Applies to a specific group
This appraisal is written for adults with prediabetes or otherwise at high risk of type 2 diabetes, and adults diagnosed within roughly the last six years who are not on insulin. If you are not in one of these groups, the recommendations here may not transfer: longer-standing or insulin-treated diabetes sits outside most of the evidence below, and the population-level picture can differ.
Two decisions face this group, and the evidence answers them differently: whether diabetes can be kept from developing, and whether diabetes caught early can be put into remission.
- An intensive lifestyle programme cut the rate of new diabetes by roughly a third to a half over three to six years in people with prediabetes, in three separate trials on three continents. Built on modest weight loss and activity most days, it beat metformin by about two to one in the one trial that compared them (Lifestyle vs Metformin for Diabetes Prevention).
- Most of that effect tracks the weight lost. Physical activity is also linked to lower risk in people who miss the weight target.
- Delaying the diagnosis is the proven gain. Two decades of follow-up in one trial, in a group at relatively low cardiovascular risk, found no fewer heart attacks or strokes.
- For someone diagnosed in the last few years and carrying excess weight, a large weight loss can put diabetes into remission. On a structured formula-diet programme, about half were in remission at one year and about a third at two. The more weight people kept off, the more likely remission was (Total Diet Replacement and Type 2 Diabetes Remission).
- Which diet produces the deficit matters much less than whether it is produced and kept. Low-carb diets lower blood glucose and drug use at six months; by a year the advantage has faded or become uncertain. They have not been shown to add remission beyond the weight they take off (Carbohydrate Restriction and Type 2 Diabetes Remission).
- In diagnosed diabetes, GLP-1 drugs carry high-certainty evidence on heart attacks and death, and moderate-certainty evidence on kidney failure; lifestyle weight loss has not shown a heart-event benefit there. The drugs have not been shown to deliver the drug-free remission that weight loss can (GLP-1 Receptor Agonists and Cardiovascular and Kidney Outcomes).
Start with prevention, where the evidence is oldest and strongest.
An intensive lifestyle programme cuts the rate of new diabetes by a third to a half
Three randomized trials, on three continents, put adults with prediabetes through a structured programme of diet change, weight loss and physical activity. All three cut new diagnoses of diabetes by between roughly a third and a bit over half compared with usual advice, over three to six years (Lifestyle vs Metformin for Diabetes Prevention).
What the programme was
The US Diabetes Prevention Program (DPP) enrolled 3,234 adults with both raised fasting and raised post-load glucose (mean age 51, mean BMI 34.0, 68% women, 45% from minority groups). The lifestyle arm aimed for at least 7% weight loss and at least 150 minutes of activity a week, delivered through a 16-lesson curriculum with one-to-one case managers and a low-calorie, low-fat diet. Comparators were metformin and placebo; average follow-up was 2.8 years (Knowler, 2002).
The Finnish Diabetes Prevention Study randomized 522 overweight, middle-aged adults with impaired glucose tolerance. Its goals were a weight loss of 5% or more, fat below 30% of energy, fibre of at least 15 g per 1000 kcal, and at least 30 minutes of moderate exercise a day (Tuomilehto et al., 2001). The Da Qing study in China randomized 33 clinics, not individuals, covering 577 people with impaired glucose tolerance, to diet, exercise, diet plus exercise, or control, and followed them for six years (Pan et al., 1997).
Three trials, one direction, different sizes
Each trial reported a relative reduction in new diabetes against a usual-care or placebo comparator. The numbers are comparable in kind but not identical:
| Trial | Follow-up | Reduction in new diabetes vs control |
|---|---|---|
| US DPP | mean 2.8 yr | 58% (95% CI 48-66) |
| Finnish DPS | mean 3.2 yr | 58% (HR 0.4, 95% CI 0.3-0.7) |
| Da Qing | 6 yr | 31% diet, 46% exercise, 42% diet plus exercise |
(Knowler, 2002) (Tuomilehto et al., 2001) (Pan et al., 1997)
The defensible summary is a 31-58% reduction, consistent in direction, not 58% everywhere. Da Qing randomized clinics, and each active arm beat the control clinics; its percentages come from an individual-level model adjusted for baseline differences in BMI and fasting glucose (Pan et al., 1997). Its six-year horizon also gives more of the control group time to progress, which shrinks a relative reduction. The trials are separate experiments on separate populations. They are not fully sealed off from each other: a Da Qing author sits in the DPP research group and DPP’s lead author helped plan the Finnish study. The cleanest independent pair is Finland and Da Qing (Lifestyle vs Metformin for Diabetes Prevention).
What it means in absolute terms
In DPP, new diabetes ran at 11.0, 7.8 and 4.8 cases per 100 person-years on placebo, metformin and lifestyle. By three years, 28.9% of the placebo group had diabetes, against 21.7% on metformin and 14.4% on lifestyle. About 7 people (6.9; 95% CI 5.4-9.5) had to follow the programme for three years to prevent one case, against 13.9 for metformin (Knowler, 2002).
Those absolute numbers belong to a high-risk group. DPP selected people with three strong risk factors and tested their glucose often, which the authors suggest may explain why its placebo rate «was higher than we had anticipated and was higher than the incidence in observational studies» (Knowler, 2002). Someone at lower baseline risk can expect a similar relative cut but fewer cases prevented per person treated — absolute benefit scales with starting risk (Baseline Risk and the Relative-Absolute Split).
The other trials give their own absolute frames, each on its own clock. In Finland, cumulative incidence at four years was 11% (95% CI 6-15) with the programme against 23% (17-29) without, and the authors put the number needed to treat at 5 people for five years (Tuomilehto et al., 2001). In Da Qing, 67.7% of the control group had diabetes at six years, against 41.1-46.0% in the active arms (Pan et al., 1997). These NNTs run over different horizons and should not be ranked against each other.
Who it works for
The lifestyle effect held across nearly every subgroup DPP tested. «The lifestyle intervention was highly effective in all subgroups.» It did not differ by sex or ethnic group and was at least as effective in older participants as in younger ones. Its one significant modifier was baseline post-load glucose, with a larger effect at lower levels (Knowler, 2002).
Da Qing extends the finding to leaner people. Its participants averaged a BMI of 25.8, and «The relative decrease in rate of development of diabetes in the active treatment groups was similar when subjects were stratified as lean or overweight (BMI < or >25 kg/m2).» Among the lean, the reduction reached significance in the exercise and combined arms but not the diet arm (Pan et al., 1997). Lifestyle worked across body sizes in these trials; metformin, as the drug section shows, did not (Baseline Risk and the Relative-Absolute Split).
What the 58% assumes
DPP’s result came from a programme people only partly kept up. Half the lifestyle group reached the 7% weight goal by 24 weeks; 38% were still there at their latest visit. The activity goal fell from 74% to 58% over the same span (Knowler, 2002). The 58% is therefore the effect of the programme as people actually followed it.
It is also the effect of a well-resourced programme. The investigators describe it as «systematic and intensive, with the study participants receiving detailed, individualized counseling» (Knowler, 2002). A thinner version — a leaflet, a single appointment — is a different exposure, likely weaker, and these trials do not estimate its effect (Lifestyle vs Metformin for Diabetes Prevention).
The programme also has a limit of scope. It delays the diagnosis; whether that prevents heart attacks or death is a separate question, taken up below.
Harms
Harms were modest. Musculoskeletal symptoms, mostly aches and joint pain, ran at 24.1 events per 100 person-years on lifestyle against 21.1 on placebo. No deaths were attributed to any study intervention (Knowler, 2002).
Which part of the programme did the work? Mostly the weight — but not only.
Weight loss does most of the work, and activity adds a partly separate effect
How much the weight explains
The DPP lifestyle arm lost about 5.6 kg on average (Knowler, 2002), and a later re-analysis of that arm alone (1,079 people, mean BMI 33.9, about 3.2 years) asked which part of the programme carried the prevention. Weight loss came out as the dominant predictor: «For every kilogram of weight loss, there was a 16% reduction in risk, adjusted for changes in diet and activity.» Per 5 kg lost, the hazard ratio was 0.42 (0.35-0.51) (Hamman et al., 2006). Weight change explained about 8% of the variation in who developed diabetes; self-reported fat intake and activity explained under 0.1% each, and only 0.15% and 0.10% even after correcting those measures for their likely imprecision (Hamman et al., 2006).
This is one trial’s cohort analysed again, not a fourth trial. It shows association within the lifestyle arm, and it does not add independent weight to the three-trial result in the previous section (Lifestyle vs Metformin for Diabetes Prevention). The per-kilogram figure is a hazard ratio on new diabetes. It is not the same quantity as the per-kilogram odds ratio for remission quoted later, and the two should not be compared.
