Blood sugar does not jump: it climbs for years, then accelerates

When does the road to type 2 diabetes begin, and what does it look like? A single blood test says whether a person is above or below a threshold today. The researchers wanted the shape of the curve: how fasting blood sugar, blood sugar after a glucose drink, the body's sensitivity to insulin and the pancreas's insulin output change in the years before a diagnosis.

The Whitehall II study follows British civil servants in London. For this analysis the researchers used 6,538 people without diabetes at the start (71% men, 91% white), examined up to three times at intervals of about five years, with a fasting blood sample and a standard glucose tolerance test each time: 11,790 examinations in all. Over a median of 9.7 years, 505 of them developed diabetes. The researchers then lined up everyone's measurements backwards from the moment of diagnosis (or from the last visit, for those who stayed healthy) and fitted curves to the group, up to 13 years before. Insulin sensitivity and insulin secretion were not measured directly: they were estimated from fasting glucose and insulin with a standard formula (the homeostasis model assessment, HOMA).

In people who never developed diabetes, everything moved slowly and in a straight line: fasting blood sugar crept up a little, insulin sensitivity fell a little, insulin output stayed flat. In people who later developed diabetes, the picture was different for years before anyone could have made a diagnosis. Fasting blood sugar rose in a straight line but faster, then from about three years before diagnosis climbed steeply, from 5.79 to 7.40 mmol/L. Blood sugar two hours after the glucose drink shot up in the last years, from 7.60 to 11.90. Insulin sensitivity dropped sharply in the five years before diagnosis. And the pancreas told a story of its own: between four and three years before diagnosis it produced more insulin (its estimated function rose from 85.0% to 92.6%), as if trying to compensate, then its output collapsed to 62.4% at diagnosis. The authors read this as the signature of a long compensation that eventually fails.

There is no "effect size" here in the usual sense, because the study does not test a treatment or a predictor: it describes curves. The figures above are averages for the group of 505 future cases, drawn through up to three measurements per person. The differences between the two groups were already visible 13 years before (future cases had higher glucose and lower insulin sensitivity throughout), but the changes that stand out start three to six years before diagnosis. How much any single person's curve resembles the group's, the study cannot say.

The authors draw a careful conclusion: there seem to be different windows for screening and prevention. Prevention programmes usually target people with "prediabetes", who, on these curves, are already on the accelerating part. The authors hypothesise that acting earlier, while the curve is still flat, might work better, and they say plainly that this needs more research. The study itself tests no intervention.

This is the study closest to the idea of the Digital Twin, and the one that shows its limit most clearly. The data are ordinary: fasting glucose, insulin, a glucose tolerance test, the kind of results a person receives from routine check-ups and can keep in the Twin. What the study shows is that a series of such results carries information a single result does not. What it does not provide is a model: the authors say that risk models using repeated measurements "could" be more accurate, and that this is something to build and test, not something built. The Twin lines up the values; turning that series into a prediction for one person is work that has not been done.

It does not show that a person's own values can be read against these curves to tell how far they are from diabetes: with at most three measurements each, individual trajectories could not be studied. It gives no accuracy figures, no threshold, no model. The group was an occupational cohort, mostly white men in London, and a quarter of the original sample was left out for missing data; the authors judged this unlikely to bias the shapes, but the numbers belong to that group. Insulin sensitivity and secretion were estimated with a formula, not measured. And a detail that matters for a reader checking sources: the author manuscript archived in PubMed Central reports a median follow-up of 8.2 years, the published article 9.7; this page uses the published figure and flags the difference.

Prospective occupational cohort (Whitehall II, London) with up to three clinical examinations per participant (phases 3, 5, 7) and questionnaire phases; retrospective trajectory modelling anchored at diagnosis (cases) or last follow-up (controls); multilevel longitudinal models (linear and piecewise), adjusted for age, sex, ethnicity and phase; sensitivity analyses.

