同一锅汤,换个下料顺序就换个结论:独立假设与相关假设下谐波方差份额的偏差

Written by

in

Language

Evidence level: C

“几何占四成、弹性占三成”——这句话漏了一个状语:在假定三者互不相关的前提下。换掉这个前提,四成可能变两成。

正文

一、一个数字少了状语,就等于没有数字

前几篇把脉象谐波的能量分配拆成”几何 / 弹性 / 边界条件”三份账,并且已经点出一个坑:分解顺序会改变各家的份额。但那一步只到方法学处方为止——改用 Shapley 效应、分层报告、附上差值——始终没有回答”差多少”。

而”差多少”恰恰是这个课题能不能继续的前提。因为方差基分解有一个几乎从不被写出来的前提:输入之间统计独立。前提成立,”占几成”是一个确定的数;前提不成立,同一个模型、同一批人、同一个采样路径,能算出好几个都”正确”的份额。此时任何单设定份额都不可比、不可复现,也不能拿去和别人的数字对照。

二、独立假设在建模圈里是常态,而常态不等于成立

这个假设的普遍程度有书目计量证据:过去十年里绝大多数已发表的敏感性分析仍是局部法或单因子轮换(OAT),并且建立在未经检验的线性与可加性假设上,全局方法只有少数人在用(PMID 26934843)。

更直接的一条来自 2026 年的方法学文章:传统方差基全局敏感性分析依赖输入独立的假设,而这个假设对经贝叶斯校准的模型几乎从不成立;近年的 Rosenblatt 变换与 Shapley 效应扩展理论上解决了这一限制,但实施时需要对相关联合分布做条件采样,而现有方案对输入依赖结构仍带限制性假设(PMID 41838759)。

换句话说:不是没人知道该怎么做,是做起来贵,所以默认不做。

三、脉象这条线上,”输入相关”不是可能性,是实测

如果三个输入真的互不相干,独立假设只是省事;但这条线上它们明显同向走动。

谐波失真(HD,基频以上能量与基频能量之比)在 24 小时连续遥测里随收缩压线性下降,同时与切向模量、脉搏波速度线性相关,且两个饮食组的斜率显著不同(PMID 35433650)——一个指标同时随血压、硬度、时间走。

传输线模型那一侧更直接:增强指数(AI)、脉搏波速度(PWV)、谐波失真三者都主要对结构刚度敏感,而 HD 与 AI 还与几何参数强相关(PMID 40097550)。三个”独立的解释变量”,被同一个观测量同时读取。

四、更麻烦的一层:弹性本身不是常数

份额分解默认每个参数是一个人身上的固定值。但血管壁的相对粘性在同一个下午就能被改:抑制一氧化氮合成后从 34.9±8.9% 降到 28.9±8.6%,阻断钙激活钾通道后从 31.7±6.6% 升到 35.7±8.0%,两者合用升幅更大(PMID 29158357)。

这不是”个体差异”,是急性调控。把一个受内皮调控的量当成常量塞进方差分解,等于把一部分本属”状态”的方差强行记到”参数”名下——份额自然被抬高。

五、这件事已经被正面做过一次,说明可行

生物力学领域已有现成路径:对典型的脉搏波传播模型做考虑输入相关性的全局敏感性分析,用代理模型把计算成本压到理论上的 1/27000,结论明确写着”输入相关性会显著影响敏感性分析结果,因此在模型开发阶段必须彻底调查输入相关性的影响”(PMID 38116742)。

技术门槛已经不高了。真正缺的是一句判据:差值到多大才算不能忽略。

六、把”差值”本身做成一把尺子

这里给一个可执行的判据,也是本文最想提出的一点:不要再问”几何占几成”,改问”在什么相关性结构下占几成“,并把两种设定之差当作”共线强度”的定量指标直接发表。

  • 差值 < 5 个百分点:共线弱,既往独立假设下的结论可保留,只需在方法里注明;
  • 差值 5–15 个百分点:必须双设定并列报告,任何单设定份额不得单独引用;
  • 差值 > 15 个百分点(尤其在 C5 以上高阶):既往基于独立假设的份额结论一律作废、须按相关输入重算。

