IIT-Madras Wants Its India-Trained Foetal-Age AI in Scanning Rooms by 2027-28

Doctors in India estimate a foetus's gestational age from ultrasound measurements using formulas built largely on Western populations, the Hadlock formula being the common one. IIT-Madras wants to replace that with a model trained on Indian pregnancies, and is now working on getting it into hospitals.

Prof Balaraman Ravindran, head of IIT-Madras's Wadhwani School of Data Science and AI, told DT Next the aim is to have something deployable by 2027-28.

What has been built

• The model is called Garbhini-GA2. It was developed and externally validated on data from Indian pregnancies.
• Published research shows it significantly reduced gestational-age estimation errors compared with Hadlock.
• External validation used data from Christian Medical College, Vellore, with work also involving Faridabad.
• Next steps are further validation and integration with ultrasound scanning systems. Army hospitals are being explored for more evidence.

Ravindran was clear that the model is ready and deployment is the hard part. Government hospitals already spend about 90 seconds per scan and find it stressful, so any extra time per patient is a real cost.

Why I'm watching this one

I see this pattern again and again with applied AI in India. The research result is the easier half. A model built on local data beats an imported formula, the papers are out, the validation is done. What decides whether it helps anyone is whether it fits into a scan that has to finish in 90 seconds on equipment the hospital already owns.

That is a product and integration problem, not a modelling one, and it is the same problem enterprise AI teams hit every day. Ravindran said the focus is on "translating the research into a deployable hospital technology", and that's the right way to put it.

The Wadhwani school is also working on Indian-language AI tools for teachers, some designed to run on mobile devices with little cloud dependence. It's the same idea: design around the constraint that's really there, whether that's a 90-second scan or a patchy network.

Source: DT Next: Indigenous tech to determine foetal age in 2027-28, says IIT-Madras