Activity helps people who miss the weight target
People who missed the weight target but met the activity target still developed diabetes less often. In DPP, «Among 495 participants not meeting the weight loss goal at year 1, those who achieved the physical activity goal had 44% lower diabetes incidence». These people still lost a little weight, and the authors place the activity effect as independent of «the small weight loss (−2.9 kg) that occurred» (Hamman et al., 2006).
The Finnish trial found the same pattern in a separate population: among intervention participants who lost less than 5% of their weight, those exercising more than four hours a week had an odds ratio for diabetes of 0.2 (0.1-0.6) compared with those who stayed sedentary, and 0.3 (0.1-0.7) after adjusting for baseline BMI (Tuomilehto et al., 2001). Two separate trial populations agreeing makes this sub-finding sturdier than either alone, though the two trials are not fully independent (DPP’s lead author helped plan the Finnish study).
Both results come from comparisons inside the trial arms, not from randomizing activity, so people who kept up their exercise may differ in other ways. Hamman adds a further limit. Measured as a continuous self-reported quantity, activity had no effect independent of weight change; only meeting the defined goal (about 150 minutes a week) showed one (Hamman et al., 2006).
Diet and exercise together did not beat either alone
Da Qing is the only one of the three trials that randomized the components separately (by clinic). In its adjusted results (the table above), the combined arm did no better than exercise alone. The authors read this as «the efficacy of diet was similar to that of exercise, and there was no additional benefit of combining the interventions» (Pan et al., 1997). The same paragraph notes that, by intention to treat, the three active arms did not differ significantly from one another. So the trial shows no added benefit from stacking, not proof that stacking is useless.
Read together, weight loss is the main route and the thing to aim for, and activity is a partly independent second channel. Activity is linked to lower risk in people who cannot reach the weight target, and it helps hold the weight once lost. In Da Qing, combining diet with exercise bought no extra reduction in new diabetes. from Lifestyle vs Metformin for Diabetes Prevention. Why a few kilograms matter so much is often explained by fat leaving the liver and pancreas. That is a hypothesis, set out in the remission sections and on Ectopic Fat and Depot-Specific Risk.
Strength training, sitting and sleep: smaller, observational signals
Beyond the trials, cohort data point to three further levers. All three rank well below weight loss and none has a randomized trial on new diabetes behind it.
- Muscle-strengthening activity. Any versus none was associated with a 17% lower incidence of diabetes, RR 0.83 (0.77-0.89), adjusted for aerobic activity, from 5 studies at GRADE low (Momma et al., 2022). Risk fell steeply up to about 60 minutes a week and more gradually after that, with no upturn over the volumes studied (Muscle-Strengthening Activity and Mortality).
- TV time. Each extra hour a day of TV was associated with an RR of 1.09 (1.07-1.12) for diabetes, after adjustment for physical activity. Total sitting showed almost nothing, 1.01 (1.00-1.01) (Patterson et al., 2018). Reverse causation is the main caveat: «An estimated 27% of those with the condition have no formal diagnosis», so undiagnosed diabetes may come first (Patterson et al., 2018). The authors’ 29% attributable fraction assumes causation (Sedentary Behaviour and Chronic Disease Risk).
- Sleep. Below 7 hours, each hour less carried an RR of 1.09 (1.04-1.15), about 68 extra cases per 100,000 people a year at US incidence (Shan et al., 2015). Sleep-restriction experiments support the short-sleep arm, though on glucose tolerance and insulin sensitivity rather than new diabetes. The long-sleep arm has not been tested against illness as the cause, so it stays unadjudicated (Sleep and Metabolic Health).
If the weight carries most of the effect, the next question is whether what you eat matters beyond how much.
Food choices shift diabetes risk at the margins
Some foods track lower diabetes risk and others higher, and the pattern is consistent across large cohort pooling. But every estimate in this section comes from observational cohorts, and no trial held here has tested a food swap against new diabetes. The relative risks per serving are modest next to the 31-58% reduction that the weight-loss programme produced in trials. Food Groups and Health Outcomes - A Dose-Response Matrix
The food-group map
A dose-response meta-analysis of prospective cohorts gives each food group a risk ratio per daily serving increment, graded for certainty with NutriGrade (a nutrition-specific rating of how trustworthy the pooled evidence is). (Schwingshackl et al., 2017)
| Food (daily increment) | Risk of new diabetes, RR (95% CI) | Certainty |
|---|---|---|
| Whole grains (30 g) | 0.87 (0.82-0.93) | High |
| Dairy (200 g) | 0.97 (0.94-0.99) | Moderate |
| Refined grains (30 g) | 1.01 (0.99-1.03) | Moderate |
| Red meat (100 g) | 1.17 (1.08-1.26) | High |
| Sugar-sweetened beverages (250 mL) | 1.21 (1.12-1.31) | High |
| Processed meat (50 g) | 1.37 (1.22-1.55) | High |
Whole grains, red meat and processed meat (Schwingshackl et al., 2017); sugar-sweetened beverages (Schwingshackl et al., 2017).
Refined grains sit at the null; whole grains carry the protective signal. The whole-grain curve flattens: the authors found little further risk reduction beyond about 2 servings (60 g) a day, where the RR was 0.75 (0.71-0.79); they give no interval on where the curve flattens. Sugar-sweetened drinks show no safe floor within the data: risk rose throughout the studied range of 0-748 mL a day. (Schwingshackl et al., 2017) A later pooling of 19 cohorts found almost the same slope, RR 1.19 (1.13-1.25) per 250 mL a day, and a straight-line shape. (Qin et al., 2020) One plausible route is extra calories the body does not offset, which ties back to weight (Free Sugars Intake).
Fibre points the same way: people in the highest intake group had a lower risk of diabetes than those in the lowest, RR 0.84 (0.78-0.90), at moderate certainty. (Reynolds et al., 2019) A high glycaemic index diet tracked higher risk, RR 1.27 (1.21-1.34) highest vs lowest, graded low certainty by GRADE and moderate by NutriGrade. Only 12 of the 69 glycaemic-index analyses adjusted for fibre, though in those that did, the authors report that adjusting made little difference (Jenkins et al., 2024). Glycaemic index may partly stand in for the fibre and whole-grain pattern, since the three markers move together in real diets (Glycaemic Index and Glycaemic Load and Chronic Disease, Dietary Fibre and Health, Whole Grains Refined Grains and Pulses).
Why these numbers stay at the margin
Food-frequency questionnaires measure intake poorly, and people who eat whole grains and avoid processed meat also tend to exercise, smoke less and follow medical advice. A high NutriGrade rating certifies that the association is consistent and precise; it does not rule out that healthier people are doing the eating. The red-meat signal also shifts with body weight: one Shanghai cohort found lower risk among normal-weight women who ate more red meat and higher risk among obese women. (Schwingshackl et al., 2017) Artificially sweetened drinks show a near-identical association (RR 1.15 per 250 mL a day). The authors read this as possible harm and advise limiting these drinks too (Qin et al., 2020); people at risk switching to diet drinks could also produce it. (Non-Sugar Sweeteners)
One genetic study (Mendelian randomization, which uses gene variants as a natural experiment) found no significant processed-meat effect on diabetes, OR 1.090 (0.714-1.664). Its interval includes no effect and is wide enough to hold a cohort-sized harm (the two estimates sit on different scales), so the genetic test neither confirms nor clears the meat association. (Li et al., 2024)
To put a ratio in absolute terms: for someone with a 10% chance of developing diabetes over ten years, an RR of 1.37 would mean about 14 cases per 100 instead of 10. That holds only if the whole association were causal, which these studies cannot show.
Several smaller levers point in the same direction, each observational: Coffee Consumption and Health (high vs low intake RR 0.70, though the genetic studies its review cites found no causal signal), Dietary Magnesium and Cardiometabolic Outcomes (RR 0.81), Ultra-Processed Food and Health Outcomes (RR 1.12 per 10% of the diet), Dairy and Cardiometabolic Health and Heme Iron and Cardiometabolic Risk. High blood levels of branched-chain amino acids predict diabetes but mostly reflect insulin resistance, so they give no reason to cut protein (Does Circulating BCAA Cause Type 2 Diabetes or Merely Predict It).
Heme iron, the form of iron in meat, has a genetic check that points the same way. Highest against lowest dietary heme-iron intake carried an RR of 1.31 (1.21-1.43) (Zhao et al., 2012). A Mendelian-randomization study found 7% higher diabetes risk per SD of transferrin saturation, though not for other iron markers (Liu et al., 2024). The genetic test tracks body iron, not dietary heme as such, so it supports the direction, not the size .