七、怎么测

≥60 人的个体化一维血流模型,用同一条生理范围采样路径,跑两遍:① 各参数按边缘分布独立抽样;② 按实测的人群几何—弹性—边界相关性构造联合分布(或用后验相关)。两遍各算逐阶谐波能量的 Shapley 方差分解,分离主效应与交互项,逐阶比对份额差。

注意证据薄弱:以上全部是方法学推演加间接证据。L1 对应网络里没有这条边,一维模型份额能否外推到真实脉搏尚无直接验证,本结论只能当作”份额数字必须带假设标签”这一操作规范的论证,不能当作”几何占几成”的答案。

证据与局限

已验证:输入相关性会显著改变生物力学模型的敏感性分析结果,且代理模型可把相关输入的全局分析成本降到 1/27000(PMID 38116742);脉象侧三个常用指标对刚度、几何、血压存在重叠依赖(PMID 35433650、PMID 40097550);血管壁粘性受急性内皮调控而非常数(PMID 29158357)。

部分验证:独立输入假设在建模实践中普遍存在且常未被检验(PMID 26934843、PMID 41838759)——这是”问题存在”的证据,不是”偏差有多大”的证据;5 / 15 个百分点两条线目前无实证出处,是暂定阈值。

待验证:两设定份额差在真实脉象上究竟落在哪一档;逐阶差值是否随阶次单调放大;共线强度能否作为可复现的个体特征。

相关阅读

证据等级: C
参考文献: PMID 38116742, PMID 41838759, PMID 26934843, PMID 35433650, PMID 40097550, PMID 29158357
L1 网络状态:「节点表按『易学/中医侧』『物理侧』『生理/体验侧』三类域分派;物理侧仅列『相位 / 节律 —— 周期性耦合的关键』,生理侧八项全为可穿戴单点横断面观测」——该主题(谐波方差份额的输入相关性 / 共线强度)尚未进入 L1 对应网络,边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」,现代证据栏原文「尚无任何可测来源」。


免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

Same Soup, Different Conclusion: Bias in Harmonic Variance Share Under Independent versus Correlated Assumptions

Evidence level: C

“Geometry accounts for 40%, elasticity for 30%” — this statement omits a crucial adverbial clause: under the assumption that the three factors are mutually uncorrelated. Altering this premise could shift the 40% figure to 20%.

Main Text

I. A missing adverbial clause renders a number meaningless

Previous articles decomposed the energy allocation of pulse wave harmonics into three accounts: “geometry / elasticity / boundary conditions,” and highlighted a pitfall: the decomposition order alters each party’s share. That step stopped at the methodological prescription — using Shapley effects, layered reporting, and appending deltas — without ever answering “by how much”.

Yet “by how much” is precisely the prerequisite for continuing this inquiry. Variance-based decomposition rests on an almost never-stated premise: statistical independence of inputs. If this premise holds, “what share” is a determined number; if it fails, the same model, same cohort, and same sampling path can yield multiple “correct” shares. Under such conditions, any single-assigned share becomes incomparable, irreproducible, and unusable for cross-validation against others’ numbers.

II. The independent assumption is standard in modeling circles, but standard does not mean valid

The ubiquity of this assumption is evidenced by bibliometrics: over the past decade, most published sensitivity analyses remain local methods or single-factor rotation (OAT), and rely on untested linearity and additivity assumptions, while global methods are used by only a minority (PMID 26934843).

A more direct source comes from a 2026 methodological article: traditional variance-based global sensitivity analysis depends on the input independence assumption, which rarely holds for Bayesian-calibrated models; recent Rosenblatt transformations and Shapley effect extensions theoretically resolve this limitation, but implementation requires conditional sampling of correlated joint distributions, while existing schemes still impose restrictive assumptions on input dependency structures (PMID 41838759).

In other words: it is not that no one knows how to proceed, but it is expensive to do so, so it is defaulted out.

III. On the pulse wave line, “input correlation” is not a possibility but an empirical fact

If the three inputs were truly uncorrelated, the independent assumption would merely be a convenience; but along this line they clearly move in tandem.

Harmonic distortion (HD, the ratio of energy above fundamental to fundamental energy) linearly decreases over 24-hour continuous telemetry with systolic pressure, while simultaneously showing linear correlation with tangential modulus and pulse wave velocity (PWV), with significant differences in slopes between the two dietary groups (PMID 35433650) — one indicator thus varying with blood pressure, stiffness, and time.

