A new Nature Communications study suggests that insulin resistance tracks with higher risk for 12 cancer types. An international research team applied an unusual approach — instead of measuring insulin resistance directly, they developed an algorithmic proxy, opening the door to population-scale risk analyses. Here’s why that matters.
Insulin resistance is a key metabolic risk marker, but gold-standard testing is impractical and can not be performed routinely. The Japan–Taiwan team treated insulin resistance as a digital biomarker, using a machine-learning model to predict the likelihood of insulin resistance from routine blood tests and basic anthropometrics.
The authors call this metric AI-IR — AI-calculated insulin resistance.
What’s important is that this is one of the first studies to show, at a cohort size of hundreds of thousands, that AI-derived insulin resistance is associated not only with future diabetes and cardiovascular events, but also with specific cancer types.
Insulin resistance is now common in many people because the prevalence of obesity and metabolic syndrome is rising. Yet it is usually not checked directly at population scale — not because nobody needs it, but because direct methods are impractical for screening: the gold standard is complex, and even a simpler option like HOMA-IR requires a fasting insulin test.
Why this is a problem: insulin resistance is hard to measure at the population level
Insulin resistance is a state in which the body’s tissues respond less effectively to insulin, forcing the pancreas to produce more of it to keep glucose within the normal range. Over time, this compensation stops working, and the risk of type 2 diabetes and cardiovascular disease increases.
Why might this be linked to cancer? Researchers have several lines of explanation.
First, when tissues respond poorly to insulin, the body often compensates by keeping chronically higher-than-normal insulin levels. And insulin functions as a “build and grow” signal: it promotes nutrient uptake and supports cellular growth and repair. At the same time, the related IGF (insulin-like growth factor) pathway may become more active, and it too is involved in cell growth and survival. Under normal conditions these signals support healthy function, but when they’re persistently elevated, they may create conditions that favor the emergence and progression of tumor cells.
Second, obesity and metabolic disturbances are often accompanied by chronic low-grade inflammation, which is considered a potential contributor to cancer development.
At the epidemiological level, the association between diabetes/obesity and the risk of several cancers has long been well described. Agency for Research on Cancer (IARC) have also emphasized that excess adiposity is associated with higher risk of multiple cancer types.
But there is a methodological trap here: diabetes and obesity are end states, or coarse markers. Insulin resistance often appears earlier, can exist in people with normal body mass, and is not always documented in medical records. And the gold standard for measuring insulin resistance — the hyperinsulinemic euglycemic clamp — is virtually impossible to use at scale. As a result, large cohorts typically do not have a direct label of “IR present / IR absent.”
There is another widely used tool as well: HOMA-IR, an index based on fasting insulin and glucose. But even it runs up against a practical barrier — fasting insulin is not routinely measured in many countries.
This is where AI can be used to full effect: if a direct variable is hard to measure, you can estimate its likelihood from a set of available features, and then test whether this indirect metric predicts future outcomes.
Study design and key findings
Data: a large biobank and “long” medical records. The authors applied AI-IR to data from the UK Biobank — one of Europe’s largest projects that follows people’s health over many years: participants were assessed at baseline, and subsequent diagnoses and hospitalizations are then identified through official medical databases.
This is an important nuance for interpretation: the study didn’t rely on self-reported cancer, but on linked registry records with diagnoses and dates. That reduces recall bias and self-reporting errors.
Model: predicting HOMA-IR from routine measures. Here’s how it works: it takes standard lab results and questionnaire data and tries to infer whether a person has marked insulin resistance — that is, whether they would have had a high HOMA-IR if we had measured it directly.
The authors chose a HOMA-IR threshold of 2.5: if the probability estimated by the model is above a preset level, the person is classified as “AI-IR positive.” The calculation requires nine familiar variables: age, sex, self-identified race/ethnicity (as a demographic factor), body mass index, glucose, HbA1c, triglycerides, total cholesterol, and “good” HDL cholesterol. The model itself had previously been validated on other large datasets (NHANES in the US and Taiwan’s MJ cohort), and in this paper the team simply applied the already-built algorithm to the UK Biobank and examined whether it predicts future diagnoses and events. They also addressed a practical issue: in the UK Biobank, blood samples were not always collected in the fasting state, so the authors tested results across groups with different fasting durations and showed that the association with diabetes risk held throughout, so they did not exclude part of the cohort from the analysis.
Sanity check: diabetes and cardiovascular outcomes. Before turning to cancer, the authors test whether their metric behaves the way an insulin-resistance marker should. The logic is straightforward: if AI-IR truly reflects metabolic dysfunction, then people with high AI-IR should be more likely, over time, to develop diabetes, and more likely to experience cardiovascular complications, which typically track alongside insulin resistance. That is exactly what they observe.
AI-IR predicts future diabetes better than coarser proxies such as BMI alone or simple lab-derived indices. They then show that even among people without diabetes at baseline, high AI-IR is associated with a higher risk of heart attacks/strokes and with higher mortality. In other words, the metric detects a metabolic problem early, before a clinician formally diagnoses diabetes.
This can be viewed as a plausibility test: if AI-IR did not predict diabetes and MACE, it would be hard to trust the subsequent conclusions about cancer.
Key result: which cancers are linked to AI-IR, and how strong the association is
This is where things get most interesting. The cancer analysis included hundreds of thousands of participants who had no cancer diagnosis at baseline. Overall, across all cancers combined, there was almost no difference. But when the data were broken down by site, the pattern became heterogeneous: AI-IR was associated with increased risk for a number of specific tumors.
