Researchers built the first medical dataset for Urdu — a language spoken by 230 million people with almost no AI training data

A team of researchers created a labeled dataset of 153,000 medical terms in Urdu by collecting text from health news sites, prescriptions, and hospital websites, then having three medical annotators tag the entities by hand. This is the first time anyone has built a standard benchmark dataset for teaching AI systems to recognize medical concepts in Urdu, which means AI medical tools can now be evaluated and trained on actual Urdu language text instead of approximations built from other languages.

Most AI medical tools work in English because that's where the training data exists. Urdu speakers — 230 million people across Pakistan, India, and diaspora communities — have been locked out. This dataset is small, but it removes the first bottleneck: now researchers can build and measure AI that actually works in Urdu medical text, instead of relying on English models that fail on non-English input. The next question is whether anyone actually uses it to build Urdu medical AI, and whether companies building medical tools for the region will adopt it as a standard instead of building their own proprietary datasets.