The transmission line model side is even more direct: augmentation index (AI), PWV, and HD all primarily depend on structural stiffness, while HD and AI also strongly correlate with geometric parameters (PMID 40097550). Three “independent explanatory variables” are simultaneously read by the same observed quantity.

IV. A further complication: elasticity itself is not constant

Share decomposition defaults to each parameter being a fixed value for an individual. Yet the relative viscosity of the vessel wall can be altered within the same afternoon: inhibiting nitric oxide synthesis reduces it from 34.9±8.9% to 28.9±8.6%, blocking calcium-activated potassium channels raises it from 31.7±6.6% to 35.7±8.0%, and combining both yields an even larger increase (PMID 29158357).

This is not “inter-individual variation,” but acute regulation. Plugging a quantity governed by endothelial control into variance decomposition as a constant equals forcing part of the variance inherently belonging to “state” onto the account of “parameter” — naturally inflating the share.

V. This has been positively addressed once, proving feasibility

The biomechanics field already possesses a ready path: performing global sensitivity analysis considering input correlation on typical pulse wave propagation models, using surrogate models to reduce computational cost to theoretically 1/27000, with conclusions explicitly stating “input correlation significantly affects sensitivity analysis results, so input correlation effects must be thoroughly investigated during model development” (PMID 38116742).

The technical threshold is now low. What truly lacks is a criterion: to what extent must the difference be considered non-negligible.

VI. Turning “difference” itself into a ruler

Here we provide an actionable criterion, which is the most important point of this paper: stop asking “what share does geometry hold?” and instead ask “under what correlation structure does it hold,” and publish the difference between the two settings as a quantitative indicator of “collinearity strength.”

  • Difference < 5 percentage points: weak collinearity; conclusions under previous independent assumptions may be retained, provided the method notes this;
  • Difference 5–15 percentage points: both settings must be reported side-by-side; no single-set share may be cited alone;
  • Difference > 15 percentage points (especially at C5 and higher orders): all previous share conclusions based on independent assumptions are void and must be recalculated using correlated inputs.

VII. How to measure

≥60 individuals’ one-dimensional blood flow models, sampled along the same physiological range path, run twice: ① each parameter independently sampled from marginal distributions; ② joint distribution constructed from measured population geometry-elasticity-boundary correlations (or posterior correlation). For each run, compute Shapley variance decomposition of harmonic energies, separating main effects and interaction terms, then compare share differences order-by-order.

Note on weak evidence: All above are methodological deduction plus indirect evidence. L1 network lacks this edge; whether one-dimensional model shares can be extrapolated to real pulse waves remains unverified directly. This conclusion serves only as argument for the operational norm that “share numbers must carry assumption labels,” not as an answer to “what share does geometry hold.”

Evidence and Limitations

Verified: Input correlation significantly alters biomechanical model sensitivity analysis results, and surrogate models can reduce global analysis cost of correlated inputs to 1/27000 (PMID 38116742); three common pulse wave indicators exhibit overlapping dependence on stiffness, geometry, and blood pressure (PMID 35433650, PMID 40097550); vessel wall viscosity is acutely regulated by endothelium and not constant (PMID 29158357).

Partially verified: Independent input assumption is widespread in modeling practice and often untested (PMID 26934843, PMID 41838759) — this is evidence that “the problem exists,” not evidence of “how large the deviation is”; the 5/15 percentage point thresholds currently lack empirical basis and are provisional.

To be verified: Where exactly the difference between two settings falls on real pulse waves; whether incremental differences monotonically increase with order; whether collinearity strength can serve as a reproducible individual characteristic.

Related Reading

Evidence level: C
References: PMID 38116742, PMID 41838759, PMID 26934843, PMID 35433650, PMID 40097550, PMID 29158357
L1 network status: Nodes assigned across three domains — “Yijing Studies / TCM side,” “physics side,” and “physiological/experience side”; the physics side lists only “phase / rhythm — key periodic coupling,” while the physiological side includes eight items, all cross-sectional wearable observations. This topic (harmonic variance share input correlation / collinearity strength) has not yet entered the L1 network; the edge table stops at E-050 “Sixty Jiazi Na Yin | Correspondence | 60 frequencies,” with the modern evidence section originally stating “no measurable source yet.”


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *