「哪些参数压不下去是正常的」这张名单,没人拉过:脉象模型参数敏感性排序的跨个体稳定性

Written by

in

Language

Evidence level: C

「末位参数压不下去是正常的」这条豁免逻辑,只有在一张统一、稳定的敏感性排序表上才成立——而这张表在脉象模型里从来没人拉过,更没人检查过它在不同人之间是不是同一张。

正文

一、一条好用的豁免逻辑,背后站着一张没验过的名单

上一篇文章讲过三格判据:残差压不到噪声底时,先看这个参数在敏感性排序里排第几——排在末位的压不下去是正常的,属于「没测够」,不构成「模型错了」的证据;只有结构层的缺失才需要判大修。

这条豁免逻辑在闭环集总参数模型上有硬数字撑腰:在 5% 噪声下,最不敏感的那四个参数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)。「末位 = 豁免」由此而来。

但这里藏着一个没有验证的前提:那张「哪些参数排在末位」的排序表,在脉象模型里从来没有被拉出来过。更关键的是,没有人检查过它在不同人之间是不是同一张。敏感性排序是个体化的——如果甲的末位参数在乙那里排到中游,那么「末位豁免」就是按人分配的,无法形成一条统一判据。

二、排序不是「模型的属性」,是「个体 × 模型」的属性

参数敏感性排序有一个被反复踩过的坑:它不是模型固有的性质,而是依赖你在参数空间的哪个点取值。系统模型的评价框架里有一个直接演示:同一个模型、同一个输出指标,局部敏感性分析(LSA)在参数空间的两个不同取值点上,给出的排序完全不同;换成全局 Sobol 分析,排序又与其中一点相近、与另一点相反(PMID 34921743)。

这件事对脉象模型是致命的:几何类、弹性类、边界类参数在人群里的变异系数差异很大(血管长度、壁厚、外周阻力都随年龄、体型、血压水平系统性漂移),那么每个人对应的「参数空间位置」就不同,敏感性排序在人群里很可能真的会换位。也就是说,即使甲身上「弹性类某参数排末位」,乙身上它完全可能排中游——「末位豁免」名单如果只有一张,对乙就是不成立的。

三、个体化排序还有一个更基本的坑:排序本身要能重测出来

就算每个人各自建表,还有第二层问题:排序能不能稳定地测出来?同一个人的敏感性排序,隔一天再测一次,是不是排得一样?

这层在波形测量的层面上有过正面的参照:用双阻抗心动图(d-ICG)测脉搏波速度,在社区样本里表现出跨数天的高重测信度与优秀的评估者间信度(PMID 33496965)。这说明「波形指标本身可以测稳」。但「指标测稳」不等于「敏感性排序测稳」——排序是把多个参数的估计结果拿来排队,每个参数的估计都带噪声,排名就会在噪声带里晃动。靠隔日重测估计排序的重测信度(ICC),正是把「名单稳不稳」从猜测变成数字的唯一办法。

四、判据分三档,先做仿真再上实测

把上面两层拆开,就得到一条可执行的裁决路径,分三档:

第一档:若跨个体 Kendall τ 中位数 <0.6(排序大面积换位),则「末位豁免」不能跨人套用,必须每人各自建表,#1179 第二格判据须改为「个体化排序 + 个体化豁免名单」;

第二档:若 τ 中位数落在 0.6–0.85,则只能按参数大类(几何类 / 弹性类 / 边界类)分组豁免,不能精确到单个参数——因为大类内部的换位仍然频繁,但大类之间相对稳定;

第三档:若 τ >0.85 且重测信度 ICC >0.75,则一张统一的末位名单可用,原判据可整体沿用。

两条轨道并行:仿真轨道用 §504 的 1-D 模型在 60 例合成个体(参数按实测分布抽样)上逐个算 Sobol 总效应指数,得到 60 张排序表,算 Kendall τ 的跨个体分布与每个参数的排名区间;实测轨道用 ≥30 人 × 隔日 3 次,估各自的敏感性排序,看排序的重测信度。反解「要把某参数稳健判入末位组,需要多少次重复测量」——这是把豁免逻辑从「拍脑袋」变成「有标定」的最后一步。

五、为什么这条边值得先于一切实验被拉出来

这套判据的意义不在方法本身,而在它决定前面所有结论的记账方式。上一篇文章的三格判据已经默认「末位豁免」存在;但如果排序随人换位,那么既往所有「单参数个体化成功」的案例里,都可能混着「恰好落进别人的豁免名单」的假成功。先验排序稳定性,等于先给整条共振腔假说的统计账本做一次审计——账没算清之前,任何「这个参数个体化成功了」的结论都该打个问号。

