账本对不平的那几笔钱,是谁花的:高谐波残差的归因顺序与阶次指纹

Written by

in

Language

Evidence level: B

方差分解告诉你”谁占几成”,但先查谁、后查谁,会改变谁占几成。残差归因真正难的不是算份额,是定顺序。

正文

上一篇把谐波预测误差拆成几何、弹性、边界三方份额,却留了个包袱:拆份额前得先决定”按什么顺序查账”。顺序不是技术细节——同一笔钱,先查甲的账甲就背,先查乙就乙背。这篇专门处理这被跳过的一步。

先说清为什么顺序会出事。方差分解的前提是输入彼此独立,但生理上这三项天然共变:年纪大的人血管又粗又硬,几何与弹性同涨;交感张力上去,粘性和出口阻力一起变。输入相关时,经典主效应分解会把”两个变量共同造成的那块方差”记不清归属,甚至让它凭空消失。近年的方法学综述点得很直白:绝大多数已发表敏感性分析仍是单因子轮换,默认模型线性可加,而这些假设往往站不住(PMID 26934843)。更关键的一条——传统方差基方法假设输入独立,这在校准过的模型上几乎从不成立;要公平分配就得用 Shapley 效应,而它又要求能对相关联合分布做条件采样(PMID 41838759)。技术含义是:“三方各占几成”在方法选错时是个假数字,假的方式取决于查账顺序。

第二件事,残差本身带指纹。参数误差随机散布,漏掉的结构项有方向性偏移,两者能分开看。非线性弹性管实验给了最干净的对照:同一根管、两种生理脉幅,线性方法在高阶散射明显增大,且随脉幅升高,三点法在低阶开始给出负的衰减估计(PMID 10468357)——负衰减物理上不成立,说明那部分”残差”根本不是血管的,是方法与非线性打架的产物。所以第一个筛子不是”谁占几成”,而是”这笔残差跨重复测量能不能复现”。测量链这条漏账线也不是虚的:用指尖光电容积波重建中心波形时,研究者得准备三套传递函数、按输入信号的谐波轮廓挑其中一套,才把形状误差压到 1.8%(PMID 39479582)——误差大小取决于这一拍的谐波谱长什么样,本身就是谐波依赖的系统偏移。不可复现的残差属于测量链,归因无意义。

第三件事,三条非几何候选的证据成色并不一样,混着用会互相顶账。

弹性这条最像”已结案”,其实没结。体外 37 条动脉全参数实测、零拟合下,纯弹性壁会在末梢产生欠阻尼高频振荡,加入粘弹性后第 5-20 谐波压力误差从 0.7% 降到 0.5%、流量侧 7.0% 降到 3.3%(PMID 21724188)。数字干净,但前提是”弹性已知”;一旦从实测变成反解,账就还回去了。更要紧的是弹性不是物性常数——人体实验里抑制一氧化氮合成,桡动脉壁相对粘性从 34.9% 掉到 28.9%;阻断钙激活钾通道则从 31.7% 升到 35.7%,两者叠加升到 41.5%(PMID 29158357)。同一根血管、同一个下午,只改内皮信号,粘性就动七八个百分点。

边界这条影响确实大,但大不等于可分离。换成多孔弹性管代表微循环后,作者明确说边界参数对压力与流量分布有”深刻影响”(PMID 30900050)。问题在于,”深刻影响”与”能唯一反解”是两件事,前者成立恰恰让后者更难。

交感这条最容易被漏账,因为它在模型里长得像边界参数。综述汇总微神经电图证据,结论是交感能在独立于当前血流动力学与血管运动张力的条件下调节大动脉硬度(PMID 33035441)。也就是说它既能改弹性也能改边界,是个跨两栏的项。跨栏项被硬塞进某一栏,另一栏就会显得解释力不足。

第四件事,也是最要命的一件:共线已在实测层面被看见了。谐波失真这个指标被系统检验过——24 小时里变异很大,随收缩压升高线性下降,同时与切向模量、脉搏波速度线性相关(PMID 35433650);传输线模型的敏感性分析则显示,增强指数、脉搏波速度、谐波失真三者主要对结构硬度敏感,而谐波失真与增强指数同时与几何参数强相关(PMID 40097550)。同一张谱,既随硬度走,也随几何走,还随当下血压走。任何”先固定几何、再扫弹性、再扫边界”的单向顺序,后扫者都会把前面所有的账吸进自己身上——这叫参数补偿,不叫归因。

所以方法序必须改。建议五步:第零步先做残差可归因性体检,看跨重复测量是否复现、是否与压力心率相关;第一步改用 Shapley 效应分配;第二步按压力与心率分层,把幅值依赖的非线性单拎出来(PMID 10468357 的负衰减即此层证据);第三步用可辨识性筛掉反解不出来的床;第四步报份额,并同时报”相关”与”独立”两种输入设定下的差值——差值本身就是共线强度。

