把账做平不等于账对了:残差是「调个参数」能救,还是模型少了一个维度

Written by

in

Language

Evidence level: C

残差能被某个参数吃掉,不代表这个参数是对的——它可能只是在替一个还不存在的维度打工。

正文

一、同一个「对上了」,有两条来路

一段脉搏波形拟合不好,剩下一截残差。摆在面前只有两条路:要么某个参数取值不对,调准就行;要么模型本身就少了一个维度,怎么调也调不到。前者是小修,后者是大修。

这两条路通向相反的结论。若只是参数没个体化,共振腔那条假说还站着;若缺的是结构性维度,那么在补上结构项之前,「几何占几成、弹性占几成」这类份额数字全部悬空。麻烦在于两条路表面一模一样,都是「调完曲线更贴了」。分开它们靠的是换一套判据。

二、可辨识性本来就分两层

心血管药理建模里有个干净案例:一个非线性血流动力学反馈模型总有参数估不准,结构可辨识性分析确认模型本身不可辨识,重参数化(削减参数表)后才变成全局可辨识。关键在于,重参数化前后的两个模型在输入-输出行为上完全无法区分(PMID 22343490)。这句话很重:只比对曲线,永远分不出「参数多的版本」和「参数少的版本」哪个是真的。

Profile likelihood 这条路线把两层拆得更干脆:结构不可辨识表现为功能相关的参数组合——换一组参数给出同样输出;实践不可辨识则来自数据量和质量不够(PMID 19505944)。前者加样本救不回来,后者能救。所以「残差压不下去」必须先分两层问,否则后面全是猜。

三、实践层有硬数字

闭环集总参数模型的系统性测试给了一组能直接用的参照:无噪声时所有参数都能标到万分之一以下误差;只要加上 5% 正态噪声,最敏感那五个参数误差被限在 10% 上下,最不敏感那四个用波形数据无法可靠估计,最不敏感那三个「特别难估」——作者的结论是优先去测它们,而不是去调它们(PMID 34758345)。同一份研究还发现,用压力波形做代价函数,比用每搏量、脉压这类临床指标准几个数量级。

这三个数字给了一条能搬的判据:残差压不到噪声底,先看这个参数在敏感性排序里排第几。末位那几个压不下去是正常的,属于「没测够」,不构成「模型错了」的证据。

四、结构性错误的签名是方向不对

非线性弹性管的实验给出了一个很干净的签名。同一根管、两种生理脉幅:线性方法在高阶散射明显增大,而且随脉幅升高,三点法在低阶给出物理上不成立的负衰减(PMID 10468357)。负的衰减系数没有任何物理解释——那是模型形式的错,不是参数值的错。

这就是两种错误的分水岭。参数错,残差有界、方向大体随机;结构错,残差会越界到物理不可能的象限。所以判据第一问不该是「残差多大」,而是「它有没有跑到不可能的象限里」。

五、反过来更危险:越能对上,越可能在跨栏补偿

谐波失真在 24 小时内随收缩压升高线性下降(PMID 35433650),说明幅值依赖是一个独立维度;模型里没有它,它就会以「某个参数随血压漂移」的形式冒出来。而交感能在独立于血流动力学的条件下调节大动脉硬度(PMID 33035441),等于多配了一个自由旋钮,专门用来与别的参数互相抵消。更底层:小动脉在肌源性与代谢调节受限时,局部阻力由血液表观粘度而非管径决定,表观粘度本身随灌注压与血流改变(PMID 26011432)——「阻力」在生理上本就不是常数。

所以「参数被调准了」必须附加条件:它在什么生理状态下准、能准多久。一个几何参数若随血压做小时级漂移,它不是被调准了,是在替缺失的幅值依赖维度打工。相关输入下的多解与参数补偿还会让份额估计进一步失真(PMID 41838759)。

六、判据三格

第一格,结构层先判:重参数化后输入-输出不可区分,或残差落进物理不可能象限(负衰减、非单调)→ 判结构性缺维度,属大修。份额数字一律作废,须先补压力依赖、频率依赖粘性这类结构项再谈分账。

