Churuangsuk 2022 (Diabetologia) is the gold-tier umbrella review for the T2D-specific version of the “which diet?” question — 19 published meta-analyses of RCTs of weight-loss diets in people who already have type 2 diabetes, each pooled estimate AMSTAR-2-rated for MA quality and GRADE-rated for certainty, paired with a systematic review of remission across 7 diet types. It is the T2D-stratum counterpart to Named Diet Programs Compared (Ge, general population) and the map above the single-diet pages (Low-Carbohydrate vs Balanced-Carbohydrate Diets, Carbohydrate Restriction and Type 2 Diabetes Remission, Total Diet Replacement and Type 2 Diabetes Remission). (Churuangsuk et al., 2021)
The answer: no macronutrient profile wins — but the energy-delivery FORMAT appears to
The umbrella’s own conclusion separates two claims the “all diets are equal” slogan blurs together: meta-analyses «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)
So the finding is not flat interchangeability. It is:
- Macronutrient composition does not matter — LCD, high-protein, Mediterranean, high-MUFA, vegetarian, low-GI all sit at 0-2 kg from their controls, which the review calls «of little clinical significance» (Churuangsuk et al., 2021);
- The format that best controls energy intake appears to matter — VLED (−6.6 kg) and formula meal replacement (−2.4 kg) beat conventional self-administered low-energy diets, which the review reads as «generally providing less energy» rather than any compositional advantage.
Two cautions on the format-over-composition read. First, the mediation is the review’s interpretation, not a demonstrated pathway: its planned analysis «to differentiate effects of energy restriction and dietary regimen[] proved impossible from the published information» (Churuangsuk et al., 2021), so “wins by delivering less energy” is inference. The VLED contrast is near-definitional (it is compared against a higher-energy prescription), and the formula-meal-replacement −2.4 kg appears in a meta-analysis whose arms were isoenergetic (~1500 kcal both) — pointing to programme structure / adherence at equal prescribed energy, not simply less energy. Second, the format/composition split is certainty-driven, not magnitude-driven: formula meal replacement (−2.4 kg) separates from the composition diets not because 2.4 > 2.0 kg but because it rests on a high-quality MA at GRADE moderate, whereas the similar −1.6 to −2 kg for MUFA/vegetarian rest on critically-low MAs. Read this way it is the energy-deficit-plus- adherence thesis, the free-living T2D face of the isocaloric-feeding null on What Drives Fat Gain - Energy Balance vs the Carbohydrate-Insulin Model.
The certainty-graded weight-loss map
Every row is a pooled meta-analytic estimate vs the stated comparator, with AMSTAR-2 MA quality and GRADE certainty kept as separate axes. (Churuangsuk et al., 2021)
| Approach | Effect vs comparator | MA quality | GRADE certainty |
|---|---|---|---|
| VLED (400-500 kcal, 8-12 wk) vs low-energy diet | −6.6 kg (−9.5, −3.7) at 3 mo; −5.7 at 6 mo | high | low (2 RCTs, completer data) |
| Formula meal replacement (replace 1-3 meals) vs low-energy diet | −2.4 kg (−3.3, −1.4), 12-52 wk | high | moderate |
| LCD vs higher-carbohydrate | no difference (<1 to <2.5 kg) | high | low to high across 4 MAs (certainty of NO difference) |
| High-protein (>20%E) vs lower-protein | −1.2 kg (−2.2, −0.2) | critically low | very low |
| Mediterranean vs control | −0.3 to −1.8 kg (greater loss) | low / crit-low | very low / moderate |
| High-MUFA · vegetarian vs control | −1.6 to −2 kg | low / crit-low | (low) |
| Low-GI vs control | no difference | low / crit-low | — |
| Intermittent fasting (5:2 · TRE) | 5:2 no diff; TRE −1.4 kg | SR only, high-RoB | not gradable |
Two structural reads of this table.
