营养学的认知危机：从群体指南到个体失效

第 1 章 营养学的认知危机：从群体指南到个体失效

1.1 营养学的循证困境：当"最佳证据"遇见分子现实

If you have ever recommended a low-sodium diet to a hypertensive patient and watched their blood pressure barely budge, or prescribed a Mediterranean diet for cardiovascular risk reduction only to see LDL-cholesterol paradoxically rise, you have already encountered the central paradox of modern nutritional science: the recommendations we issue with the highest possible evidence grade (Grade A) are also, in clinical practice, the recommendations that fail most frequently at the bedside. This paradox is not a failure of diligence — it is a failure of inference. The evidence-based nutrition paradigm inherited from epidemiology has systematically mistaken population-level statistical regularities for individual-level causal mechanisms, producing what we might call *the fallacy of the average patient* (平均患者谬误).

To understand the depth of this crisis, we need to confront an uncomfortable fact: of the major chronic diseases that dominate mortality in the 21st century — type 2 diabetes, cardiovascular disease, certain cancers, and neurodegeneration — nutrition is implicated in roughly 40–60% of preventable cases according to the Global Burden of Disease 2019 estimates. Yet when a clinician or researcher pulls down the latest iteration of the Dietary Guidelines for Americans (DGA), the Chinese Dietary Guidelines (中国居民膳食指南, 2022), the WHO recommendations, or the EAT-Lancet planetary health diet, they are reading documents whose internal evidence base, upon methodological audit, often fails to meet the standards we would demand of a Phase III pharmaceutical trial. The American Heart Association's "Life's Essential 8" (2022) explicitly grades its own dietary recommendations as Class I/Level B-R in some domains and Level C-EO (expert opinion) in others — yet the document is delivered to clinicians and the public as if it were a unified, action-ready algorithm.

This chapter builds the methodological foundation for the entire book. We will perform a four-dimensional audit of the evidence architecture underpinning population-level dietary guidelines: (1) the evidence hierarchy itself and where nutrition science sits within it; (2) the magnitude of inter-individual variability in metabolic response, as quantified by the PREDICT-1 and ZOE studies; (3) the chronic misuse of surrogate endpoints as proxies for hard clinical outcomes; and (4) the measurement error structure of dietary exposure assessment. Each of these four domains reveals a structural deficiency that cannot be patched with larger sample sizes or longer follow-up. They require a categorical re-framing — a paradigm shift from reductionist, one-size-fits-all recommendations to mechanism-based, multi-omic, individualized prescriptions.

> 核心论断 (Central Thesis): 现行膳食指南的有效性受到四项结构性限制的复合压制——证据等级的结构性偏低、个体响应的巨大异质性、替代终点的因果外推不可靠、膳食暴露的测量误差巨大。这四项限制相互放大，共同构成"群体指南在分子层面失效"的根本原因，而精准营养的范式革命正是对这四项限制的同步回应。

The rest of this chapter is organized as a forensic examination of these four constraints, followed by a practical decision tree that integrates them. By the end, the reader should be able to defend, with methodological rigor, the position that *any* dietary recommendation issued at the population level must be considered a hypothesis to be tested in the individual — not a prescription to be followed.

---

1.2 膳食指南的证据等级与方法学审计

The evidence-based medicine (EBM) hierarchy, originally codified by the Canadian Task Force on the Periodic Health Examination (1979) and later refined through the GRADE working group (2004), ranks evidence from systematic reviews of randomized controlled trials (RCTs) at the top to expert opinion at the bottom. When this hierarchy is applied to nutrition, however, the result is, to put it gently, embarrassing. Let us audit the four major evidence categories that support the current generation of dietary guidelines.

1.2.1 观察性研究：混杂控制的先天不足

The bulk of nutritional epidemiology — and therefore the bulk of the recommendations in the DGA, the Chinese Dietary Guidelines, and the WCRF/AICR reports — rests on prospective cohort studies and case-control studies. These designs are vulnerable to three classes of confounders that are difficult or impossible to fully control:

1. Residual confounding by socioeconomic status (SES). Higher-quality diets correlate strongly with higher income, education, and access to healthcare. Adjustment for SES indicators reduces but does not eliminate confounding, and unmeasured dimensions of SES (e.g., neighborhood food environment, occupational physical activity) continue to bias effect estimates upward. The famous "French paradox" (high saturated fat, low CVD mortality) is a classic example of an effect estimate inflated by unmeasured lifestyle and cultural factors.

2. Reverse causation (反向因果). Individuals who are already ill — prediabetic, hypertensive, obese — often *change* their diet in response to early symptoms, producing the illusion that the "healthy" diet preceded disease when in fact the prodromal disease preceded the dietary change. This is particularly acute in studies of older cohorts, where subclinical disease may be driving both the dietary pattern and the outcome. The observational literature on multivitamin supplementation and cardiovascular outcomes is especially contaminated by this bias: supplement users are sicker at baseline.

3. Confounding by indication for plant-based diets. Vegetarian and vegan cohorts consistently show lower mortality, but they also show lower rates of smoking, lower alcohol consumption, higher physical activity, and lower BMI. Multivariable adjustment typically accounts for 30–50% of the apparent effect, leaving the residual susceptible to unmeasured confounding. Mendelian randomization studies (e.g., the use of *LCT* gene variants as instruments for dairy consumption) have generally produced much smaller effect estimates than observational studies, suggesting substantial confounding inflation.

The magnitude of these biases is not trivial. In a famous 2018 simulation study by Schoenfeld and Ioannidis, hypothetical "healthy" lifestyle factors were generated randomly and shown to predict cardiovascular events with hazard ratios of 0.60–0.80 in observational analyses — purely due to residual confounding. When the same factors were tested in RCTs, the protective effect typically shrank to 0.90–0.95. The lesson: when an observational nutrition study reports a hazard ratio of 0.70, a 30% risk reduction, the true causal effect may be closer to 5%. This is not a marginal correction; it is a fundamental re-calibration of how we should read the literature.

1.2.2 随机对照试验：外部效度的系统性塌缩

The RCT, gold standard of clinical research, has been applied to nutrition with great methodological care but limited ecological validity. The four largest dietary RCTs in history — the Women's Health Initiative Dietary Modification Trial (WHI-DM, n = 48,835), the PREDIMED trial (n = 7,447), the Lyon Diet Heart Study (n = 605), and the Dietary Approaches to Stop Hypertension (DASH) trials — each have taught us something important, but each has also revealed a structural limitation.

WHI-DM randomized postmenopausal women to a low-fat dietary pattern versus usual diet. The intent-to-treat analysis showed no significant reduction in breast cancer, colorectal cancer, or cardiovascular disease over 8.1 years of follow-up. The methodological response from the nutrition community was to point out that the intervention group failed to achieve a meaningful reduction in fat intake (only 8% absolute reduction at year 1, decaying to 4% by year 6), and that contamination of the control group was substantial. This is fair criticism — but it is also a confession: the RCT could not deliver the dietary contrast it was designed to test, and the *pragmatic* trial design that was used (free-living, self-administered dietary advice) did not produce adherence that would allow causal inf