第二格,实践层:单参数扫描能把该阶残差压到重复测量噪声底,且压完后参数值在隔日重复中稳定 → 判参数未个体化,属小修,假说保留。

第三格最易误判:残差压到噪声底,但参数值随瞬时血压或心率显著漂移 → 判吸收成功但归因错误。这一格必须单列,因为它表面「修好了」,实质是把结构缺失的方差记到了假参数名下。

七、测法

≥30 名受试者,每人隔日重复 3 次,估计每阶残差的噪声底与参数值的重测稳定性。对每一阶残差依次做三件事:单参数扫描(能否压到噪声底)、多参数联合反解(解是否沿脊线漂移、是否唯一)、残差与瞬时血压心率的逐阶相关(幅值依赖的直接读数)。敏感性排序先算出来,把末位参数预先归入「没测够」一类,避免污染「模型错了」的判断。

一句话收尾:残差是一张收据,它告诉你账没平,但不告诉你钱花在哪儿。要判小修还是大修,得看参数在什么条件下才准,而不是它准的时候有多准。

证据与局限

已验证:结构可辨识性与实践可辨识性必须在方法上分开处理,且重参数化前后的模型可以在输入-输出行为上完全无法区分(PMID 22343490、PMID 19505944);带 5% 噪声时,心血管集总参数模型中最不敏感的几个参数确实无法靠波形数据可靠估计(PMID 34758345);非线性弹性管在高脉幅下线性方法会给出物理不成立的负衰减(PMID 10468357);动脉硬度可被交感独立调节、表观粘度由灌注压与血流决定(PMID 33035441、PMID 26011432)。这些是「判别该怎么做」的方法学证据,强度较高。

部分验证:「残差压到噪声底但参数随血压漂移」这一格在脉象数据上尚未被直接观测到,目前由谐波失真随收缩压线性下降(PMID 35433650)间接推出,属合理推断而非实测。30 人 × 隔日 3 次的样本量与重复次数取自同类信度研究惯例,未针对脉象谐波做功效计算。

待验证:脉象谐波残差在各阶的可吸收性分布(多少阶能被单参数吸收、多少阶须多参数补偿);参数解脊线的实际形状与宽度;「末位参数」在脉象模型里的具体名单与排序;判据三格在真实数据上的落入比例。

相关阅读

证据等级: C
参考文献: PMID 10468357, PMID 19505944, PMID 22343490, PMID 26011432, PMID 33035441, PMID 34758345, PMID 35433650, PMID 41838759
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(HRV、皮温/皮电、呼吸率/呼吸模式、EMG、EEG、胃肠蠕动、气感报告、体质分型)全为可穿戴单点横断面观测量」——该主题(谐波残差的参数可吸收性与结构缺维度判别)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」;L1 对「残差 / 可辨识 / 反解 / 参数 / 血管 / 弹性 / 硬度 / 脉象 / 谐波 / 模型 / 血压 / 校准 / 补偿」十三个关键词全份文件零命中。


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。

Balancing the ledger does not mean it is correct: can residuals be fixed by tuning parameters, or does the model lack a dimension?

Evidence level: C

A residual being absorbed by a parameter does not mean that parameter is correct—it may simply be compensating for a dimension that does not yet exist in the model.

Main Text

I. Two Paths to the Same “Fit”

When a pulse waveform fits poorly, leaving behind a residual segment, only two paths lie before us: either a parameter value is incorrect and needs adjustment, or the model itself lacks a dimension, making it impossible to fit regardless of tuning. The former is a minor repair; the latter requires major revision.

These two paths lead to opposite conclusions. If the issue is merely lack of individualization of parameters, the resonance cavity hypothesis remains standing. However, if the missing element is a structural dimension, then ratios such as “what proportion is geometric and what proportion is elastic” remain entirely undefined until the structural term is added. The difficulty lies in the fact that both paths appear identical on the surface: “after adjustment, the curve fits better.” Distinguishing them requires a new set of criteria.