Food choice is worth getting right, but it ranks well below weight loss for preventing diabetes.
Metformin prevents less, and mainly in heavier, more hyperglycaemic people
Where an intensive lifestyle programme is not available, or not enough, metformin is the realistic alternative. Only DPP, described above, tested the two head to head; its metformin arm took 850 mg twice daily. Against placebo, metformin cut new diabetes by 31% (95% CI 17 to 43) and the programme by 58% (48 to 66); the programme beat metformin directly by 39% (24 to 51) (Knowler, 2002). The Finnish and Chinese prevention trials had no metformin arm, so this comparison rests on DPP alone (Lifestyle vs Metformin for Diabetes Prevention).
Who metformin helps
The 31% average hides a steep gradient. Metformin did little in people with a BMI under 30 or a fasting glucose under 110 mg/dl, and roughly halved new diabetes at BMI 35 or more and at fasting glucose 110-125 mg/dl (Knowler, 2002):
| DPP subgroup | Placebo | Metformin | Lifestyle | Metformin vs placebo (95% CI) |
|---|---|---|---|---|
| BMI 22 to <30 | 9.0 | 8.8 | 3.3 | 3% (-36 to 30) |
| BMI 30 to <35 | 8.9 | 7.6 | 3.7 | 16% (-19 to 41) |
| BMI 35 or more | 14.3 | 7.0 | 7.3 | 53% (36 to 65) |
| Fasting glucose 95-109 mg/dl | 6.4 | 5.5 | 2.9 | 15% (-12 to 36) |
| Fasting glucose 110-125 mg/dl | 22.3 | 12.3 | 8.8 | 48% (33 to 60) |
New diabetes per 100 person-years; last column is the relative reduction.
In the heaviest group the two arms finished level: lifestyle was no better than metformin at BMI 35 or more (-4%, -47 to 26). Lifestyle stayed effective in every subgroup, so its lead over the drug was largest where metformin was weakest: «The ad- vantage of the lifestyle intervention over metformin was greater in older persons and those with a lower body-mass index than in younger persons and those with a higher body-mass index.» At 60 or older, metformin cut new diabetes by 11% (-33 to 41), and the programme beat it by 69% (47 to 82) (Knowler, 2002).
How far to trust the split
This is positive evidence that the drug’s relative effect changes with BMI and fasting glucose — the bar for treating a subgroup differently, not just the larger absolute benefit that higher baseline risk always brings (Baseline Risk and the Relative-Absolute Split). The heterogeneity tests were significant, and «Neither interaction was explained by the other variable or by age.» But it comes from one trial, and the authors flag its limits: «The study had inadequate power to assess the significance of effects within the subgroups, nor were such tests planned.» (Knowler, 2002). Read it as plausible but unreplicated, and imprecisely sized: metformin earns its place in a heavier person with fasting glucose near the diabetic range, and buys little for a lean person with near-normal fasting glucose. Which drug, at what dose, is a prescriber’s decision.
Why the drug does less, and what it costs
Metformin and the programme lowered fasting glucose by similar amounts, but the programme lowered HbA1c further and left more people with normal glucose after a glucose drink. That fits what metformin does: it mainly suppresses glucose output from the liver, which sets fasting levels (Knowler, 2002). The mechanism reading is the trial’s own, not separately tested. DPP required both fasting and post-load glucose to be raised, so the results for people with only one of the two are, in the authors’ words, «likely but unknown» (Knowler, 2002).
Gastrointestinal symptoms ran at 77.8 events per 100 person-years on metformin, against 30.7 on placebo and 12.9 on lifestyle. Metformin’s weight effect is small: 2.1 kg in DPP, against 5.6 kg on the programme (Knowler, 2002), and -1.9% of body weight (-2.5 to -1.2) beyond lifestyle alone in a network meta-analysis of obesity drugs, which classed metformin as «not convincingly different from lifestyle modification alone» (Nong et al., 2026) (Comparing Obesity Drugs).
Other drugs that move the diagnosis
Semaglutide, one of the GLP-1 drugs (GLP-1 receptor agonists), delays diabetes by a wide margin in people without it. In SELECT — adults 45 or older with a BMI of at least 27 and established heart disease, two-thirds with prediabetes — 3.5% on semaglutide versus 12.0% on placebo reached an HbA1c of 6.5% or more over a mean 39.8 months (HR 0.27, 0.24 to 0.31; about 8-9 fewer per 100 people). That was a supportive secondary end point, not corrected for multiple testing (Lincoff et al., 2023). In STEP-1, adults with obesity (or overweight plus a weight-related condition) and no diabetes, 84.1% of those with prediabetes reverted to normal glucose at 68 weeks versus 47.8% on placebo, an exploratory end point (Wilding et al., 2021).
The gain shrinks once the drug stops. In 327 STEP-1 participants followed for a year off treatment, reversion among those with prediabetes fell from 93.6% versus 41.5% at week 68 to «43.3% versus 34.0%» at week 120 (Wilding et al., 2022). Both trials measured a glucose threshold, not a hard outcome; neither compared the drug with an intensive lifestyle programme; and SELECT’s population had established heart disease. These numbers do not show semaglutide beating lifestyle — they show a strong, maintenance-dependent effect on the same diagnosis (Semaglutide for Cardiovascular Risk in Obesity; the drug class in full: GLP-1 Drugs).
Vitamin D sits far behind. A pooled analysis of individual data from three prevention trials in prediabetes found a hazard ratio of 0.85 (0.75 to 0.96) over a median of about three years, «the number needed to treat was 30» — against about 7 for the DPP programme and 14 for metformin (Pittas et al., 2023). The effect appeared only below a BMI of 31.3 in the trials that used cholecalciferol. On these numbers it is at most a low-cost add-on, not a substitute for either (Vitamin and Mineral Supplements for Disease Prevention).
Statins nudge some people over the diabetes line
Statins raise new diabetes slightly. In blinded trials, low- or moderate-intensity statins raised it by 10% (RR 1.10, 1.04 to 1.16; about 1 extra case per 1,000 people per year, +0.12%/yr) and high-intensity statins by 36% (RR 1.36, 1.25 to 1.48). The excess falls on people already near the cut-off: «approximately 62% of new-onset diabetes cases were among participants who were already in the top quarter of the baseline distribution» — the prediabetic group this section is about. The trialists judged that «The diabetes-related risks arising from the small changes in glycaemia resulting from statin therapy are greatly outweighed by the benefits of statins on major vascular events when the direct clinical consequences of these outcomes are taken into consideration» (Reith et al., 2024). On that reading, the diabetes risk does not change whether someone with a heart-risk indication should take a statin (Statins and New-Onset Diabetes).
All of these move the diagnosis. Whether that prevents what the diagnosis predicts is a separate question.
Preventing the diagnosis has not been shown to prevent heart attacks
The Diabetes Prevention Program’s trial group kept following its three arms for a median of 21 years (the DPP Outcomes Study, DPPOS) to ask the next question: did delaying diabetes also prevent heart attacks, strokes and cardiovascular deaths? It did not. «Neither metformin nor lifestyle intervention reduced the primary outcome: metformin versus placebo hazard ratio 1.03 (95% CI, 0.78–1.37; P = 0.81) and lifestyle versus placebo hazard ratio 1.14 (95% CI, 0.87–1.50; P = 0.34).» In absolute terms, the placebo group had 5.28 first major cardiovascular events per 1,000 person-years, the metformin group 5.51 and the lifestyle group 6.10, across 310 first events among 3,234 middle-aged, ethnically diverse adults with prediabetes. (Goldberg et al., 2022)
A bounded null, not proof of no effect
The trial was not a clean test by the time the events accrued, and its authors say so. When the original trial ended, every participant was offered group lifestyle sessions; study metformin use fell «from 77% at the end of DPP to 41%» while out-of-study metformin rose in all arms; and family doctors put 68–74% on blood-pressure drugs and 56–62% on statins. Statin use was significantly lower in the lifestyle arm than on placebo (56% vs 62%). The authors suggest these factors «may have diluted the effects of the interventions», although adjusting for medication use did not change the results.
The authors also call the result «a relatively low-risk cohort from the standpoint of the prevention of CVD»: mean HbA1c was 6.0–6.1% at 21 years. (Goldberg et al., 2022) So the confidence interval rules out a large benefit in this setting, not a modest one: the lifestyle estimate runs from a 13% reduction to a 50% increase.