判据提一个新的:“小修还是大修”的分界线不是份额大小,而是残差能不能被已有参数吸收。若某个非几何参数扫过一遍残差就掉到噪声底,共振腔假说只需小修——把该参数个体化即可;若单参数都吸收不了、必须几个参数同时调且互相补偿才勉强对上,说明模型形式错了,缺的是结构性维度(如压力依赖、频率依赖的粘性),这就不是调参能救的。

证据与局限

已验证: 输入相关时经典方差分解的独立性假设不成立,须改用 Shapley 效应并做条件采样(PMID 41838759);已发表敏感性分析多数仍用单因子轮换并默认线性可加(PMID 26934843);管壁粘弹性显著压低第 5-20 谐波误差(PMID 21724188);内皮 NO 与钙激活钾通道可急性改变人桡动脉壁相对粘性(PMID 29158357);交感可独立于血流动力学调节大动脉硬度(PMID 33035441);非线性弹性管中线性方法在高阶散射增大、随脉幅出现负衰减(PMID 10468357);谐波失真随收缩压线性下降且与硬度、几何同时相关(PMID 35433650、PMID 40097550);边界模型参数对压力流量分布影响深刻(PMID 30900050)。

部分验证: “残差可分噪声型与结构型”目前是从非线性管实验与敏感性方法学两头推出来的,尚未在”同个体 + 逐阶谐波输出”上直接做过判别检验;三条候选的相对份额(弹性/边界/交感)从未在相关输入设定下同时估计;”参数补偿导致后扫者背账”是机制推论,缺逐阶实证。

待验证: ①同批受试者上”独立输入 vs 相关输入”两种设定下三方份额的差值(共线强度的直接度量);②残差跨重复测量的可复现性分层(可复现/不可复现各占多少阶次);③残差与瞬时血压、心率的逐阶相关结构;④单参数扫描能否把残差压到噪声底的”可吸收性”测试——这是小修/大修判据的唯一直接检验。

相关阅读

证据等级: B
参考文献: PMID 10468357, PMID 21724188, PMID 26934843, PMID 29158357, PMID 30900050, PMID 33035441, PMID 35433650, PMID 40097550, PMID 41838759, PMID 39479582


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。

Who Spent the Discrepancies in the Ledger: Attribution Sequence and Order Fingerprints of High Harmonic Residuals

Evidence level: B

Variance decomposition tells you “what share each factor occupies,” but the order in which you audit determines that share. The true difficulty of residual attribution is not calculating shares but establishing the sequence.

Main Text

The previous article decomposed harmonic prediction errors into three shares: geometry, elasticity, and boundary. However, it left an unresolved issue: before assigning shares, one must decide “in what order to audit the accounts.” Order is not a technical detail—for the same debt, if you audit Account A first, A takes responsibility; if you audit B first, B takes responsibility. This article addresses this skipped step.

First, clarify why order causes errors. Variance decomposition assumes inputs are independent, but physiologically these three factors naturally covary: older individuals have thicker, stiffer vessels where geometry and elasticity rise together; increased sympathetic tension causes viscosity and outlet resistance to change simultaneously. When inputs are correlated, classical main-effect decomposition fails to correctly attribute the variance caused jointly by two variables, sometimes even making it vanish. Recent methodological reviews state plainly that most published sensitivity analyses still use single-factor rotation, assuming linearity and additivity—assumptions that often fail (PMID 26934843). A more critical point: traditional variance-based methods assume independent inputs, which almost never holds for calibrated models; fair attribution requires Shapley effects, which in turn require conditional sampling over correlated joint distributions (PMID 41838759). The technical implication is: “What share each party occupies” becomes a false number when the method is wrong, and the nature of the falsehood depends on the audit order.

Second, residuals themselves carry fingerprints. Parameter errors scatter randomly, while missed structural terms exhibit directional bias; these can be distinguished. Nonlinear elastic tube experiments provided the cleanest control: with the same tube and two physiological pulse amplitudes, linear methods showed significantly increased higher-order scattering, and as pulse amplitude rose, the three-point method began yielding negative attenuation estimates at lower orders (PMID 10468357)—negative attenuation is physically impossible, indicating that portion of the “residual” was not from the vessel but a product of conflict between method and nonlinearity. Thus, the first filter is not “what share,” but “can this residual be reproduced across repeated measurements.” The leak in the measurement chain is also not fictitious: when researchers reconstruct central waveforms using fingertip photoplethysmography, they must prepare three transfer functions and select one based on the harmonic profile of the input signal to reduce shape error to 1.8% (PMID 39479582)—error magnitude depends on the harmonic spectrum of this specific beat, constituting a harmonic-dependent systematic offset. Irreproducible residuals belong to the measurement chain; attribution is meaningless.

Third, the evidentiary quality of the three non-geometric candidates varies; mixing them causes mutual offsetting.

