Can AI Detect Diabetes With Just Five Questions?

Almost half of all diabetes patients worldwide are not diagnosed until serious complications have already developed. Master’s student Kavin Varadharajulu (Erasmus University) has developed an AI system that can assess whether someone is at risk of diabetes using only five targeted questions. The system performs better than traditional screening methods and costs just $1,531 per correctly identified patient – a total savings of $60.9 million compared with having no early screening at all.

From Personal Experience to an AI Solution

“In my South Indian community, diabetes was accepted as something normal from one generation to the next,” Varadharajulu explains. “Family members ignored warning signs like weight gain until severe symptoms, such as swollen ankles, forced them to see a doctor. By then, treatment focused mainly on managing complications rather than prevention.”

These personal experiences inspired his research. Having grown up in California and later moving to the Netherlands to study Data Science, he saw an opportunity to use modern AI technology for early diabetes detection.

Artificial Intelligence as a Virtual Doctor’s Assistant

Varadharajulu tested five different AI approaches on a U.S. health dataset containing information from more than 400,000 people. The data was self-reported – similar to what you would enter at home in an online questionnaire.

“Imagine you’re sitting on the couch at home, worried about diabetes,” the master’s student says. “You open your hospital’s chatbot. It starts with basic questions about your age and lifestyle. But instead of running through a long standard list, the system uses a ‘contextual bandit’ – a smart technique that selects only the most relevant follow-up questions for each person.”

The result: while traditional questionnaires require dozens of questions, the AI system needs an average of just five to detect diabetes risk with 90% accuracy.

The Grocery Store Test

“A contextual bandit works much like choosing fruit at the grocery store,” Varadharajulu explains. “You keep trying different fruit, remember which ones were good, and use that experience to make better choices next time. The AI system does the same: it learns which questions are most informative for your specific situation.”

This means faster and more personalized screening. The system adapts to your answers – a 62-year-old smoker will receive different follow-up questions than a 30-year-old athlete.

Challenges and Fairness

The research also shows limitations. “The system identifies certain groups less accurately, such as people who are underweight or those over eighty,” Varadharajulu admits. “This is because these groups are underrepresented in the training data. AI should therefore always be seen as a supportive tool, never a replacement for medical staff.”

Ironically, the system performs less effectively for underrepresented groups, which often overlap with those lacking healthcare access  – the very group in which undetected diabetes is most common. “This highlights that technology alone is not enough. We also need better healthcare access and more diverse data.”

Future Outlook

Varadharajulu hopes his research will lead to pilot projects in hospitals. “Early detection saves lives – and with modern AI, we can make accessible, affordable diabetes screening available to everyone, before complications develop.”

But technology remains a tool. “Only when AI systems are carefully designed, monitored, and deployed fairly can they contribute to improving healthcare for everyone.”