After correction for multiple comparisons (Bonferroni), a statistically significant association remained for six cancer types: uterine cancer, kidney cancer, esophageal cancer, pancreatic cancer, colorectal cancer, and breast cancer. For another six cancer types, the authors report nominal associations — signals that are weaker and do not formally survive the strict correction: tumors of the renal pelvis, small intestine, stomach, liver and gallbladder, leukemia, and bronchial and lung cancer.
The researchers then combine several tumor types into a composite — a set of cancers for which risk rises with AI-IR. For this composite, among people without diabetes but with AI-IR positivity, the hazard ratio (HR) is about 1.25 after adjustment for age and sex.
The next step is an attempt to separate what is “just obesity” from what may reflect broader metabolic dysfunction. The authors add BMI to their models and show that weight explains part of the effect, but not all of it. In particular, for bronchial and lung cancer, the association with AI-IR becomes even stronger after accounting for BMI — an effect they interpret as a BMI-independent component of risk linked to other markers of metabolic ill health (glucose, lipids).
Smoking is another important layer. Because smoking sharply increases lung-cancer risk and also affects body weight and metabolism, it can easily confound the statistics. The authors adjust for smoking status and analyze interactions. They report that the AI-IR effect on lung-cancer risk is most noticeable in former smokers, while in current smokers it is “drowned out” by the enormous absolute risk from smoking itself.
What was known about the insulin resistance–oncology link before this
The medical community has already accumulated a large body of evidence linking obesity, diabetes, and metabolic dysfunction to cancer, but it has long debated what’s to blame — adipose tissue as such, chronic inflammation, high insulin, lipid changes, related factors like smoking and physical activity, or all of the above.
A large meta-analysis in The Lancet back in 2008 showed that BMI is associated with the risk of multiple cancer types, and this line of evidence has been reinforced by more recent work. In 2016, the International Agency for Research on Cancer (IARC) systematized the evidence and identified a set of sites for which excess adiposity is considered a risk factor at the level of international consensus.
But those studies mainly discuss weight and body composition, not insulin resistance per se. Diabetes research then entered the picture. Umbrella reviews that synthesize meta-analyses have shown that type 2 diabetes is associated with increased risk of a range of tumors and with cancer mortality, although the strength of evidence varies by site and confounding remains a problem.
Within the UK Biobank, studies have also appeared linking insulin-resistance indices to specific tumors. For example, a 2025 Scientific Reports paper examined associations between several IR surrogates and the risk of small-cell lung cancer, again emphasizing the role of smoking and the need to adjust for it. Against this backdrop, the 2026 Nature Communications paper looks like the next step in an evolution rather than a bolt-from-the-blue discovery.
It does three things at once.
First, it proposes a metric that integrates several routine parameters into a single score and shows that, for diabetes prediction, it outperforms single indices. This follows the logic of a “composite biomarker,” where the weaknesses of individual measures partially offset one another.
Second, it extends that metric to cancer at the scale of the full cohort and identifies a cluster of tumors for which risk increases. This aligns well with broader knowledge about which cancers are more often linked to obesity and diabetes, but adds an important nuance: weight alone does not always depict the state of a person’s metabolism. Two people can have the same BMI yet very different metabolic health. That’s why a metric built from multiple lab tests can sometimes give a sharper risk signal than a single weight measure.
Third, the study explicitly accounts for smoking, because it strongly affects lung-cancer risk while also influencing weight and metabolism. If you don’t address this, you can mix up causes: it may look like metabolism is driving the association when part of the effect is due to smoking. After adjustment, the authors find that the link between insulin resistance and lung cancer is most evident in former smokers. That makes sense: in current smokers, risk is already very high because of cigarettes, and against that background the contribution of metabolic factors can be “lost.”
If you zoom out to the broader arc of the last few years, it looks like this: the field used to ask more often, “How is obesity linked to cancer?”, then “How is diabetes linked to cancer?”, and now an increasing number of studies are trying to isolate the intermediate link — insulin resistance/hyperinsulinemia — and learn how to measure it at scale. The AI approach in this paper is an attempt to do exactly that, using what many high-income countries actually have in their datasets: standard lab tests and registries.
What may change, and what is still too early to promise
This work raises a practical question that healthcare systems tend to like most: can we identify — more cheaply and earlier — a group of people who need closer follow-up. The authors explicitly discuss the potential use of AI-IR to flag high-risk individuals and enable more targeted screening for diabetes and for a range of cancers.
But it’s important to understand the caveats. The study shows associations, not direct causality. Even after adjustment for BMI and smoking, there are other lifestyle and health factors that could partly explain the relationship. In addition, the model was trained against the surrogate indicators, rather than the gold standard — clamp, and its performance depends on input-data quality and on the context of laboratory measurements. If such a metric were to be implemented as a digital marker in clinical practice, it would need validation across different populations and in more diverse groups by age and ancestry.
Transparency is another issue. The authors did not release the model code openly because they filed a patent; instead, they provide an online calculator that computes the metric from user-entered data. That doesn’t mean the results are wrong, but it does create a downside: it is harder for other scientists to fully verify exactly how the score is computed and to reproduce the analysis one-to-one. As a result, confidence in the conclusions will depend more on whether independent groups can confirm the same findings in other datasets.