding advice that is not in the record, such as a diet restriction.
• Dose and timing errors. "1-0-1" becoming the wrong schedule; "before food" and "after food" confused.
• Drug names. Brand names transliterated wrongly, or generic and brand names mixed up.
• Register. Translations too formal or too literal, so patients still do not understand.
• Numbers and units. mg and ml confused, decimal points lost.

A safe workflow

The safest approach separates the steps and keeps a clinician in the loop.

Three design choices do most of the safety work:

• Structured medication tables, not prose. Build the medication section from structured pharmacy data and show it as a table with fixed, pre-translated phrases for timing. Do not let the model rewrite doses.
• Pre-approved phrase libraries. Warning signs, activity advice and common instructions translated once by professionals and reviewed by clinicians, then reused.
• Back-translation for checking. Translate the output back to English automatically and show both to the reviewer to catch meaning changes.

Voice and WhatsApp delivery

Many patients and caregivers prefer listening to reading. Text-to-speech in Malayalam, Hindi and other languages is now usable, and Bhashini offers speech services. A short voice note on WhatsApp explaining medicines and warning signs, alongside the printed sheet, can help families who care for elderly patients.

Keep delivery tied to the hospital record: log what was sent, when and in which language. Our guide to WhatsApp automation for hospitals and clinics covers templates, consent and opt-outs, and our patient engagement work covers follow-up journeys.

Worked example: a cardiac discharge in Malayalam

Consider a 68-year-old patient in Kochi discharged after an angioplasty, cared for at home by a daughter who reads Malayalam comfortably and English less so. Here is how the workflow above plays out.

1. Source data. The HIS supplies diagnoses, the procedure note, discharge medications from the pharmacy module (with new, changed and stopped flags), follow-up appointment and wound care orders.
2. Plain English draft. The LLM produces a one-page summary in fixed sections: what happened, your medicines, caring for the wrist or groin site, activity, warning signs, follow-up. It is told to write at a simple reading level and to say "not in record" for anything missing.
3. Clinical review. The resident reviews the draft in two to three minutes, corrects anything wrong and approves. The medication table is not editable here; it comes from pharmacy data.
4. Translation. Free-text sections are translated to Malayalam. The medication table uses pre-translated labels: drug name, strength, morning-afternoon-night pattern, before or after food, and duration. Warning signs such as chest pain, breathlessness and bleeding from the puncture site come from the approved phrase library.
5. Language check. A Malayalam-speaking nurse reads the translated sheet alongside the back-translation and confirms it says the same thing.
6. Delivery and teach-back. The nurse goes through the sheet with the daughter, asks her to explain the medicines back, and sends a copy and a short voice note on WhatsApp.

The model is doing useful work, but every safety-critical part (doses, warning signs, approval) is either structured data, a pre-approved phrase or a human check.

Measuring whether it works

Choose a few measures before you start:

• Time doctors spend on discharge documentation.
• Errors found during review (omissions, hallucinations, translation issues), tracked by type.
• Patient or caregiver understanding, using teach-back at discharge.
• Calls to the hospital with questions after discharge.
• Seven- and 30-day readmissions and emergency visits for relevant conditions.

Data protection and consent

Discharge information is sensitive health data under the DPDP Act. If you use a cloud LLM or translation API, know where data is processed, whether it is retained, and whether it is used for training. Running open models such as IndicTrans2 on your own infrastructure is an option when data residency matters. Our DPDP Act checklist for hospitals covers consent and processor contracts.

How to start

1. Pick one ward and two languages. For a Kochi hospital, perhaps general medicine in Malayalam and Hindi (for migrant workers).
2. Build the template with your clinicians: sections, reading level, medication table format.
3. Create the phrase library for timing, warning signs and common advice, translated professionally.
4. Test on 50 past discharges. Doctors and bilingual staff grade English and translated versions for omissions, errors and readability.
5. Pilot with review. Every patient version is approved before release.
6. Measure and expand. Add languages and wards once error rates and review time are acceptable.

At Algon we build this kind of pipeline using natural language processing on top of the hospital's own records, with review screens built into the discharge workflow.

Mistakes to avoid

• Translating the clinician-facing summary directly instead of first creating a patient version.
• Letting the model invent advice that is not in the record.
• No bilingual review for languages where machine translation is weaker.
• Delivering only on paper to patients who cannot read the script well.
• No logging of what was sent to whom.

Conclusion

AI discharge instructions in Indian languages can close a real gap between what doctors write and what patients understand. The evidence shows large readability gains, alongside omissions and uneven translation, and India's language infrastructure, from Bhashini to open models, is ready to build on. The safe pattern is structured data for medicines, an LLM for plain language, clinician approval, pre-approved phrases for critical instructions, and bilingual checks.

Start small, measure errors and understanding, and expand by evidence. For related patient-facing AI, see our guide to the AI chatbot for hospitals.