Guideline-locked AI for type 1 diabetes

Answers a family can trust at 2 a.m.

MedEDiabuddy is an AI companion for parents and teenagers living with type 1 diabetes. It is locked to paediatric guidelines and DiaBuddy's validated dosing algorithms, adapted to Indian food and culture, and it sends emergencies to hospital before any AI is involved.

See the validation How it stays safe

Shortlisted for the Kaichi Kida Award Paper Session, APPES 2026, Brisbane

Can't see the video? Watch it on YouTube

96% agreement with a blinded paediatric endocrinologist across 50 clinical scenarios

90% for ChatGPT and for Gemini on the same scenarios

56% for caregivers with an average of seven years' experience of diabetes

0 of 30 real family questions where both blinded reviewers preferred the nurse educator's answer

01 · The need

Families' questions don't wait for clinic hours, and general chatbots aren't built for children with diabetes.

About 301,000 children and adolescents under 20 in India live with type 1 diabetes (IDF Diabetes Atlas 2025), with very few diabetes educators to support them. Parents turn to search engines and general-purpose AI, which are not tied to paediatric guidelines, do not know Indian food or fasting practices, and can miss an emergency.

The knowledge gap persists Caregivers whose children had lived with diabetes for 7.3 years on average answered only 56% of everyday scenarios correctly. Mean HbA1c in their children was 10.0%.

Human support doesn't scale Diabetes educators are scarce across the Asia-Pacific; one educator cannot answer every family's message the moment it arrives.

Generic AI isn't accountable ChatGPT and Gemini scored 90% on our scenarios: good, but without guideline grounding, local context or a hard safety layer.

02 · How it works

A language model on a short leash: safety rules first, validated algorithms second, curated guidelines third.

1 Red-flag safety layerHard-coded checks run before any AI call. Signs of ketoacidosis or severe hypoglycaemia go straight to an emergency instruction.
2 Six deterministic DiaBuddy algorithmsMeal bolus, correction, basal titration, sick-day matrix, hypoglycaemia and exercise. Doses come from validated rules, not from the language model.
3 Curated guideline knowledgeISPAD 2022 consensus guidelines and their 2024 updates, ISPAE 2017 and ICMR type 1 diabetes guidelines, and the clinic's family booklet. A commercial language model answers only from this curated collection, and parent questions are de-identified before they reach it.
4 Indian contextA registry of 30 common Indian diabetes myths with evidence-based replies, carbohydrate values for more than 10,000 foods, and guidance for festivals and religious fasts.
5 An answer pitched to the readerParent, Teen and Doctor modes change the language and depth, not the clinical content.

Emergency question · Parent mode

My daughter has been vomiting since morning, she is breathing fast and her urine ketones are 3+. Should I give extra insulin at home?

This needs hospital care now. Repeated vomiting, fast breathing and high ketones can be signs of diabetic ketoacidosis. Please take her to the nearest emergency department immediately and call your diabetes team on the way. Do not try to manage this at home.

Triggered by the safety layer before any AI response.

Eight domains covered in validation

Sick days Exercise Hypoglycaemia Insulin dosing Myths Nutrition Insulin storage Adolescence

Everyday question · Parent mode

I want to go to a dance party for an hour, and my insulin dose is due now. What precautions should I take?

Great that you want to enjoy the party! Your sugar reading right now guides everything. Before you go

Below 70 mg/dL: treat the low first and wait until it is above 90
70–125 mg/dL: eat 15–20 g of fast carbs before you start
126–270 mg/dL: fine to dance; keep carbs with you
Above 270 with ketones: if ketones are above 1.5 mmol/L, skip the party and follow your sick-day plan

Your insulin dose An hour of dancing is moderate-to-vigorous activity, so the meal dose usually needs to be reduced by 25–75%. Use the DiaBuddy Insulin Wizard to calculate it, or call your diabetes team. After the party Dancing can cause a low 7–11 hours later, often during sleep. Check your sugar when you get home and have a slow-carb snack, such as roti with dal or a glass of milk, before bed.

Condensed from a MedEDiabuddy answer, based on ISPAD exercise guidance.

Topics on the home screen

Families can type a question in their own words or start from one of fourteen everyday topics.