The elasticity line appears “closed,” but is not. Under zero-fitting conditions with full-parameter measurement of 37 in-vitro arteries, pure elastic walls produce under-damped high-frequency oscillations at the periphery; adding viscoelasticity reduced pressure error at harmonics 5–20 from 0.7% to 0.5% and flow-side error from 7.0% to 3.3% (PMID 21724188). The numbers are clean, but only under the premise that “elasticity is known.” Once moving from measurement to inversion, the ledger returns. More importantly, elasticity is not a constant property—in human experiments, inhibiting nitric oxide synthesis dropped relative wall viscosity from 34.9% to 28.9%; blocking calcium-activated potassium channels raised it from 31.7% to 35.7%, with both mechanisms combined raising it to 41.5% (PMID 29158357). In the same vessel, same afternoon, altering only endothelial signals shifts viscosity by nearly ten percentage points.

The boundary line has a significant impact, but significance does not imply separability. When replaced by porous elastic tubes representing microcirculation, the authors explicitly state that boundary parameters have a “profound influence” on pressure and flow distributions (PMID 30900050). The issue is that “profound influence” and “unique invertibility” are two different matters; the former holds precisely makes the latter harder.

The sympathetic line is most easily overlooked because it resembles boundary parameters within the model. A review summarizing micro-neurographic evidence concludes that sympathetic activity can regulate large artery stiffness independently of current hemodynamics and vascular motion tension (PMID 33035441). That is, it can alter both elasticity and boundary conditions, acting as an item spanning two columns. Forcing a cross-column item into a single column causes the other column to appear insufficiently explanatory.

Fourth, and most critical: collinearity has been observed at the empirical level. The harmonic distortion metric has undergone systematic testing—variation is high over 24 hours, decreases linearly with rising systolic pressure, and correlates linearly with tangential modulus and pulse wave velocity (PMID 35433650); sensitivity analysis of transmission line models shows that enhancement index, pulse wave velocity, and harmonic distortion are primarily sensitive to structural stiffness, while both harmonic distortion and enhancement index strongly correlate with geometric parameters (PMID 40097550). The same spectrum moves with stiffness, geometry, and current blood pressure simultaneously. Any unidirectional sequence such as “fix geometry first, then sweep elasticity, then sweep boundary” causes the latter to absorb all preceding accounts—this is parameter compensation, not attribution.

Therefore, the methodological sequence must change. We propose five steps: Step Zero performs a residual attribution viability check, verifying reproducibility across repeated measurements and correlation with pressure and heart rate; Step One adopts Shapley effect allocation; Step Two stratifies by pressure and heart rate to isolate amplitude-dependent nonlinearity (the negative attenuation in PMID 10468357 serves as evidence for this layer); Step Three uses identifiability screening to eliminate beds that cannot be inverted; Step Four reports shares while simultaneously reporting the difference under “correlated” and “independent” input settings—the difference itself quantifies collinearity strength.

We propose a new criterion: The dividing line between “minor repair” and “major overhaul” is not share size, but whether existing parameters can absorb the residual. If a non-geometric parameter reduces the residual to noise floor upon a single sweep, the resonance cavity hypothesis requires minor repair—individualizing that parameter suffices. If no single parameter absorbs it, and multiple parameters must be adjusted simultaneously with mutual compensation to barely match, the model form is incorrect; it lacks structural dimensions (such as pressure-dependent or frequency-dependent viscosity), which cannot be saved by parameter tuning alone.

Evidence and Limitations

Verified: Independence assumptions in classical variance decomposition fail under correlated inputs; Shapley effects with conditional sampling are required (PMID 41838759); most published sensitivity analyses still use single-factor rotation assuming linearity and additivity (PMID 26934843); wall viscoelasticity significantly suppresses errors at harmonics 5–20 (PMID 21724188); endothelial NO and calcium-activated potassium channels can acutely alter relative wall viscosity in human radial arteries (PMID 29158357); sympathetic activity can regulate large artery stiffness independently of hemodynamics (PMID 33035441); linear methods show increased higher-order scattering and negative attenuation with rising pulse amplitude in nonlinear elastic tubes (PMID 10468357); harmonic distortion decreases linearly with systolic pressure and correlates with both stiffness and geometry (PMID 35433650, PMID 40097550); boundary model parameters profoundly influence pressure and flow distributions (PMID 30900050).

Partially Verified: The distinction between “noise-type” and “structural-type” residuals is currently inferred from nonlinear tube experiments and sensitivity methodology, lacking direct discriminant testing in “same individual + stepwise harmonic output”; relative shares of the three candidates (elasticity/boundary/sympathetic) have never been estimated simultaneously under correlated input settings; “parameter compensation causing posterior absorption of accounts” is a mechanistic inference lacking stepwise empirical validation.

To Be Verified: ① Differences in three-party shares under “independent input” vs. “correlated input” settings within the same cohort (direct measure of collinearity strength); ② Stratification of residual reproducibility across repeated measurements (proportion of reproducible vs. irreproducible orders); ③ Stepwise correlation structures between residuals and instantaneous blood pressure and heart rate; ④ “Absorbability” testing via single-parameter scanning to reduce residuals to noise floor—the only direct test for the minor/major repair criterion.

Related Reading

Evidence Level: B
References: PMID 10468357, PMID 21724188, PMID 26934843, PMID 29158357, PMID 30900050, PMID 33035441, PMID 35433650, PMID 40097550, PMID 41838759, PMID 39479582


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 *