II. Identifiability Has Two Layers

In cardiovascular pharmacokinetic modeling, there is a clean case: a nonlinear hemodynamic feedback model always has parameters that cannot be estimated accurately. Structural identifiability analysis confirms the model itself is structurally unidentifiable; only after reparameterization (reducing the parameter table) does it become globally identifiable. Crucially, the two models before and after reparameterization are completely indistinguishable in their input-output behavior (PMID 22343490). This statement is significant: comparing curves alone can never distinguish between “the version with more parameters” and “the version with fewer parameters.”

The Profile Likelihood approach splits these two layers more cleanly: structural unidentifiability manifests as functionally related parameter combinations—swapping one set of parameters yields the same output; practical unidentifiability arises from insufficient data quantity and quality (PMID 19505944). The former cannot be saved by adding samples, while the latter can. Therefore, “residuals that cannot be suppressed” must first be divided into two layers of inquiry, otherwise everything that follows is pure speculation.

III. Practical Layer Has Hard Numbers

Systematic testing of closed-loop lumped-parameter models provides a set of directly usable reference values: without noise, all parameters can be estimated with errors below one ten-thousandth; once 5% Gaussian noise is added, the errors of the five most sensitive parameters are limited to around 10%, while the four least sensitive parameters cannot be reliably estimated using waveform data, and the three least sensitive ones are “particularly hard to estimate”—the authors conclude prioritize measuring them rather than tuning them (PMID 34758345). The same study also found that using pressure waveforms as the cost function is several orders of magnitude better than using clinical indices such as stroke volume or pulse pressure.

These three numbers provide a movable criterion: if residuals cannot be suppressed to the noise floor, first check where this parameter ranks in the sensitivity ordering. Those at the bottom being unsuppressible is normal, belonging to “not measured enough,” and does not constitute evidence that “the model is wrong.”

IV. The Signature of Structural Error Is Wrong Direction

Experiments on nonlinear elastic tubes provide a very clean signature. Using the same tube with two different physiological pulse amplitudes, linear methods show a significant increase in higher-order scattering, and as pulse amplitude rises, the three-point method yields physically untenable negative attenuation in lower orders (PMID 10468357). Negative attenuation coefficients have no physical explanation—that is an error in the model form, not an error in parameter values.

This marks the dividing line between two types of errors. Parameter errors result in bounded residuals with generally random direction; structural errors cause residuals to cross into physically impossible quadrants. Therefore, the first question of the criterion should not be “how large is the residual,” but rather “has it run into impossible quadrants.”

V. Conversely, It Is More Dangerous: The Better the Fit, the More Compensation Across Bars

Harmonic distortion linearly decreases over 24 hours as systolic pressure rises (PMID 35433650), indicating amplitude dependence is an independent dimension; without it in the model, it emerges as “some parameter drifting with blood pressure.” Meanwhile, sympathetic activity independently regulates aortic stiffness under conditions unrelated to hemodynamics (PMID 33035441), effectively adding another free knob specifically to cancel out other parameters. More fundamentally: when arteriolar myogenic and metabolic regulation is limited, local resistance is determined by blood apparent viscosity rather than vessel diameter, and apparent viscosity itself varies with perfusion pressure and blood flow (PMID 26011432)—”resistance” is not constant physiologically.

Therefore, “parameters being tuned correctly” must come with conditions: under what physiological state they are correct and for how long. If a geometric parameter drifts on an hourly scale with blood pressure, it is not being tuned correctly; it is working to compensate for the missing amplitude-dependence dimension. Multiple solutions under correlated inputs and parameter compensation further distort ratio estimates (PMID 41838759).

VI. Three-Grid Criteria

First grid, structural layer judgment first: after reparameterization, input-output behavior is indistinguishable, or residuals fall into physically impossible quadrants (negative attenuation, non-monotonicity) → judge structural dimension deficiency, requiring major overhaul. All ratio numbers are void; must first add structural terms such as pressure dependence, frequency-dependent viscosity before discussing allocation.

