AI Model Links Insulin Resistance to 12 Types of Cancer in Half-Million-Person Study

Researchers led by the University of Tokyo used a machine-learning model to show that insulin resistance is a risk factor for 12 types of cancer in a study of half a million UK Biobank participants, offering the first population-scale evidence of this long-suspected link. The findings were published in Nature Communications in a paper titled "Machine learning-predicted insulin resistance is a risk factor for 12 types of cancer."

Insulin resistance—when the body doesn't properly respond to insulin, a hormone that helps control blood glucose levels—is one of the fundamental causes of type 2 diabetes and is tightly associated with obesity. In addition to diabetes, it is widely known that insulin resistance can lead to cardiovascular, kidney and liver diseases. Despite its broad impact, insulin resistance has been notoriously difficult to measure directly in clinical settings, limiting researchers' ability to understand its full consequences.

That challenge prompted Yuta Hiraike, a researcher from the University of Tokyo Hospital, and colleagues at the University of Tokyo to turn to artificial intelligence. The team recently developed a machine-learning tool called AI-IR, which predicts insulin resistance using nine standard clinical measurements collected during routine health checkups. "We recently made a tool, AI-IR, for predicting insulin resistance in individuals based on nine different pieces of medical information," Hiraike said. "It proved successful and made us think we could apply this tool to related concerns."

One of those concerns was cancer. Although scientists have long suspected a link between insulin resistance and certain cancers, gathering large-scale evidence has been difficult because direct measurement requires specialized testing available only in advanced diabetes clinics. While a possible link between insulin resistance and cancer has been suggested, large-scale evidence has been limited due to the difficulty of evaluating insulin resistance in the clinic. By applying AI-IR to 500,000 UK Biobank participants, the team was able to estimate insulin resistance at a population level and examine its relationship to cancer incidence.

"With AI-IR, we have provided the first population-scale evidence that insulin resistance is a risk factor for cancer," Hiraike said. Because the model relies on routine clinical data, he added, "AI-IR could be easily implemented to identify high-risk individuals and enable focused screening of diabetes, cardiovascular disease and cancer."

The study also highlights the limitations of relying on body mass index (BMI) as a proxy for metabolic health. It's common at present for BMI, a measure for body fat, to predict an individual's insulin resistance and knock-on susceptibility to related cancers. But with that there are false positives, where some obese people are considered metabolically healthy and don't suffer the ill effects of obesity to the same degree as others, and false negatives, where people with an ideal BMI end up suffering from insulin resistance or related concerns usually connected with obesity. By combining nine clinical parameters into a single metric, AI-IR can detect insulin resistance that BMI alone cannot explain.

Part of the challenge Hiraike and team faced was convincing reviewers of the paper that AI-IR could overcome these shortcomings in a reliable, repeatable way. Thankfully, they demonstrated not only its predictive power but also that their model is robust under various conditions. "When compared with directly measured insulin resistance in validation datasets, AI-IR achieved strong predictive performance," Hiraike said. "Directly measuring insulin resistance is impractical except for where patients are treated in specialized diabetes clinics. AI-IR provides a robust and scalable alternative for evaluating insulin resistance at the population scale."

The team now plans to explore how genetic differences influence insulin-resistance-related cancer risk and to integrate large-scale human data with molecular biology studies. "We are now working to understand how genetic differences between individuals influence this risk, and ultimately to link large-scale human data with molecular biology studies to develop better strategies to overcome insulin resistance," Hiraike said.