Higher-risk people may be different
The DPPOS authors place their null beside the Finnish study’s 10-year and Da Qing’s 20-year follow-ups, which also found no cardiovascular effect. Only a longer, 30-year Da Qing follow-up showed a lifestyle benefit on major cardiovascular events, «hazard ratio 0.74 [95% CI, 0.59–0.92]», in what they describe as «a higher-risk population with a greater proportion of smokers, a higher prevalence of hypertension, and diabetes with more severe hyperglycemia». (Goldberg et al., 2022) These results need no contradiction to coexist. The fewer heart attacks a group is heading for, the fewer a programme can remove, so the same relative effect can be invisible in a well-treated, low-risk cohort and detectable in a high-risk one — the point worked on Baseline Risk and the Relative-Absolute Split. The 30-year Da Qing report itself is not held here; its numbers come through the DPPOS paper.
Metformin’s selective benefit did not carry over
Metformin’s prevention effect on diabetes was concentrated in heavier, more hyperglycaemic people (the previous section). On cardiovascular events the drug had no overall effect, and the prespecified subgroups «showed no significant heterogeneity by age, sex, race/ ethnicity, or diabetes development for either metformin or lifestyle». (Goldberg et al., 2022) The subgroups tested differed from those that carried the diabetes result (BMI and fasting glucose), so this is not the same modifier found null. A drug that helps one group on one outcome has to earn that claim again on the next outcome. Baseline Risk and the Relative-Absolute Split
What this changes
New diabetes is a diagnosis, nearer to what people care about than a lab value, yet in DPPOS it still did not translate into fewer events over two decades — a surrogate disconnect one rung above the cholesterol cases on Surrogate Outcomes. Prevention remains worth doing for the diagnosis it delays; the DPPOS authors report that the diabetes-prevention benefit itself lasted as long as 15 years. (Goldberg et al., 2022) For someone with prediabetes whose concern is heart disease, the direct levers on LDL and blood pressure, heavily used in this very cohort, are the ones to look to. The primary DPPOS report on long-term diabetes incidence is not held here, so how durable the diagnosis-level benefit is rests on the authors’ summary alone. Lifestyle vs Metformin for Diabetes Prevention
For someone already diagnosed, the question changes from delay to reversal.
Losing 10-15 kg soon after diagnosis can put diabetes into remission
What counts as remission
Every remission rate below uses one definition: HbA1c below 6.5% (48 mmol/mol) after at least two months off all glucose-lowering drugs. The DiRECT trial set it as a co-primary outcome: «remission of diabetes, defined as glycated haemoglobin (HbA1c) of less than 6·5% (<48 mmol/mol) after at least 2 months off all antidiabetic medications, from baseline to 12 months» (Lean et al., 2018). A normal HbA1c reached while still on a drug does not count. Studies that use looser definitions appear later in this guide, labelled as such (Total Diet Replacement and Type 2 Diabetes Remission).
The DiRECT trial
DiRECT randomised 49 UK primary-care practices, not individual patients, and enrolled 298 people aged 20-65, diagnosed within the past six years, with a BMI of 27-45 and not on insulin. Nearly all were white (98-99% in each arm). The intervention arm stopped all diabetes and blood-pressure tablets on day one; the control arm received best-practice care and kept its drugs (Lean et al., 2018).
The programme had three phases. A formula total diet replacement (TDR) of 825-853 kcal a day replaced all food for 3 months (up to 5). Its macronutrients were 59% carbohydrate, 13% fat and 26% protein, so this was a low-energy diet, not a low-carb one. Food came back in steps over 2-8 weeks, followed by monthly visits to hold the weight off: «All oral antidiabetic and antihypertensive drugs were discontinued on day 1 of the weight management programme, with standard protocols for drug reintroduction under national clinical guidelines, if indicated by regular monitoring of blood glucose and blood pressure.» (Lean et al., 2018)
What happened at one year
At 12 months, 46 of every 100 people started on the programme were in remission, against 4 of every 100 on usual care — about 42 more per 100. The trial reported «Diabetes remission was achieved in 68 (46%) participants in the intervention group and six (4%) participants in the control group (odds ratio 19·7, 95% CI 7·8–49·8; p<0·0001).» (Lean et al., 2018) Average weight fell 10.0 kg on the programme and 1.0 kg on usual care. Quality of life (EQ-5D visual analogue scale) rose 6.4 points more (95% CI 2.5-10.3), and 109 of the intervention arm (73.6%) took no oral diabetes drug at 12 months, against 18.2% of controls (Lean et al., 2018).
Not everyone stayed the course. In the intervention arm, 6 people (4%) never engaged and 26 (17%) withdrew from treatment during the first year; all were kept in the intention-to-treat analysis. Weight fell 14.5 kg (95% CI 13.4-15.5) during the formula phase, then crept back by 1.0 kg during food reintroduction and 1.9 kg during maintenance (Lean et al., 2018).
The comparison is the whole programme against usual care, and it is tilted. Only the intervention arm had its drugs withdrawn, and a control patient with a good HbA1c on tablets counts as a non-remitter under the definition. The between-arm gap is therefore larger than a comparison with drugs handled equally in both arms would show (Total Diet Replacement and Type 2 Diabetes Remission).
More weight off, more remission
Remission tracked the kilograms kept off. Pooling both arms at 12 months, «Remission varied with weight loss in the whole study population, with achievement in none of 76 participants who gained weight, six (7%) of 89 participants who maintained 0–5 kg weight loss, 19 (34%) of 56 participants with 5–10 kg loss, 16» (Lean et al., 2018) (57%) of 28 with 10-15 kg loss, and 31 (86%) of 36 who lost 15 kg or more. Over the studied range — weight gain to 15 kg or more lost — the rate rose at every step, with no plateau shown; above 15 kg the band is open-ended.
These band rates are not what one person should expect from starting the programme. They sort people by the weight they ended up losing, which is not a randomised contrast: those who lost the most may differ in ways that also favour remission. The trial’s randomised answer is the 46% vs 4% above. from Total Diet Replacement and Type 2 Diabetes Remission.
A separate design points the same way. In ADDITION-Cambridge, a cohort of 867 newly diagnosed people followed for five years, 30% were in remission, and «Every 1 kg of weight loss was associated with 7% higher chance of remission at 5 years, regardless of specific diet regimens or lifestyle interventions» (Churuangsuk et al., 2021). This supports the direction, not the size, and it is not fully independent: it reaches this guide through a review co-authored by DiRECT’s lead investigator, and it shares the same responder problem. This per-kilogram figure measures remission, not new diabetes, so it does not combine with the prevention trials’ per-kilogram estimate.
How firm the finding is
Formula TDR is the best-supported route to remission held here. An umbrella review graded it the only high-certainty cell in its remission map: two low-risk-of-bias trials (DiRECT and DIADEM-I, 445 people) gave a median 54% remission at one year (range 46-61%) against 4-12% on standard care. Formula meal replacement, Mediterranean and ketogenic diets sat at moderate to very-low certainty (Churuangsuk et al., 2021). The review’s own reading: «The main contributor to HbA1c reduction and remission appears to be weight loss, irrespective of diet type.» (Churuangsuk et al., 2021) (Diets for Weight Management in Type 2 Diabetes).
The explanation favoured by DiRECT’s investigators is that weight loss drains fat from the liver and pancreas, letting insulin action and insulin secretion recover. Taylor puts it as «people with recent onset T2DM could regain normal glucose control and normal β-cell function when the fat content of the liver and the pancreas was decreased by a weight loss diet- ary regimen» (Taylor & Holman, 2014). The idea that each person has a personal fat threshold remains a hypothesis under test, not an established quantity (Ectopic Fat and Depot-Specific Risk).
The same liver fat links the two decisions in this guide. Fatty liver diagnosed on imaging is linked to nearly twice the risk of new diabetes (HR 1.85, 95% CI 1.47-2.22, in the fully adjusted studies) (Mantovani et al., 2018), and weight loss is the shared target for both (Fatty Liver MASLD and Weight Loss).
Who else it may apply to
None of the points below shows the programme works better in one group; they say who is more likely to remit, or where the trial evidence stops.
- Earlier is better. The review concludes «remission should be attempted as early as possible from diabetes diagnosis» (Churuangsuk et al., 2021). The highest rates, up to 75%, came from single-arm studies in people with diabetes for under two years — uncontrolled, so not a randomised estimate. Within DiRECT’s six-year window, duration’s effect was imprecise; the next section covers it.
- Normal-weight diabetes is outside the trial. DiRECT required a BMI of 27 or more, yet 36% of newly diagnosed UKPDS patients had a BMI under 25. In UKPDS’s 3-4-month diet-only run-in, among people presenting with plasma glucose of 8-10 mmol/l, a normal fasting glucose (below 6.0 mmol/l) came with a mean loss of 13% of body weight in those of normal weight, against 21% across the whole cohort (Taylor & Holman, 2014). That is fasting normoglycaemia on diet, not drug-free remission. The mechanism suggests the effect may transfer; no trial has tested remission in this group.