Second grid, practical layer: single-parameter scanning can suppress the residual of that order to the repeat-measurement noise floor, and after suppression, parameter values remain stable across overnight repeats → judge parameters not individualized, requiring minor repair, hypothesis retained.

Third grid is most prone to misjudgment: residuals suppressed to the noise floor, but parameter values drift significantly with instantaneous blood pressure or heart rate → judge successful absorption but wrong attribution. This grid must be listed separately because although it appears “fixed,” in reality it attributes the variance of missing structure to the name of a false parameter.

VII. Measurement Method

With ≥30 subjects, each undergoing repeat measurements three times overnight, estimate the noise floor and retest stability of parameter values for each order residual. For each order residual, sequentially perform three tasks: single-parameter scanning (can it be suppressed to the noise floor), multi-parameter joint inversion (does the solution drift along the ridge, is it unique), and correlation between residuals and instantaneous blood pressure/heart rate (direct reading of amplitude dependence). Sensitivity ranking is calculated first, with bottom-ranked parameters pre-classified as “not measured enough,” to avoid contaminating the judgment of “model is wrong.”

One sentence conclusion: residuals are receipts, telling you the accounts are not balanced, but not where the money was spent. To judge minor or major repair, one must look at under what conditions parameters are correct, rather than how correct they are when they happen to be right.

Evidence and Limitations

Verified: Structural identifiability and practical identifiability must be handled separately in methodology, and models before and after reparameterization can be completely indistinguishable in input-output behavior (PMID 22343490, PMID 19505944); with 5% noise, the least sensitive parameters in cardiovascular lumped-parameter models indeed cannot be reliably estimated using waveform data (PMID 34758345); nonlinear elastic tubes yield physically untenable negative attenuation under high pulse amplitudes using linear methods (PMID 10468357); aortic stiffness can be independently regulated by sympathetic activity, and apparent viscosity is determined by perfusion pressure and blood flow (PMID 33035441, PMID 26011432). These are methodological evidence for “how to judge what to do,” with high strength.

Partially Verified: The third grid—”residuals suppressed to noise floor but parameters drift with blood pressure”—has not been directly observed in pulse data; currently inferred indirectly from harmonic distortion linearly decreasing with systolic pressure (PMID 35433650), representing reasonable inference rather than direct measurement. The sample size of 30 subjects × 3 repeats overnight and the repetition count follow conventions from similar reliability studies, without power calculation specifically for pulse harmonic analysis.

To Be Verified: The absorbability distribution of pulse harmonic residuals across orders (how many orders can be absorbed by single parameters, how many require multi-parameter compensation); the actual shape and width of parameter solution ridges; the specific list and ranking of “bottom-ranked parameters” in pulse models; the proportion of cases falling into each of the three grids under real data.

Related Reading

Evidence Level: C
References: PMID 10468357, PMID 19505944, PMID 22343490, PMID 26011432, PMID 33035441, PMID 34758345, PMID 35433650, PMID 41838759
L1 Network Status: “Node tables are assigned across three domains: ‘Yijing Studies / TCM side’, ‘Physics side’, and ‘Physiological/Experience side’; the Physics side lists only ‘Phase / Rhythm — key periodic coupling elements’, while the Physiological side includes eight items (HRV, skin temperature/skin conductance, respiration rate/respiratory pattern, EMG, EEG, gastrointestinal motility, qi-sensation reports, constitution typing) as wearable single-point cross-sectional observables.” This topic (parameter absorbability of harmonic residuals and structural dimension deficiency judgment) has not yet entered the corresponding L1 network. The edge table stops at E-050 “Sixty Jiazi Na Yin | Correspondence | 60 Frequencies,” with the original text of the modern evidence column stating “no measurable source yet available.” The L1 network has zero hits for all thirteen keywords: residual, identifiability, inversion, parameter, vessel, elasticity, hardness, pulse, harmonic, model, blood pressure, calibration, compensation.


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 *