Low sugar (hypo) Sick day Missed insulin Indian food carbs Exercise and sport Birthday parties School and diabetes Night-time lows Post-meal highs CGM basics Sugar targets Travel tips Festivals Teen burnout

03 · Evidence

Tested against experts, caregivers, leading AI models and a real diabetes educator

Kaichi Kida Award Paper Session · APPES 2026 · Brisbane · 5 October 2026 Development and validation of MedEDiabuddy, a guideline-curated AI companion for paediatric type 1 diabetes Shah R, Agarwal N, Raithatha D, Sachan N, Pai K, Shukla P, Bajpai A. Presented by Dr Riya Shah for the Regency CDER and MedEClasses team. Shortlisted from abstracts across the Asia-Pacific; not yet peer reviewed.

Watch the video

Dr Riya Shah at the podium presenting the MedEDiabuddy development and validation study, with the title slide on screen — Dr Riya Shah presenting MedEDiabuddy at the Kaichi Kida Award Paper Session, APPES 2026, Brisbane.

STUDY 1 · BLINDED VIGNETTE VALIDATION

50 scenarios, scored against a blinded expert

MedEDiabuddy, ChatGPT, Gemini and 30 caregivers answered the same 50 vignettes across eight domains. Answers were scored for expert concordance, guideline adherence and completeness.

Correct answers (concordant with the expert)

MedEDiabuddy

96%

ChatGPT

90%

Gemini

90%

Caregivers (n = 30)

56%

Safety review

48 of 50 answers matched the expert, each scoring 5/5 for guideline adherence and completeness. Of the two that didn't, one (advising a doctor's review) carried no harm. One — declining an extra insulin dose during a high reading without measurable ketones — was rated potentially unsafe through under-treatment. Both errors erred on the side of caution, and the rule was corrected in the next version.

STUDY 2 · REAL-WORLD COMPARISON

AI versus the clinic's diabetes nurse educator

30 real questions from caregivers were answered by both MedEDiabuddy and our diabetes nurse educator. Two paediatric endocrinologists, blind to the source, scored every answer.

SafetyHigherP < 0.001

CompletenessHigherP < 0.001

Guideline adherenceHigherP < 0.001

Effect sizes were large on all three measures. In none of the 30 encounters did both reviewers rate the educator's answer above MedEDiabuddy's.

The reviewers used the scales differently (ICC 0.49–0.67; 77–92% of scores within one point), but the direction of effect was identical for both. A larger arm of about 120 encounters is under way.

Children's mean age14.9 years

Diabetes duration7.3 years

Mean HbA1c10.0%

Lower-middle income families53%

These results were presented at APPES 2026 and are not yet peer reviewed. They are not evidence of safety in unsupervised use: independent safety evaluation is planned before wider release, and MedEDiabuddy is used only under the child's diabetes care team.

04 · Key observations

What we learned

Curation beats scale A model restricted to paediatric guidelines and validated rules scored higher than general-purpose chatbots (96% vs 90%) on the questions families actually ask.
It fails safe Both errors were conservative, never over-treatment, and the red-flag layer means emergencies never depend on the AI.
It narrows a 40-point gap MedEDiabuddy scored 40 percentage points above caregivers with years of experience, on the same everyday scenarios.
It extends the educator On 30 real questions, blinded reviewers rated its answers safer and more complete than the nurse educator's. The aim is for one educator to supervise support for many more families.

05 · Future plan

From validated prototype to a companion in every DiaBuddy app

DONE · 2026

Vignette validation: 96% expert concordance
30-encounter blinded comparison with a nurse educator
Presented at APPES 2026 (Kaichi Kida session)

NEXT · 3–6 MONTHS

100 real-life queries with a further independent adjudicator, then publication
Complete the ~120-encounter educator comparison
Launch as the chat option inside the DiaBuddy app

2027 · TRIAL

Five-centre randomised trial of DiaBuddy with MedEDiabuddy
Retrieval over the full guideline library
Copyright, trademark and CDSCO regulatory pathway

SCALE

Hindi and other Indian languages
Educator dashboard for supervising AI conversations
Extension to other paediatric endocrine conditions

06 · Developers

Clinicians building AI they would trust with their own patients

Dr Anurag Bajpai

Lead developer and principal investigator · FRACP

Consultant Paediatric and Adolescent Endocrinologist at Regency Centre for Diabetes, Endocrinology & Research, Kanpur, and President of ISPAE 2025–26. Developer of DiaBuddy, whose algorithms form MedEDiabuddy's clinical core, and of the MedEClasses suite of decision-support tools.

Trained at AIIMS New Delhi and the Royal Children's Hospital, Melbourne
Founder, MedEClasses Private Limited (DPIIT-recognised) and GROW Society; MedEClasses courses have reached more than 1,700 health professionals in over 50 countries
Series Editor, Advanced Pediatric Endocrinology; h-index about 24

Development and research team