- Ethnicity is untested in trials. Pooled studies, mostly non-randomised, reported Asian participants more likely to remit; this came from an exploration of heterogeneity, not a randomised comparison, and DiRECT’s authors warn against «unqualified extrapolation» to groups such as South Asians (Churuangsuk et al., 2021) (Lean et al., 2018).
- Insulin users and longer-standing diabetes were excluded from DiRECT; the evidence here does not cover them.
Side effects, conflicts and the guideline
Side effects during the formula phase were common and mostly mild or moderate: constipation, feeling cold, headache and dizziness. Nine serious adverse events occurred in 7 of 157 intervention participants; two, in the same person (biliary colic and abdominal pain), were possibly related, and none led to withdrawal (Lean et al., 2018). Blood pressure falls during rapid weight loss, so continuing blood-pressure or glucose-lowering drugs risks dizziness on standing or low blood sugar (Churuangsuk et al., 2021). Drug changes belong with the prescribing clinician.
Diabetes UK funded DiRECT; Cambridge Weight Plan donated the formula, and several authors, including the lead author, reported payments from Counterweight or the formula maker. The paper states that neither Diabetes UK nor Cambridge Weight Plan had any input into design, analysis or interpretation. It makes no such statement about Counterweight, whose programme was the intervention and which employed two authors during the study (Lean et al., 2018).
NICE makes a weak recommendation to consider a low-energy TDR (800-1200 kcal a day) for overweight adults with type 2 diabetes. It caps the formula phase at 12 weeks and places it within long-term specialist support (Diets for Weight Loss - What NICE Recommends). The committee extended it «irrespective of diabetes duration», noting it was not aware of wider evidence on duration and remission — a judgement beyond the trial data (National Institute for Health and Care Excellence, n.d.).
A one-year rate says what is possible; the two-year data say what lasts.
Remission lasts about as long as the weight stays off
The two-year result
DiRECT followed the same 298 people for a second year. At 24 months, 53 of 149 people started on the programme (35.6%) were in remission, against 5 of 149 on usual care (3.4%), on the same definition as before: HbA1c below 6.5% with at least two months off all glucose-lowering drugs. The adjusted odds ratio was 25.8 (95% CI 8.3-80.8). These two-year figures come from a pre-publication draft of the trial report (Draft 7, dated 7 February 2019, deposited in the university repository), which is the version held here, not the published journal text. (M. E. J. Lean et al., 2019)
In absolute terms the gap is about 32 percentage points, so roughly 3 people need to start the programme for one extra person to be in remission at two years, in this stratum (age 20-65, BMI 27-45, diabetes under six years, not on insulin, UK primary care). That number counts drug-free remission at two years. It cannot be set beside the NNTs for preventing diabetes, which count a different outcome over different spans. (M. E. J. Lean et al., 2019)
The counts are the stable part of that result. With only five remitters in the control arm, the odds ratio swings with the model: 25.8 in the main analysis, 15.87 in the trial’s own figure. (M. E. J. Lean et al., 2019) The intervention arm also fell from 46% at one year to 36% at two, roughly a fifth fewer remissions. Total Diet Replacement and Type 2 Diabetes Remission
Remission tracked the kilograms kept off
Across both arms pooled, remission at two years still rose with the weight people had kept off: 5.2% of those under 5 kg down, 28.8% at 5-10 kg, 60% at 10-15 kg and 70% (14 of 20) at 15 kg or more. Each extra kilogram lost carried an odds ratio of 1.25 (1.16-1.35) for remission at two years, against 1.32 (1.23-1.41) at one. (M. E. J. Lean et al., 2019)
These bands describe an association, not proof that holding the weight causes the remission. They pool both arms and compare people who differ in more than weight: someone who keeps 15 kg off may also differ biologically from someone who regains. The 70% at 15 kg is also not a fall from the one-year 86% in the same people, because each visit re-sorts people into bands. And neither figure is what a person starting the programme should expect; the ITT rate of about one in three is that number. -> Total Diet Replacement and Type 2 Diabetes Remission
The people who relapsed point the same way, but weakly. Among 63 remitters with weight at both visits, the 48 who stayed in remission regained 4.25 kg in year two, the 15 who relapsed 7.09 kg (p=0.073, short of conventional significance). (M. E. J. Lean et al., 2019) Weight was not the whole story either: 4 of the 50 people who gained weight stayed in remission at both visits, all with small gains and a baseline HbA1c just above the threshold (6.5-6.63%). (M. E. J. Lean et al., 2019)
Keeping the weight off took repeated support
The two-year rate is the rate for a supported programme, not for a diet done once. About half the intervention arm needed at least one rescue phase: anyone who regained more than 2 kg was offered 2-4 weeks of partial meal replacement, and more than 4 kg a fresh round of total diet replacement and food reintroduction, with the offer of orlistat. Of 143 people with data, 71 (49.7%) never needed one, 49 had one, 15 had two and 8 had three or more. (M. E. J. Lean et al., 2019)
People attended an average of 7.7 of the 12 monthly visits in year two, and the intervention arm regained 2.6 kg on average between 12 and 24 months. (M. E. J. Lean et al., 2019) Whether remission holds as well without monthly visits and rescue is untested.
Who remitted
Within DiRECT’s range, baseline BMI did not predict two-year remission (OR 0.99 per kg/m2, 0.92-1.06). The authors report that diabetes duration did not influence remission (OR 0.92 per year, 0.76-1.11); the interval is wide enough that this is insufficient evidence rather than a demonstrated null. Men and older participants remitted more often. (M. E. J. Lean et al., 2019) These are post-hoc associations across both arms. They say who was more likely to remit, not that the programme works better in men or older people, and they are no basis for turning anyone away.
The biases run in opposite directions for different numbers. The ITT comparison is likely biased against the programme: 9 control participants lost more than 10 kg on their own in year two, which the authors say considerable media coverage of the first-year results may have encouraged, and more intervention participants missed the two-year visit (77.9% vs 94.6% attended). Those with no data from the visit or from GP records were counted as not in remission. (M. E. J. Lean et al., 2019) The weight-band and relapser analyses lean the other way, because they condition on who stayed and how much they lost.
Why regain is the expected path
Regain is the norm after diet-induced weight loss, and the body appears to defend the higher weight. In a single-arm study of non-diabetic adults, analysed in those who completed it, the appetite hormones that push towards eating were still shifted a year after an eight-week very-low-energy diet: «One year after initial weight reduction, levels of the circulating mediators of appetite that encourage weight regain after diet-induced weight loss do not revert to the levels recorded before weight loss.» (Sumithran et al., 2011) No held evidence shows that holding the lower weight for longer wears that defence down. -> Weight-Loss Maintenance and Metabolic Adaptation, Mitigations of the Defended Setpoint
That makes a maintenance plan part of the treatment, not an optional extra. Whether a failed attempt adds diabetes risk is unsettled. In a meta-analysis of 14 cohorts, people with a history of weight cycling had a higher risk of developing diabetes overall (RR 1.23, 1.07-1.41). In people with obesity (BMI 30 or more) the association was not significant (RR 1.47, 0.96-2.24, three reports with high heterogeneity), though it was in the overweight range (RR 1.62, 1.31-2.00). The authors leave open that cycling may be a marker of illness rather than its cause. (Zou et al., 2020) -> Weight Cycling and Cardiometabolic Risk Those data are observational and concern new diabetes, not relapse after remission, so they do not settle the worry. On balance, regain is a reason to plan for maintenance rather than a demonstrated reason to skip the attempt.
What two years cannot tell you
Nothing beyond 24 months is held for this route. Total Diet Replacement and Type 2 Diabetes Remission
If the kilograms decide remission, does it matter which diet takes them off?
No diet composition beats energy control; low-carb is one route to the deficit
Which diet takes the weight off matters far less than whether it comes off and stays off. That holds for weight, for HbA1c and for remission, in people who already have type 2 diabetes. Low-carb diets do one thing better for a while: they lower blood glucose and drug use at six months, an edge that has faded or become uncertain by a year. They have not been shown to add remission beyond the weight they remove (Diets for Weight Management in Type 2 Diabetes, Carbohydrate Restriction and Type 2 Diabetes Remission).
