Risk Stratification in Diabetes Prevention

Type 2 diabetes prevention is often described as a choice between lifestyle intervention and medication. A prior question is just as important: who is likely to benefit, on what timescale, and from which intensity of support?

Risk stratification separates people into groups with different expected outcomes or management needs. In prevention, it can concentrate a limited programme where absolute benefit is likely to be greatest. But a tier is useful only if it is valid in your target population, tied to an action, and designed so that people with fewer resources are not screened out of effective care.

Four questions that look similar but are not#

Screening asks who should receive a test. Diagnosis asks whether current measurements meet criteria for diabetes or prediabetes. Prognosis estimates the chance of developing diabetes over a stated period. Allocation decides what action follows from that estimate.

A questionnaire based on age, body size, family history, or blood pressure can identify people for laboratory testing. It cannot diagnose diabetes. An A1C or glucose result can classify current glycaemia, but one value alone does not supply a complete personalized probability of progression. A prognostic model can estimate that probability, but it does not establish that a particular intervention will work better in one tier than another.

Confusing the tasks creates hidden errors. A model validated for five-year diabetes incidence may be used as a diagnostic substitute. A prediabetes label may be treated as if everyone within a broad glycaemic range has the same prognosis. A high-risk cutoff may be selected because it sounds clinically meaningful, without showing that the programme capacity or treatment evidence supports it.

Screening defines who enters the pathway#

Current recommendations differ because they answer related questions with different evidence standards: the USPSTF recommends screening adults ages 35 to 70 who have overweight or obesity and referring people with prediabetes to effective preventive interventions. It notes that earlier screening may be considered in populations with higher prevalence or risk.

The American Diabetes Association's 2026 Standards of Care provide a broader clinical framework, including risk-based testing and testing beginning at age 35 for other adults. Details such as pregnancy, medications, pancreatic disease, symptoms, and prior results can place a person on a different pathway.

These are policies, not biological boundaries. Turning 35 does not cause an abrupt risk change. A body-mass-index threshold is an imperfect proxy across ancestry, body composition, and life course. So check locally who is missed, who completes testing, and whether a positive screen actually leads to prevention support.

Prediabetes is a heterogeneous risk state#

Common laboratory definitions include A1C from 5.7% to 6.4%, fasting plasma glucose from 100 to 125 mg/dL, or two-hour plasma glucose from 140 to 199 mg/dL during a 75-g oral glucose tolerance test. These measures capture overlapping but nonidentical physiology and populations.

Progression risk is generally higher with values nearer the diabetes range, multiple abnormal tests, rising values, younger age with substantial metabolic burden, prior gestational diabetes, and other cardiometabolic factors. Regression to normoglycaemia also occurs. Measurement variation, haemoglobin conditions, acute illness, and medicines can affect interpretation. So the category is a starting point, not a complete prognosis, and an effective pathway records the assay, threshold, confirmation rules, comorbidities, and planned retesting rather than leaving you with a binary label.

What the Diabetes Prevention Program established#

The US Diabetes Prevention Program randomized 3,234 adults at high risk, selected with elevated fasting glucose and impaired glucose tolerance, to intensive lifestyle intervention, metformin, or placebo. Over an average 2.8 years, diabetes incidence was 58% lower with lifestyle intervention and 31% lower with metformin than with placebo.

Those are relative reductions in a selected high-risk trial population, and the lifestyle programme targeted at least 7% weight loss and at least 150 minutes of weekly physical activity, supported by an intensive curriculum and individual coaching. It was not equivalent to brief advice to “eat better and exercise.”

Absolute benefit depends on untreated risk. If two groups experience the same relative reduction but one has twice the baseline event rate, the higher-risk group will avert more cases per 100 people treated over the same time. This is the principal efficiency argument for stratification.

Long-term follow-up adds nuance. A 2025 report through 21 years found reductions in diabetes incidence of about 24% for the original lifestyle group and 17% for the metformin group, with median diabetes-free survival extended by 3.5 and 2.5 years, respectively, compared with placebo. Group differences attenuated as interventions and care evolved after the randomized phase.