- The LCD null is a positive no-meaningful-difference finding, not “insufficient evidence.” The comparison was run at scale (four high-quality MAs) and came back null; its certainty is not uniformly high: «GRADE assessment ranged from low to high certainty of evidence. Just one meta-analysis reported greater weight loss with LCD, by 3.5 kg, using complete case data for pooled results [41].» (Churuangsuk et al., 2021) — that fifth MA is also one of the review’s high-quality set (LCDs, n = 5), so “the high-quality MAs” are null on LCD except for this complete-case outlier. (corrected 2026-10-06: the single highest-certainty cell (GRADE high) -> GRADE low-to-high across the four MAs) This is the T2D-stratum confirmation of the pairwise Cochrane result on Low-Carbohydrate vs Balanced-Carbohydrate Diets.
- Outside the energy-controlling formats, the impressive point estimates rest on the worst MAs. 12 of 19 MAs were AMSTAR-2 «critically low»/«low»; high-protein, Mediterranean and vegetarian advantages rest on low / critically-low MAs, while the high-quality MAs found either no difference (LCD) or an advantage only for the formats that control energy (VLED, formula meal replacement). reading of the table above. (corrected 2026-10-06: the best-graded diets are the ones with null/negative findings -> dropped; VLED and meal replacement are high-quality-MA positives, Churuangsuk chunk 02)
HbA1c adds nothing separable: «HbA1c reduction broadly followed weight loss, and differences between diet types assessed over 3-12 months were small», and the data «do not permit an individual-level regression analysis to quantify weight loss-independent effects on HbA1c.» (Churuangsuk et al., 2021) So no diet is shown to buy a separable glycaemic edge — weight-loss-independent effects were small and could not be quantified. This is insufficient evidence for a composition-specific glycaemic effect, not a demonstrated null: the review could not separate the two, so read it as “not shown,” not “shown absent.”
The remission map — same structure, one clear winner on certainty
The remission SR (14 studies) is graded on Table 3. Here, unlike weight loss, one approach does dominate on certainty — but the winner is a format (total diet replacement), not a macronutrient class. (Churuangsuk et al., 2021)
| Approach | Remission at 1 yr (vs control) | Evidence | GRADE |
|---|---|---|---|
| Total diet replacement (formula induction) | 54% (range 46-61) vs 4-12% | 445 (2 low-RoB RCTs) | HIGH |
| Formula meal replacement | 11% vs 2% | 4503 (1 RCT, ancillary) | moderate |
| Mediterranean | 15% vs 4% LFD | 215 (1 RCT) | low |
| VLCKD / ketogenic | 20% vs 0 | 349 (1 non-RCT) | very low |
| Food-based VLED | 22% | 9 (1 uncontrolled) | very low |
The verdict the review draws: «The main contributor to HbA1c reduction and remission appears to be weight loss, irrespective of diet type.» (Churuangsuk et al., 2021) The single decisive GRADE-high cell is TDR — the intervention that delivers the largest sustained weight loss — which is why the map points to weight, not composition, as the causal variable. Held in full on Total Diet Replacement and Type 2 Diabetes Remission, where an independent population cohort corroborates the per-kilogram gradient.
The load-bearing gap: «No RCT has evaluated LCDs/ketogenic diets for type 2 diabetes remission.» (Churuangsuk et al., 2021) The much-marketed keto-for-remission figure (20%) is a single non-RCT at GRADE very low with serious risk of bias — so the popular low-carb-remission claim rests on the weakest cell in the table, sharpening the reading on Carbohydrate Restriction and Type 2 Diabetes Remission.