6,538 civil servants free of diabetes at phase 3 (1991–1993), 70.7% men, 91.3% white, aged 35–55 at recruitment; 505 incident cases (49.1% by oral glucose tolerance test, 35.4% doctor-diagnosed, 15.4% on medication) over a median 9.7 years (published); 11,790 person-examinations (801 in future cases, 10,989 in controls).

Fasting venous samples (≥ 8 h) and 2-h 75 g oral glucose tolerance test at each screening; plasma glucose by glucose oxidase (YSI analysers); serum insulin by radioimmunoassay (phase 3) and ELISA (phases 5, 7); HOMA2-%S and HOMA2-%B from the HOMA2 calculator (model-derived estimates). Diabetes: fasting glucose ≥ 7.0 mmol/L or 2-h glucose ≥ 11.1 mmol/L, doctor diagnosis or medication.

Controls: linear trends in all measures except insulin secretion (flat). Future cases: fasting glucose linear then steep quadratic from year −3 (5.79 → 7.40 mmol/L); 2-h glucose rapid rise from year −3 (7.60 → 11.90 mmol/L); HOMA2-%S steeper decline in the last 5 years (to 86.7%); HOMA2-%B rise between years −4 and −3 (85.0% → 92.6%) then fall to 62.4% at diagnosis. Future cases had higher glucose and lower sensitivity throughout the 13 years. No discrimination, calibration or predictive metrics: the study is descriptive. Sensitivity analyses (extended fasting window, complete-data sub-cohort, midpoint timing) supported the shapes; adjustment for time-varying BMI attenuated the between-group difference in sensitivity but not the trajectory differences.

Not a prediction model: PROBAST does not apply; a risk-of-bias judgement for an observational cohort (ROBINS-I is for interventions; here a narrative judgement): selection moderate (25.8% excluded; included participants younger, more often white and male); measurement low-to-moderate (HOMA estimates, assay change across phases); outcome low (OGTT-based definition, current criteria). Generalisability limited to a white, male-dominated occupational cohort. Reporting: pre-TRIPOD; methods reported in detail.

Conceptual, not operational: the study motivates risk models based on repeated measures and earlier prevention windows, but builds neither. No evidence that knowing the trajectory changes outcomes.

A for the principle (repeated routine measures of glucose and HOMA kept in order over years); no model to apply. Required data: fasting glucose and insulin, ideally OGTT, from routine clinical testing, with dates; a dynamic risk model would have to be built and validated before any individual reading.

Published abstract (PubMed record) and the full text of the author manuscript in PubMed Central, read in full on 5 November 2026; the typeset Lancet article was not consulted. Supplementary webappendix not consulted. Two numerical differences between manuscript and published abstract are recorded above.

Background: little is known about the timing of changes in glucose metabolism before type 2 diabetes. The authors characterised trajectories of fasting and post-load glucose, insulin sensitivity and insulin secretion in people who develop the disease. Methods: prospective occupational cohort (Whitehall II) of 6,538 British civil servants without diabetes at baseline; over a median 9.7 years, 505 cases were diagnosed (49.1% by oral glucose tolerance test); retrospective trajectories were assessed from up to 13 years before diagnosis (cases) or end of follow-up (controls). Findings: in non-diabetics (10,989 measurements) all measures followed linear trends, except insulin secretion, which did not change. In the diabetic group (801 measurements), fasting glucose rose linearly then steeply from 5.79 to 7.40 mmol/L starting 3 years before diagnosis; 2-h glucose rose rapidly from 3 years before (7.60 to 11.90 mmol/L); insulin sensitivity fell steeply in the 5 years before diagnosis (to 86.7%); beta-cell function rose between years 4 and 3 before diagnosis (85.0% to 92.6%) and then fell to 62.4%. Interpretation: changes in glucose, insulin sensitivity and insulin secretion are visible 3–6 years before diagnosis; describing these trajectories could contribute to more accurate risk models using repeated measures from regular check-ups.

The authors' abstract cannot be reproduced here for rights reasons: read it at the link.