The prevention of a surrogate or intermediate event should not be silently converted into proof of every downstream benefit. A separate 21-year analysis did not find a reduction in major cardiovascular events from the original lifestyle or metformin assignments. Treatment crossover, use of cardioprotective medicines, and long follow-up complicate interpretation, but the result still limits claims that delaying diabetes necessarily prevents cardiovascular events.

Prognostic enrichment can improve efficiency#

A prediction model combines several variables to estimate an individual's outcome probability over a defined horizon. Population tools can use nonlaboratory data to identify communities or adults for further assessment, while clinical models may add glucose, A1C, lipids, medication history, or prior gestational diabetes.

Discrimination asks whether people who develop diabetes tend to receive higher predictions than those who do not, while calibration asks whether a predicted 20% risk corresponds to about 20 events per 100 similar people. For allocation, calibration is especially important because it determines expected absolute benefit and capacity.

External validation is mandatory. A model can transport poorly when the baseline incidence, ethnicity, deprivation, measurement, healthcare access, or competing risks in your population differ. A 2024 external validation of the Diabetes Population Risk Tool illustrates the need to test performance in distinct cohorts rather than treating a published equation as universal.

The prediction horizon must match the decision. A short-term model may favor people already near a diagnostic threshold. A lifetime framework may identify younger people with substantial cumulative risk who look low risk over three years; neither is inherently correct, but your programme has to say which one it is aiming at.

Prognosis is not treatment-effect prediction#

High baseline risk often produces larger absolute benefit when relative treatment effects are stable. But it does not prove that relative effect is the same in every subgroup. A variable can be strongly prognostic without modifying the effect of an intervention.

To claim differential treatment response, an analysis should test an interaction or use a validated individualized treatment-effect method, preferably prespecified and replicated. Comparing a statistically significant result in one subgroup with a nonsignificant result in another is not an interaction test.

The DPP found evidence that metformin's relative benefit varied across some baseline characteristics, with greater effects reported in younger adults, people with higher body mass index, higher fasting glucose, and a history of gestational diabetes. Subgroup findings should be read with their uncertainty and subsequent guideline interpretation. They inform selection; they do not turn a risk score into a prescribing rule.

A practical tiered pathway#

A prevention service can define tiers without pretending that biology has natural cut points:

1. Population reach: create environments that support physical activity, nutritious food, sleep, and access to primary care without requiring a risk label.
2. Screening: use age and clinical risk to offer appropriate laboratory testing, with routes for earlier testing where indicated.
3. Risk characterization: confirm glycaemic status and assess trajectory, comorbidities, pregnancy history, medicines, social constraints, and cardiovascular risk.
4. Matched support: offer an evidence-based lifestyle programme, consider metformin for selected high-risk adults under current guidance, and manage blood pressure, lipids, smoking, and other risks.
5. Follow-up: repeat measurements on a defined schedule, monitor tolerability and participation, and escalate promptly if diagnostic thresholds or symptoms emerge.

The threshold for a scarce intensive programme can reflect capacity, expected absolute benefit, patient preferences, and harms. It should not be reverse-engineered to fill slots and then presented as a biological truth.

Equity is part of model performance#

Risk tools can reproduce unequal access embedded in their development data. People with fewer opportunities for diagnosis may appear to have lower recorded incidence. Variables such as postcode can improve prediction while encoding deprivation. Digital enrolment and attendance requirements can preferentially exclude people working multiple jobs, living far from a site, or needing language or childcare support.

Audit your pathway from invitation through outcome. Useful measures include invitation, testing, eligibility, enrolment, completion, weight and glycaemic change, adverse effects, progression, and loss to follow-up, stratified by relevant demographic and access factors. Calibration should also be examined across groups, while avoiding unstable conclusions from small samples.

An equitable design can combine universal population measures with proportionate intensity for higher-need groups. Risk stratification should unlock support, not become another barrier to it.

References#

1. ADA Standards of Care 2026: prevention or delay of diabetes
2. ADA Standards of Care 2026: diagnosis and classification
3. USPSTF screening recommendation
4. Original Diabetes Prevention Program trial
5. Diabetes incidence and diabetes-free survival through 21 years
6. Cardiovascular outcomes through 21 years
7. External validation of a population diabetes risk tool