Who this changes the recommendation for (stratification)
- Route (a) — baseline / timing. Remission scales with sustained kilograms lost and with early action: «remission should be attempted as early as possible from diabetes diagnosis», the highest rates (up to 75%) being in newly-diagnosed / <2-year disease — a figure from non-randomised single-arm studies: «NRSs (non-RCT, single-arm intervention) reported remis- sion rates ranging from 3% to 75% by ITT, over various follow-up durations.» Longer duration collapses the achievable benefit — a duration-stratified prognosis, not a diet choice. (Churuangsuk et al., 2021)
- Route (a) prognostic, route (b) only a candidate — duration and ethnicity. «Participants with shorter type 2 diabetes duration, and Asian ethnicity, were more likely to achieve remission»; the review explored these as «hypothesised effect modi- fiers» via between-study heterogeneity, not a within-trial interaction test, and South Asians «may be more sensi- tive to weight loss». (Churuangsuk et al., 2021) A higher remission rate in a subgroup is prognostic (route a); positive evidence that the weight->remission slope differs (route b) is not held. (corrected 2026-10-06: route (b) effect modification -> route (a) prognosis with a route-(b) candidate)
- Route (c) — contraindications, diet-specific. Harms are not interchangeable even when efficacy is: ketogenic diets carry thiamine deficiency (heart failure, neurological) and ketoacidosis risk (amplified by SGLT2 inhibitors); replacing carbohydrate with red/processed meat raises LDL-cholesterol; high protein carries an observational kidney-disease association; rapid loss risks postural hypotension / hypoglycaemia if antihypertensive or hypoglycaemic drugs are continued. (Churuangsuk et al., 2021) So the choice among efficacy-equivalent diets is decided by the harm profile, cost, and adherence — not by weight-loss magnitude.
Decision relevance
- Choose the diet the person will sustain. Since composition does not demonstrably separate the diets on weight — Szczerba’s two standouts, meal replacement (high certainty) and plant-based (moderate, anthropometric), read as an energy format and an energy-mediated effect respectively (see The paired umbrella below;) — the operative levers are adherence, cost, harm profile and preference, with composition mattering only for the glycaemic/lipid surrogates Szczerba separates (e.g. low-carb HbA1c and triglycerides; high-protein, Mediterranean and low-GI on lipids) — «Practitioners can therefore be confident that a variety of diet types can all achieve the intended weight losses, and potentially remissions of type 2 diabetes, if their patients are able to adhere to the programme sufficiently.» (Churuangsuk et al., 2021) (corrected 2026-10-06: composition does not separate the diets on weight or HbA1c -> weight reconciled with Szczerba’s meal-replacement / plant-based findings; HbA1c caveated by Szczerba chunk 02)
- If maximal weight loss / remission is the goal, prefer a format that controls energy for the person: VLED or formula total-diet-replacement. The firmest support for this is the GRADE-high TDR remission cell, not the VLED weight-loss number (which is GRADE low, completer-only, and from COI-exposed authors) — so lead with the remission evidence and treat the VLED magnitude as low-certainty.
- The popular low-carb-for-diabetes claim is the weakest-supported one on both outcomes: LCD = higher- carb for weight (GRADE low to high across the high-quality MAs), and no RCT exists for its remission claim at all.
- The review declares the diet-comparison literature effectively saturated: «future trials of similar diet comparisons are unlikely to add useful information», redirecting the field to pragmatic / practice-based / n-of-1 evidence. (Churuangsuk et al., 2021)
The paired umbrella (Szczerba 2023) — surrogate refinement on shared evidence
Szczerba 2023 (BMJ Medicine) is the cardiometabolic-surrogate counterpart to Churuangsuk’s weight/remission map: 88 publications, 310 GRADE-rated MAs of RCTs (>=12 wk) in T2D. It refines the map to glycaemia, lipids, BP and one hard endpoint — it does not overturn it. (Szczerba et al., 2023)
Parameter table — the two umbrellas, matched (the shared-evidence + not-joined check).
| Parameter | Churuangsuk 2022 | Szczerba 2023 | Same quantity? |
|---|---|---|---|
| Meal replacement -> weight | −2.4 kg (−3.3, −1.4), GRADE moderate (Noronha 2019) | −2.37 kg (−3.30 to −1.44), n=9, GRADE high (Noronha 2019) | YES — same MA. Estimate identical; certainty grade DIFFERS. |
| Composition superiority | none, on weight / remission | none on weight / anthropometry; specific diets beat control on glycaemic / lipid markers | weight: YES (agree, null); glycaemia/lipids: different outcome -> refinement |
| Low-carb -> HbA1c | weight-loss-independent effect not separable (insufficient) | total effect −0.47% (−0.60 to −0.34), n=17, GRADE high | NO — total vs weight-independent; not opposed |
| Energy vs composition (weight) | format that controls energy appears to matter | isocaloric subgroups -> composition difference vanishes | YES, same direction — same literature/field, so F, not [E] |
Three findings fall out:
- The lone genuine divergence is a certainty clash on identical evidence. Both grade the same Noronha-2019 meal-replacement estimate, and reach moderate (Churuangsuk) vs high (Szczerba) — a worked case that GRADE certainty is rater/threshold-dependent even when the pooled number is the same -> filed on Rating Certainty of Evidence. Szczerba’s group carries no meal-replacement-industry COI (public funding), so its higher grade blunts the COI worry on the interpretation — though it shares the same underlying trials, so it is not independent confirmation of the effect. (Szczerba et al., 2023)
- On weight, Szczerba CORROBORATES energy format, not composition. Meal replacement is the sole GRADE-high weight lever; Mediterranean weight is null (−0.25, CI crosses 0); and where MAs matched calories, composition differences disappeared — «did not find differences when calories were restricted or matched with controls». Author’s own reading: «liquid meal replacements decrease energy intake and thus body weight.» (Szczerba et al., 2023) Same direction, shared evidence -> corroboration, not [E].