For weight, the format that controls energy matters; the macronutrient mix does not
An umbrella review of 19 meta-analyses of weight-loss diets in type 2 diabetes, each graded for certainty, concluded that they «do not support any particular macronutrient profile or style over others. Very low energy diets and formula meal replacement appear the most effective approaches, generally providing less energy than self-administered food-based diets.» (Churuangsuk et al., 2021)
Low-carb against higher-carb came out null on weight in four high-quality meta-analyses (differences under 1-2.5 kg; GRADE low to high), a positive finding of little or no difference rather than a gap. High-protein, Mediterranean, high-MUFA, vegetarian and low-GI diets sat 0-2 kg from their controls, mostly on low-quality meta-analyses. A very-low-energy diet (400-500 kcal) beat a standard low-energy diet by 6.6 kg (95% CI 3.7-9.5) at three months, at low certainty. Formula meal replacement beat it by 2.4 kg (1.4-3.3) over 12-52 weeks, at moderate certainty. (Churuangsuk et al., 2021)
The less energy explanation is the review’s inference, not a tested pathway. Its planned analysis to separate energy restriction from diet type «proved impossible from the published information» (Churuangsuk et al., 2021). And the meal-replacement trials gave both arms about 1500 kcal, so the structure of the programme, and how well people kept to it, likely did part of the work. (Diets for Weight Management in Type 2 Diabetes)
HbA1c moved with the weight. Across the meta-analyses, «HbA1c reduction broadly followed weight loss, and differences between diet types assessed over 3–12 months were small», but the data could not «quantify weight loss- independent effects on HbA1c» (Churuangsuk et al., 2021). A glucose effect of composition, separate from weight, is unmeasured — not shown to be absent.
Low-carb lowers glucose and drug use at six months; by a year the edge fades or blurs
A meta-analysis of 23 RCTs (n=1357) in type 2 diabetes defined low-carb as under 26% of energy or under 130 g a day. Against comparator diets, mostly low-fat: (Goldenberg et al., 2021)
| Outcome | 6 months | 12 months | Certainty (6 -> 12 mo) |
|---|---|---|---|
| HbA1c | -0.47% (-0.60 to -0.34), 17 RCTs | -0.23% (-0.46 to 0.00), 8 RCTs | high -> moderate |
| Weight | -3.46 kg (-5.25 to -1.67), 18 RCTs | +0.29 kg (-1.02 to 1.60), 7 RCTs | moderate |
| Medication reduced | RD 0.24 (0.12 to 0.35), 7 RCTs | RD 0.33 (-0.00 to 0.66), 3 RCTs | moderate -> low |
| LDL cholesterol | +0.02 mmol/L | +0.14 mmol/L (-0.00 to 0.28), P=0.05 | high -> moderate |
The drug-use result is the outcome nearest the patient, though only at moderate certainty. Szczerba’s umbrella review carries the same estimate: a low-carb diet «reduced the use of drug treatments by an additional 24 per 100 individuals» (moderate certainty), and graded triglycerides -0.30 mmol/L (-0.43 to -0.17) at high certainty (Szczerba et al., 2023).
The HbA1c figure needs reading twice. The -0.47% is the whole effect of being on the diet, including the weight it takes off. How much of it survives without the weight loss is unmeasured (insufficient evidence, not a null), as the umbrella review’s caveat above says (Churuangsuk et al., 2021). Among the meta-analysis’s secondary markers, only triglycerides and insulin resistance cleared its threshold for a clinically important change (Goldenberg et al., 2021) (Carbohydrate Restriction and Type 2 Diabetes Remission).
Against a balanced-carbohydrate weight-loss diet, the gap shrinks further. A Cochrane review found, in people with type 2 diabetes at 12 months or more, weight 0.33 kg lower (2.13 lower to 1.46 higher; 7 RCTs, n=813) and HbA1c 0.14% lower (0.38 lower to 0.10 higher), both at moderate certainty (Naude et al., 2022). Its low-carb arm reaches up to 45% of energy, a looser cut-off than the 26% above (Low-Carbohydrate vs Balanced-Carbohydrate Diets). NICE found the very-low-carb weight advantage short-lived: «a reduction in weight at 1 year, but this was not sustained at 2 years» (National Institute for Health and Care Excellence, n.d.) (Diets for Weight Loss - What NICE Recommends).
Low-carb has not been shown to add remission
The answer depends on what counts as remission. The same meta-analysis reported two definitions, and they disagree: (Goldenberg et al., 2021)
| Definition | 6 months | 12 months |
|---|---|---|
| HbA1c below 6.5% with medication allowed | RD 0.32 (0.17 to 0.47), 8 RCTs, moderate | RD 0.10 (-0.02 to 0.21), 3 RCTs, moderate |
| HbA1c below 6.5% and off all diabetes drugs | RD 0.05 (-0.05 to 0.14), 5 RCTs, low | RD -0.04 (-0.16 to 0.09), 2 RCTs, low |
Only the second row is remission in the sense used elsewhere here, and it is null at both timepoints. The first row counts people still taking glucose-lowering drugs, so it is not a remission rate. One stratum differs: in trials that excluded insulin users, drug-free remission at six months was RD 0.20 (0.03 to 0.38), about 1 extra remission for every 5 people. That result exists only at six months; no 12-month split was reported, so it says nothing about durability. Only 7 of the 23 trials allowed medication to be reduced and reported usable data (Goldenberg et al., 2021).
No randomized trial has tested low-carb with remission as its primary outcome. The umbrella review’s words: «No RCT has evaluated LCDs/ketogenic diets for type 2 diabetes remission.» The often-cited ketogenic figure comes from a non-randomized study reporting 20% remission at one year (GRADE very low, serious risk of bias) (Churuangsuk et al., 2021).
The weight explains the remission, not the carbohydrate share
DiRECT’s remission result came on a formula diet that was not low-carb (its macronutrients are given above) (M. E. Lean et al., 2018). Low-carb’s remission signal and its weight advantage fade together between six and twelve months. Both facts fit one reading: carbohydrate restriction is not necessary for remission; it is one way to lose the weight that drives it. (Carbohydrate Restriction and Type 2 Diabetes Remission)
That reading does not exclude an extra effect of cutting carbohydrate. The meta-analysis checked whether calorie-matched trials behaved differently and «identified no evidence of credible effect modification based on caloric matching or lack thereof» (Goldenberg et al., 2021). But calories were matched by prescription and self-report, not by kilograms lost, so the test is weak. An added carbohydrate effect is null on a weak test — not excluded.
Composition does move specific markers once the deficit is in place
On weight, BMI and waist, formula meal replacement (and, at moderate certainty, a plant-based diet) stood out; most other diet types gave small or no differences. On glucose and lipids, diet type separates modestly. Low-carb lowers HbA1c and triglycerides; a Mediterranean diet lowers triglycerides; a low-GI diet lowers LDL; a high-protein diet lowers total and LDL cholesterol, each at high or moderate certainty (Szczerba et al., 2023). Once the deficit is in place, a person can pick the diet by which marker they most want to move. Both umbrella reviews draw on many of the same trials, so their agreement on weight is corroboration, not independent confirmation. (Diets for Weight Management in Type 2 Diabetes)
US and European guidance read the same evidence differently on very-low-carb
On the core the two bodies agree with each other and with the evidence above. The ADA finds «there is not an ideal percentage of calories from carbohydrate, protein, and fat for all people with or at risk for diabetes» (Evert et al., 2019). The EASD grades as high certainty that «Remission of type 2 diabetes (HbA1c <48 mmol/mol [<6.5%] without glucose-lowering medication) in people who are overweight or obese can be achieved through sustained weight loss» (Diabetes and Nutrition Study Group of the EASD, 2023). Both rest on the same trials this guide uses, so their agreement is shared evidence, not independent confirmation . They split on one point.
The ADA calls low- or very-low-carb eating «a viable approach» for «select adults with type 2 di- abetes not meeting glycemic targets or where reducing anti- glycemic medications is a priority» (Evert et al., 2019). The EASD’s diabetes nutrition group states: «Very low carbohydrate intakes, such as with ketogenic diets, are not recommended.» (graded moderate) (Diabetes and Nutrition Study Group of the EASD, 2023).
The two bodies read the same evidence base and set opposite population defaults. The EASD cites no observed benefit, safety concerns (higher LDL, hypoglycaemia, ketoacidosis, micronutrient gaps), poor long-term adherence and an observational link between low carbohydrate intake and premature death in the general population. The ADA treats the cardiovascular question as unsettled. Both accommodate the individual who chooses it. The EASD: «If people with diabetes them- selves choose to reduce carbohydrate to very low intakes, it is important this is done with health professional support» (Diabetes and Nutrition Study Group of the EASD, 2023). Neither bears on remission, which no trial has tested on low-carb. (Diets for Weight Management in Type 2 Diabetes)
The harms differ even where the weight loss does not
Diets that work equally well on weight carry different risks, so harms help decide among them. Ketogenic diets have been linked to thiamine deficiency, with heart failure and neurological problems reported, and to ketoacidosis, a risk raised by SGLT2-inhibitor drugs. Swapping carbohydrate for red or processed meat raises LDL (Churuangsuk et al., 2021).
