Machine learning model links insulin resistance to 12 types of cancer in 500,000-person UK study

Insulin resistance beyond diabetes: a cancer connection hiding in routine health data

Insulin resistance - when the body's cells stop responding properly to the hormone that regulates blood glucose - is familiar as the metabolic root of type 2 diabetes. Its role in cardiovascular disease and fatty liver disease is also well-established. But the full scope of insulin resistance's health consequences has been harder to measure, partly because the condition itself is difficult to evaluate in clinical settings.

Blood glucose levels and BMI are the proxies most commonly used to assess insulin resistance risk, but both are imprecise. Some individuals with elevated BMI are metabolically healthy and show none of the expected signs of insulin dysfunction. Others with normal weight develop insulin resistance and its complications. The true picture of who has insulin resistance, and what health consequences follow from it, has been obscured by these measurement limitations.

Yuta Hiraike and his team at the University of Tokyo Hospital have approached this problem through machine learning. Their tool, called AI-IR, predicts insulin resistance from nine pieces of standard health information - the kind of metrics collected in routine checkups rather than specialized metabolic testing. After validating AI-IR in earlier work, the team applied it to data from more than 500,000 participants in the UK Biobank to ask a question that prior studies lacked the scale to answer: is insulin resistance a risk factor for cancer, and if so, which types?

What the analysis found

The results, applying AI-IR's insulin resistance predictions to the UK Biobank cohort, identified insulin resistance as a risk factor for 12 specific types of cancer. The study does not identify all 12 cancer types in publicly available summary, but the analysis was powered by the scale of the UK Biobank - one of the largest genomic and health databases available to researchers - giving it substantially more statistical authority than prior work on this question.

The potential mechanisms linking insulin resistance to cancer risk include chronically elevated insulin levels, which can act as a growth signal in tumor cells. Insulin-like growth factor 1 (IGF-1), which increases when insulin resistance is present, promotes cell proliferation and inhibits programmed cell death - processes relevant to cancer development. Chronic low-grade inflammation associated with insulin resistance may also contribute to carcinogenesis in multiple tissue types.

"We recently made a tool, AI-IR, for predicting insulin resistance in individuals based on nine different pieces of medical information. It proved successful and made us think we could apply this tool to related concerns," said Hiraike. "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. But with AI-IR, we have provided the first population-scale evidence that insulin resistance is a risk factor for cancer."

Why measuring insulin resistance matters

The current standard for diagnosing insulin resistance in clinical practice is the hyperinsulinemic euglycemic clamp - a procedure that requires intravenous glucose and insulin infusions over several hours, specialized equipment, and trained staff. It is used almost exclusively in research settings. In routine clinical care, insulin resistance is generally inferred from proxy measures like fasting glucose, hemoglobin A1c, and BMI, all of which miss a significant proportion of people with the condition.

AI-IR was designed to address this gap. The nine input parameters it uses - the team does not specify all of them publicly, but they are described as standard health checkup measurements - are available for virtually any patient who has undergone a routine physical examination. The tool could, in principle, be integrated into standard health records to flag individuals at elevated insulin resistance risk without requiring any additional testing.

"Since the nine input parameters for AI-IR are obtained through standard health checkups, AI-IR could be easily implemented to identify high-risk individuals and enable focused screening of diabetes, cardiovascular disease and cancer," Hiraike said.

Limitations and what comes next

The UK Biobank is a powerful dataset but not a perfect one. Its participants are predominantly of European ancestry and were recruited from the UK population, meaning the model's predictions and the associations it identifies may not generalize equally to all ethnic groups or geographic populations. Insulin resistance prevalence and its relationship to cancer risk may vary across populations with different dietary patterns, genetic backgrounds, and healthcare environments.

The study design is observational - it identifies associations between predicted insulin resistance and cancer incidence, but cannot establish causation with certainty. The AI-IR tool predicts insulin resistance rather than measuring it directly, introducing a layer of uncertainty. And the 12 cancer types identified as associated with insulin resistance will need validation in independent datasets before the finding can be considered definitive for each specific cancer type.

The study represents a proof of concept for applying machine learning metabolic predictors to large biobank datasets to identify disease connections that would be impossible to detect with conventional measurement approaches. The scale advantage - 500,000 participants versus the hundreds or few thousands typical of insulin sensitivity studies - is what makes this kind of analysis possible.