- On the GLYCAEMIC and LIPID surrogates, composition begins to separate (the refinement Churuangsuk’s weight-only map lacked) — but NOT on the anthropometric ones. The separation holds only for markers that are genuinely different constructs from weight: GRADE-high low-carb (<26%E) -> HbA1c −0.47% and triglycerides −0.30 mmol/L; moderate high-protein -> TC/LDL, Mediterranean -> TG, low-GI -> LDL. On the anthropometric family (weight, BMI, waist), composition does not separate — plant-based’s BMI −1.13 and waist −2.41 are weight-family measures, and plant-based (Szczerba’s ref 33) is itself one of the isocaloric subgroups where the composition difference vanished when calories were matched, so its anthropometric effect reads as energy-mediated, not composition-specific. So the honest split is: on weight/anthropometry composition is null (both umbrellas agree); only on glycaemia/lipids does the choice of diet move a different outcome — pick the diet by which of THOSE you want to move, once the energy/weight lever is set. (Szczerba et al., 2023) (Szczerba et al., 2023)
Two decision-relevant numbers beyond weight:
- Deprescribing. Low-carb (<26%E) «reduced the use of drug treatments by an additional 24 per 100 individuals» (risk difference 0.24, 0.12 to 0.35; moderate certainty). Within Szczerba’s own base its verdict is «For clinical outcomes, robust evidence exists only for a low carbohydrate diet and reduction in the use of drug treatments» — because Szczerba’s >=12-wk-MA-of-RCT filter places remission at low/very-low certainty. It does not outrank Churuangsuk’s GRADE-high TDR remission cell (a different filter admits the DiRECT RCTs); the two are complementary near-patient-important findings, drug-reduction from Szczerba and remission from Churuangsuk. (Szczerba et al., 2023)
- The one hard endpoint (the surrogate-to-outcome loop stays open). Szczerba’s own 310-MA base carries exactly one hard clinical outcome at high/moderate certainty — omega-3 -> major CV events, RR 0.94 (0.87 to 1.02), GRADE high — and its CI crosses 1: «the risk reduction was marginal». A separate, external MA (cited in Szczerba’s discussion for context, not part of its 310 base, and in a broader T1D+T2D population, >=3.9 yr) gives RR 0.93 (0.90 to 0.97), «marginal but precisely estimated». So the one in-base hard-outcome test is near-null, and the tighter figure is population-mixed and out-of-base — the surrogate-to-outcome loop is not closed, only touched at the edge. (Szczerba et al., 2023) (Szczerba et al., 2023)
Dose-signal on carbohydrate restriction (monotone, no knee shown). Deeper restriction buys more: a 10% carbohydrate decrease moves HbA1c only −0.11% (not clinically meaningful) while <26%E moves it −0.47%; Szczerba reports weight loss greater with low (<26%E) or very low (<15%E) than with moderate (<45%E) carbohydrate intake. Consistent with the domain default that every reduction pays until a knee is shown -> Low-Carbohydrate vs Balanced-Carbohydrate Diets. The low-carb effects may be under-estimated (poor adherence «especially for low carbohydrate and ketogenic diets», «potentially resulting in underestimation of their actual effect»). (Szczerba et al., 2023) (Szczerba et al., 2023)
Certainty base is thin on both. Szczerba: only «high for seven (2%), moderate for 40 (13%)» of 310 outcomes — 98% sit at moderate-or-below; AMSTAR-2 «low to very low in 77%». Corroborates Churuangsuk’s finding that the diet-comparison literature is large but mostly low-quality. (Szczerba et al., 2023)
The guidance benchmark on this question — and a cross-body emphasis type-D [2026-09-16]
This page’s two umbrellas now have a guidance-null test: the wiki holds both the US ((Evert et al., 2019)) and EU (EASD-DNSG 2023) diabetes-diet guidance. Both rest on the same evidence this page is built on — EASD’s diet-choice verdict cites the Churuangsuk umbrella directly — so agreement is shared-evidence corroboration (F), never independent [E].