A low-carb diet can also cost fibre. In people with diabetes or prediabetes, more fibre lowered HbA1c by 2.00 mmol/mol (95% CI 0.71-3.30; 33 trials, mostly 6-12 weeks) — about 0.18 percentage points — with weight 0.56 kg lower (A. N. Reynolds et al., 2020). That is small next to the effects of weight loss above. The EASD’s 35 g/day target marks the top of the cohort data rather than a located optimum. It rests on the same review, so the two count as one source, not two. For anyone eating low-carb, the review suggests fibre supplements as a stop-gap, with long-term benefit unproven. (Dietary Fibre and Health)
Small add-ons move HbA1c by tenths of a point
Once weight loss is under way, the remaining dietary and monitoring options move HbA1c by tenths of a point, measured only on surrogate outcomes. A continuous glucose monitor with feedback lowered HbA1c by 0.28% (95% CI 0.15-0.42), with high heterogeneity (I² 88%), mostly in diabetes trials (Richardson et al., 2024) (Continuous Glucose Monitoring as a Health Intervention). Cinnamon lowered it by about 0.10% (Cinnamon and Glycemic Control). Resistant starch gave a small standardized reduction (Resistant Starch and Glycemic Control). Ketone supplements have insufficient evidence on HbA1c (Exogenous Ketone Supplements and Blood Glucose), and free-living time-restricted eating left glycaemic markers unchanged (Time-Restricted Eating). Treating gum disease lowered HbA1c by about 0.43% over 3-4 months (Periodontitis and Systemic Health). None of these has a remission or hard-outcome result. (The Certainty-Importance Inversion)
Exercise sits beside the diet in established diabetes, with a narrower evidence base.
In diagnosed diabetes, exercise reliably raises fitness; its glucose edge is small
The one review held on exercise in diagnosed type 2 diabetes asks which kind, not whether. It pooled 13 randomized trials (345 adults, programmes of 11-16 weeks) comparing high-intensity interval training (HIIT) with moderate-intensity continuous training. HIIT raised peak oxygen uptake by 3.37 ml/kg/min more (95% CI 1.88 to 4.87), at moderate certainty. Higher fitness goes with lower mortality in cohort studies (Cardiorespiratory Fitness and Mortality), but this review measured no deaths or events, and whether a fitness gain from training carries that survival benefit is untested here. HbA1c fell 0.37 percentage points more on HIIT (-0.55 to -0.19), but at low certainty, while fasting glucose, fasting insulin and HOMA-IR did not differ between the two (J. Liu et al., 2018) (HIIT vs Continuous Training for Type 2 Diabetes).
The authors themselves note that an earlier meta-analysis found no significant HbA1c difference between interval and continuous training: «This is inconsistent with the findings of a meta-analysis conducted by Jelleyman et al. [18], which found that, while HIIT can reduce the levels of HbA1c in patients with diabetes and metabolic syn- drome, there is no significant difference in reduction ver- sus with continuous training.» (J. Liu et al., 2018) That review is not held. The review reported no adverse-event or adherence data, and its trials ran only 11-16 weeks. So choose the modality on preference and tolerability; neither has a firmly demonstrated glucose advantage (Is the Exercise Modality Doing Any Work, Measuring and Raising Cardiorespiratory Fitness).
The case for being active at all rests on the prevention trials and the general activity evidence (Lifestyle vs Metformin for Diabetes Prevention, Physical Activity Dose and Mortality), not on this comparison. No review held tests exercise against no exercise for glucose control in diagnosed diabetes; that is an open gap.
In diagnosed diabetes, the heart and kidney evidence comes from drugs, not lifestyle weight loss
Remission and glucose are what lifestyle demonstrably moves. Heart attacks and kidney failure are where the drugs have the evidence.
Look AHEAD: the weight came off, the events did not fall
Look AHEAD randomised 5,145 overweight or obese adults with type 2 diabetes to an intensive lifestyle programme (a goal of at least 7% weight loss) or to diabetes support and education. The programme arm lost more weight throughout: 8.6% vs 0.7% at one year, 6.0% vs 3.5% at the end. HbA1c fell further, and fitness and every cardiovascular risk factor except LDL cholesterol initially improved more. Yet heart attacks, strokes, cardiovascular deaths and hospitalised angina were no less common: 403 vs 418 people, 1.83 vs 1.92 per 100 person-years, HR 0.95 (95% CI 0.83-1.09). The trial stopped for futility at a median 9.6 years (Look AHEAD Research Group, 2013).
The null is bounded, not proof of no effect. The weight gap between arms shrank from about 8 percentage points to about 2.5, so the trial ended up testing a modest sustained difference. It was powered to detect an 18% reduction, so a smaller true benefit could pass unseen (Look AHEAD Research Group, 2013). Both arms’ physicians also adjusted medications as usual, which may narrow any gap between them (Does Weight Loss Reduce Cardiovascular Events).
The surrogates moved and the events still did not. In a Look AHEAD substudy, C-reactive protein fell 44.9% and 42.3% at one year in men and women taking statins plus the programme, against 13.7% and 21.0% on statins with usual care (Belalcazar et al., 2013). A bundle of better markers did not buy fewer cardiovascular events — the standing warning on Surrogate Outcomes.
Look AHEAD also reported at least partial remission in 11.5% of the programme arm vs 2% of controls. It counts prediabetes-level glucose as well as normal glucose, off medication for at least a year (Evert et al., 2019). That is a looser endpoint than the remission defined earlier in this guide, and not comparable with DiRECT’s rates.
A larger loss may matter — on one post-hoc analysis, not yet held
A post-hoc Look AHEAD analysis linked losing 10% of body weight in the first year to a 21% lower rate of cardiovascular outcomes over a median 10.2 years. This guide holds it only as relayed in DiRECT’s discussion (M. E. Lean et al., 2018). The primary paper is not held: — would show whether the >=10% signal survives its own analysis. Read it as a responder association, open to the usual bias: people who lose more weight may differ in other ways. It is reported, not proven, and it is still one analysis, however often it is cited Does Weight Loss Reduce Cardiovascular Events.
Across 54 trials, weight-loss diets lowered deaths; heart events did not clearly fall
A meta-analysis of 54 RCTs in 30,206 adults with obesity generalises the heart result. Weight-reducing diets did not measurably cut new cardiovascular events (RR 0.93, 95% CI 0.83-1.04). They did lower all-cause mortality: RR 0.82 (0.71-0.95), six fewer deaths per 1,000 (2 to 10 fewer) over a median two years, rated high quality (Ma et al., 2017). Look AHEAD carried 54.6% of the weight in the mortality analysis, but the benefit held without it: RR 0.78 (0.63-0.96) across the other 33 trials (Ma et al., 2017).
Two cautions travel with that number. The trials’ diets were «usually low in fat and saturated fat», so part of the benefit could run through LDL rather than weight itself; the trials cannot separate the two (Ma et al., 2017). And the trials mixed glycaemic status, from people with no comorbidity to people with type 2 diabetes, and Ma does not report what share had diabetes (Ma et al., 2017), so the figure transports to diagnosed diabetes only by extension. Bariatric surgery points the same way at a much larger dose: in the Swedish Obese Subjects study, surgery was associated with fewer deaths (adjusted HR 0.71). That study was not randomised — surgery patients were compared with matched controls on conventional treatment — and it did not report diabetes remission (Sjöström et al., 2007).
GLP-1 drugs cut heart attacks, deaths and kidney events in diagnosed diabetes
A meta-analysis of 11 cardiovascular-outcome trials of GLP-1 receptor agonists (GLP-1 drugs) pooled 67,769 people with type 2 diabetes against placebo, plus 17,604 without diabetes from SELECT. Median follow-up was 25.2 months (range 15.9-64.8), so these are two-year effects, not lifetime ones (Badve et al., 2025). In the diabetes trials, at high certainty:
- Major cardiovascular events (CV death, non-fatal heart attack, non-fatal stroke): HR 0.87 (0.81-0.93); one event prevented per 74 people treated for 25 months — roughly 14 fewer per 1,000 [INFERRED: 1,000/74].
- Death from any cause: HR 0.88 (0.83-0.93); one per 101 over 25 months — roughly 10 fewer per 1,000 [INFERRED: 1,000/101].