Where both bodies agree with the umbrellas: no single macronutrient profile wins, and the energy-controlling format (total diet replacement) carries remission. EASD: a variety of weight-loss diets «can be used equally effectively… provided they can be followed and meet recommendations for protein, fat, micronutrient and fibre intake», and low-energy total diet replacement is «the most effective dietary approach for achieving type 2 diabetes remission» (Diabetes and Nutrition Study Group of the EASD, 2023). ADA: «there is not an ideal percentage of calories from carbohydrate, protein, and fat» (Evert et al., 2019).
The one genuine divergence — very-low-carb / ketogenic diets
The bodies split on VLC/keto. The parameter table shows the split is narrow, and located in one cell:
| Parameter | ADA 2019 (US) | EASD-DNSG 2023 (EU) | Same quantity? |
|---|---|---|---|
| Instrument | ungraded expert consensus | GRADE Moderate certainty (strength via wording) | different instrument |
| Population default on VLC/keto | «a viable approach» for «select adults… not meeting glycemic targets or where reducing antiglycemic medications is a priority» | «Very low carbohydrate intakes, such as with ketogenic diets, are not recommended» | NO — opposite population defaults |
| Individual who chooses VLC | permitted, with cautions (not pregnancy/CKD/disordered-eating; monitor meds) | permitted «with health professional support» | YES — both accommodate |
| Low-carb vs higher-carb, long term | A1C benefit «not at 12 and 24 months» | «no clinically significant long-term differences» | YES — agree |
| Observational U-shaped mortality + LDL-rise harm signal | «does not appear to increase overall cardiovascular risk, but long-term studies… are needed» | folds in «low (<40% total energy) and high (>70% total energy)… greater premature mortality» as a reason to recommend against | NO — same signal, opposite weighting |
(Evert et al., 2019) (Diabetes and Nutrition Study Group of the EASD, 2023)
Not-joined checks — this is a real but narrow type-D. The two bodies agree on every effect estimate (no long-term glycemic difference; attenuation by 12 months; both permit individual choice with support). They differ on exactly two joined things: the population default, and whether an observational U-shaped mortality signal + LDL rise warrants a recommendation-against. EASD weights that observational signal into a graded «not recommended»; ADA treats the CV-risk question as not-yet-established. The scope also differs (ADA carves out a select stratum; EASD issues a population rule), giving this a divergence-class-3 core (genuine disagreement on the same evidence) with a class-1 flavor (population-graded EU standpoint vs individualized-consensus US standpoint).
The hidden insight (the type-D payoff). A reader seeing ADA: viable / EASD: not recommended would
infer the evidence conflicts. It does not. The bodies hold the same estimates and the same harm
signal; they differ on the evidential standard for a population recommendation-against and on
standpoint. So the decision-relevant question is not who is right about VLC? but do you want a population
default that acts on an observational mortality signal (EASD) or one that waits for interventional CV
outcomes (ADA)? — and for the individual who will adhere to VLC with monitoring, both bodies land in the
same place. [E-independent] does NOT apply (shared trial base). -> Carbohydrate Restriction and Type 2 Diabetes Remission, The U-Shaped Association Artifact
Self-critique [run 2026-09-16, before commit]
- Counter-passage check RUN. Read ADA’s full carbohydrate section (Evert et al., 2019) and EASD’s full carbohydrate rationale (Diabetes and Nutrition Study Group of the EASD, 2023) end-to-end before filing the contrast; the individual-choice-accommodation cell is what keeps this from being over-stated as a flat contradiction.