- Kidney composite (kidney failure, a sustained 50% fall in eGFR, or kidney death): HR 0.82 (0.73-0.93); one per 164 over 25 months — roughly 6 fewer per 1,000 [INFERRED: 1,000/164].
- Kidney failure itself: HR 0.84 (0.72-0.99), moderate certainty.
No single trial drove the MACE or kidney-composite result. The MACE effect was consistent across age, sex, BMI and kidney function; the kidney composite was split only by cardiovascular disease, kidney disease, agent, dosing frequency and follow-up, all as trial-level variables, and no subgroup analysis used individual patient data. The kidney benefit is modest in absolute terms because most trials enrolled people at low kidney risk (Badve et al., 2025); by baseline-risk arithmetic the absolute gain should be larger where kidney risk is high (Baseline Risk and the Relative-Absolute Split). A person’s own cardiovascular baseline in diabetes can be estimated with SCORE2-Diabetes SCORE2 Baseline Risk and the ESC Treatment Thresholds.
The cost side: serious adverse events did not differ, but more people stopped the drug because of side effects — 12.7% vs 9.2%, RR 1.51 (1.18-1.94), with very high heterogeneity between trials. Pancreatitis, thyroid and pancreatic cancer, severe hypoglycaemia and retinopathy were no more common than on placebo (Badve et al., 2025). In the STEP-1 extension, people who stopped semaglutide regained about two-thirds of the weight they had lost, and «Cardiometabolic improvements seen from week 0 to week 68 with semaglutide reverted towards baseline at week 120 for most variables» (Wilding et al., 2022). Whether the protection against heart attacks, strokes and deaths also fades after stopping has not been tested (Semaglutide for Cardiovascular Risk in Obesity); the wider side-effect ledger is on GLP-1 Non-Cardiometabolic Effects and Safety. Dual GLP-1/GIP agonists were not included (Badve et al., 2025), so these results do not speak to tirzepatide (GLP-1 Receptor Agonists and Cardiovascular and Kidney Outcomes).
The drug result does not show that weight loss protects the heart
These trials do not split the benefit by how much weight people lost; the meta-analysis mentions weight loss only as a possible distortion of creatinine-based eGFR [searched: weight loss / body weight / early / glucose-lowering across Badve chunk 01]. So the GLP-1 result is evidence for the drug, not for weight loss as the route. This guide’s wider reading treats the class benefit as largely independent of weight change; that is a synthesis, not something these trials measured GLP-1 Receptor Agonists and Cardiovascular and Kidney Outcomes.
Nor do the trials set the drug against lifestyle. SELECT enrolled non-diabetic adults with established heart disease; Look AHEAD enrolled people with diabetes on good background care. Opposite populations do not support a ranking of drug over lifestyle Does Weight Loss Reduce Cardiovascular Events.
What this changes, and what it does not
For an adult with diagnosed diabetes and high cardiovascular or kidney risk, a mature drug class now has high-certainty evidence on heart attacks, strokes and death, and moderate-certainty evidence on kidney failure. That shrinks how much those outcomes should rest on lifestyle weight loss Layer 1 - Ranking Interventions for a Stratum .
The shrink applies to those outcomes only. The drugs have not been shown to deliver drug-free remission, and they do not substitute for what weight loss does for fatty liver, physical function or coming off glucose-lowering medication Total Diet Replacement and Type 2 Diabetes Remission Fatty Liver MASLD and Weight Loss. Weight-reducing diets also lowered all-cause mortality, as above. Whether to take a drug, change lifestyle, or do both is the person’s choice, weighed against side effects, cost, and taking a medicine indefinitely. Starting, choosing or dosing a drug belongs with a clinician. For the drugs themselves, see GLP-1 Drugs; for weight and fat, Body Fat; for low-risk people, where most of these interventions show no hard-outcome benefit, see Cardiometabolic Interventions and Hard CV Outcomes in Low-Risk People.
That leaves several decisions the evidence cannot yet settle.
What the evidence cannot yet tell you
Each item below is a hole in the held evidence, not a finding of no effect: the answer is missing, not negative.
How long remission lasts, by which route, and for whom
- Nothing held follows DiRECT past two years. Follow-up was funded to three years, and participants consented to five (M. E. J. Lean et al., 2019). The longer signals are indirect: the ADDITION-Cambridge cohort described above, and one trial reporting at least partial remission in 5% at year six on a low-carbohydrate Mediterranean diet (Evert et al., 2019).
- Bariatric surgery is not appraised here. DiRECT’s authors argue that surgery works through the same weight and fat loss (M. E. Lean et al., 2018); no surgical remission trial is held.
- GLP-1 drugs and tirzepatide are untested here as a remission route, and no held trial pits them against a TDR or lifestyle programme. The held GLP-1 evidence covers heart, kidney and death outcomes, and excluded tirzepatide (GLP-1 Receptor Agonists and Cardiovascular and Kidney Outcomes).
- Transport beyond DiRECT’s population. Normal-weight, non-white, insulin-treated and longer-standing diabetes rest on mechanism or untested pooled signals (see Who else it may apply to above; Total Diet Replacement and Type 2 Diabetes Remission).
Prevention and the drug comparator
- How long the delay in diagnosis lasts. DPPOS’s 21-year follow-up is held for heart attacks and strokes, not for new diabetes (Lifestyle vs Metformin for Diabetes Prevention).
- A dose-response for aerobic activity. No held review maps minutes or steps of aerobic activity onto new diabetes; the held cohort data cover TV time and sitting instead (Sedentary Behaviour and Chronic Disease Risk).
- Metformin’s limitations beyond DPP, such as vitamin B12 depletion and long-term efficacy in established diabetes, are not held.
- Whether Look AHEAD participants who lost 10% or more had fewer cardiovascular events. That post-hoc analysis is held only second-hand (Does Weight Loss Reduce Cardiovascular Events).
Outcomes nobody has measured
- Whether remission prevents complications. DiRECT «was not designed or powered to evaluate effects on complications of diabetes»; its authors expect fewer microvascular complications if the lower HbA1c is maintained, which is a forecast, not a result (M. E. Lean et al., 2018) (Surrogate Outcomes).
- What life looks like after remission. Quality of life was measured to two years; function, relapse pattern and the shape of later decline are unmeasured.
Within those limits, the held evidence already supports a short list of actions.
What to do with this
- Prediabetes or high risk: start with an intensive programme aimed at losing about 7% of body weight plus regular activity. If the weight target is out of reach, keep the activity; in the trials, people who met the activity goal still developed diabetes less often (Lifestyle vs Metformin for Diabetes Prevention).
- Metformin is the fallback where a programme is unavailable or not enough. In the one trial that compared them, it paid off mostly in people with a higher BMI and higher fasting glucose.
- Prevention delays the diagnosis; it has not been shown to prevent heart attacks. Manage heart risk directly, through blood pressure and LDL.
- Recently diagnosed and overweight: a total diet replacement programme with a maintenance plan is the best-evidenced route to remission. Start early, and plan for regain: about half the people in the main trial needed a rescue phase to stay on track (Total Diet Replacement and Type 2 Diabetes Remission).
- Pick the diet you can keep. A low-carb route is reasonable if it is the one you will sustain. Protect fibre intake on it (Diets for Weight Management in Type 2 Diabetes).
- Exercise for fitness and as part of the programme. In diagnosed diabetes, interval training raised fitness more than steady cardio in short trials (11-16 weeks), with no firm glucose edge (HIIT vs Continuous Training for Type 2 Diabetes).
- Still open: whether remission lasts beyond two years, how surgery or GLP-1 drugs compare as remission routes, and whether remission prevents complications.
Changes to glucose-lowering or other medication belong with your clinician, especially before starting a formula or low-carb diet: keeping the same doses during rapid weight loss risks low blood sugar or dizziness on standing.
This guide grades the reasoning, not the outcome. Nothing here has been checked against what happened to the people who followed it.
Evidence box
Question ’For an adult at risk of, or living with early, type 2 diabetes, what does the evidence show about the modifiable lifestyle levers on incidence and on remission — direction, magnitude, durability, for whom — how do weight loss, diet composition, total diet replacement, and physical activity compare, and where does lifestyle sit against the standard first-line drug?‘ Evidence included 42 sources — 21 gold, 20 high, 1 moderate Overall certainty Medium (see Rating Certainty of Evidence) Source-selection note 1 source(s) below the gold/high evidence bar feed this page: Taylor (mechanism, moderate). Their roles and limitations are described in the text. Last updated 2026-10-06 · Independently reviewed: No · Full edit history