- Parameter table built before the prose, with a same quantity? column; the divergence is confined to two cells (default + harm-weighting), not asserted across the board.
- Independence denied explicitly — shared trial base (DiRECT, Look AHEAD, same low-carb MAs; EASD cites Churuangsuk); agreement is F, the divergence is D. No [E].
- Over-read guard: the standpoint (class-1) flavor is named, so the type-D is not laundered into one body is wrong. Coherence, not validity — the loop stays open on whether either default improves outcomes.
Fibre — a diet-quality lever under whichever weight diet is chosen [2026-10-06]
The umbrellas above rank weight-loss diets; fibre is a separate, smaller axis that sits inside any of them.
A gold SR+MA in adults with prediabetes, type 1 or type 2 diabetes (42 trials, mostly 6-12 weeks) finds
higher fibre lowers HbA1c by -2.00 mmol/mol (-3.30 to -0.71; 33 trials), weight by -0.56 kg, and LDL by
-0.17 mmol/L, with all-cause mortality RR 0.55 (0.35-0.86) highest vs lowest in 2 European cohorts
(Reynolds et al., 2020). EASD’s
«Dietary fibre intake should be at least 35 g per day (4 g per 1000 kJ). Moderate»
(Diabetes and Nutrition Study Group of the EASD, 2023) cites that review as its reference
70, and the review was commissioned through the same EASD study group — so guideline and review are one
witness, not two (same commissioning group, and Reynolds is the guideline’s correspondence contact; no [E-independent]); EASD’s reference 8 is the same team’s 2019
review.
- Small next to the energy-delivery levers. About -0.18 percentage points HbA1c, and the effect was not detected in the only 5 trials that controlled weight (a small number, as the authors note) — which the authors note conflicts with earlier energy-balanced trials, so whether fibre’s glycaemic gain is independent of weight is unresolved. It does not compete with total diet replacement for remission.
- The 35 g is the top of the cohort data, not a located optimum; the review frames a 15 g/day increase, or reaching 35 g, as «a reasonable target».
- A low-carbohydrate pattern can cost fibre. The review: «for those who choose a reduced intake of total carbohydrates, the inclu- sion of fibre supplements may provide the means of ensuring recommended intake. However, further long-term adequately powered studies will be required to establish whether fibre sup- plements will confer longer term clinical benefit that is comparable with fibre that occurs natu- rally in foods.» (Reynolds et al., 2020) A substitution point for anyone taking the low-carb route above, not a reason against it.
Limits
- Surrogate/intermediate outcomes only. Weight, HbA1c and remission — no hard endpoint (mortality, MI, stroke, microvascular events); most evidence is <=1 year, so durability and complication effects are unaddressed -> Surrogate Outcomes.
- Coherence, not validity (R1). This grades and collates evidence; it does not close the loop on whether these weight/remission changes translate to fewer complications.
- COI. Several authors hold commercial ties (meal-replacement / formula-diet industry), directly relevant to the two approaches the review names as most effective — apply symmetric scrutiny to the favourable VLED/meal-replacement findings.
- Mostly European participants; transportability to other ethnic/deprived groups is caveated by the review itself (and cuts against the very South-Asian modifier it reports).
- AMSTAR-2 downgrades are partly instrument artefacts (no-protocol / no-publication-bias with <10 RCTs), so several «low» MAs may be methodologically sounder than the label -> Is This Actually a Systematic Review.
- Paired with a second gold umbrella (Szczerba 2023) — F-refinement on SHARED evidence, not two witnesses. The two umbrellas pool the same underlying MA for meal replacement (Noronha 2019), so their agreement is corroboration-with-caveat, not independent [E]. What Szczerba adds is in the section above.
- Szczerba is equally surrogate-bound, and does NO subgroup analysis (sex / race / diabetes duration), so it supplies no route-(b) effect-modification claim; its patient-relevant outcomes (remission, QoL, CVD incidence) all sit at low/very-low certainty. The stratification above still rests on Churuangsuk. (Szczerba et al., 2023)