L1 网络状态:本条对应边在 L1 中无对应边,已登记 science-yixue-l1-gaps.md。母文 #1179(谐波残差的参数可吸收性 / 结构缺维度判别)已于 2026-10-22 登记为「测量学与外效度」簇第 17 条(原文摘录:「节点表三域只登记『观测什么』,既没有『这个量属于哪个数学空间』,也没有『这个量在给定观测下能否被唯一反解』,更没有『模型本身错了要怎么认出来』……补一条『丹田参数 ↔ 测量学与外效度维度』核心边仍可一次收敛全簇 17 条」);本条为同一簇深化第 18 条:即便补上「可辨识性层级」登记位,网络依然表达不了「敏感性排序本身是否跨个体稳定、末位豁免名单能否统一」——「末位 = 豁免」需要一个「排序稳定 → 名单统一 → 判据通用」的前提链,而 L1 没有任何位置承载「排序稳定性」这一被测量本身。

证据与局限

已验证: 集总参数模型在 5% 噪声下最不敏感参数无法可靠估计、末位豁免逻辑有量级参照(PMID 34758345);灵敏度排序依赖参数空间位置、LSA 与 GSA 排序可不同、QOI 选择改变排序(PMID 34921743);profile likelihood 可区分结构/实践不可辨识(PMID 19505944);结构不可辨识须先重参数化排除(PMID 22343490);波形指标(PWV)可做到跨天高重测信度(PMID 33496965)。

部分验证: 「脉象模型敏感性排序随个体换位」由几何/弹性/边界参数的人群变异系数差异与排序-空间位置依赖性间接推出(PMID 34921743 的排序换位演示),尚未在脉象模型上直接实测。

待验证: ① 60 例合成个体 Sobol 排序的跨个体 Kendall τ 分布——τ 中位数落哪档决定豁免名单能否统一;② ≥30 人隔日 3 次排序重测信度(ICC)——决定「名单稳不稳」是测出来的还是猜出来的;③ 「稳健判入末位组所需重复次数」的反解——把豁免从拍脑袋变成标定。

相关阅读

证据等级: C
参考文献: PMID 34758345, PMID 34921743, PMID 19505944, PMID 22343490, PMID 33496965


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。

Which Parameters Are Normally Uncompressible: Cross-Individual Stability of the Sensitivity Ranking of Pulse Model Parameters

Evidence level: C

The exemption logic that “failure to suppress the least sensitive parameter is normal” holds only if there exists a unified, stable sensitivity ranking table. Such a table has never been constructed for pulse models, nor has it ever been checked for consistency across different individuals.

Body

I. Behind the useful exemption logic stands an unverified list

The previous article discussed the three-tier criterion: when residuals cannot be suppressed to the noise floor, first check where this parameter ranks in the sensitivity ordering. If it ranks last, failure to suppress it is normal—it belongs to “insufficient measurement,” not evidence that “the model is wrong”; only structural-layer deficiencies require major revision.

This exemption logic has hard numbers backing it for closed-loop lumped-parameter models: under 5% noise, the four least sensitive parameters cannot be reliably estimated from waveform data, and the three least sensitive ones are “especially hard to estimate.” The authors concluded that it is better to measure them rather than tune them (PMID 34758345). “Last place = exemption” thus arises.

But here lies an unverified premise: the ranking table of “which parameters fall last” has never been pulled out for pulse models. More critically, no one has checked whether this table remains the same across different individuals. Sensitivity rankings are individualized—if Person A’s least sensitive parameter ranks mid-tier for Person B, then “last-place exemption” becomes person-dependent, preventing a unified criterion.

II. Sensitivity ranking is not “the model’s property,” but “individual × model” property

A repeatedly overlooked pitfall of parameter sensitivity ranking is that it is not an intrinsic property of the model, but depends on where in the parameter space you evaluate. The system model evaluation framework provides a direct demonstration: for the same model and same output metric, local sensitivity analysis (LSA) yields completely different rankings at two different points in parameter space; switching to global Sobol analysis, the ranking differs from one point and aligns with another (PMID 34921743).

This is fatal for pulse models: geometric, elastic, and boundary parameters exhibit large variation coefficients across populations (vascular length, wall thickness, peripheral resistance all systematically drift with age, body type, blood pressure levels). Consequently, each individual occupies a different “parameter space position,” and sensitivity rankings may genuinely swap across the population. That is, even if “an elastic parameter ranks last” for Person A, it could rank mid-tier for Person B—”last-place exemption” fails if the list is only one.

III. Individualized ranking faces an even more fundamental pitfall: the ranking itself must be retestable

Even if each person builds their own table, a second layer of problem arises: can the ranking be stably measured? If the same person’s sensitivity ranking is retested one day later, does it remain identical?

This aspect has positive reference in waveform measurement: dual-impedance photoplethography (d-ICG) measuring pulse wave velocity showed high cross-day retest reliability and excellent inter-rater reliability in community samples (PMID 33496965). This demonstrates “waveform metrics can be measured stably.” But “metric stability” does not equal “sensitivity ranking stability”—ranking involves ordering estimates of multiple parameters, each estimate carrying noise, so the rank order fluctuates within the noise band. Estimating retest reliability (ICC) of the ranking via cross-day retesting is the only way to transform “whether the list is stable” from guesswork into numbers.

IV. The criterion splits into three tiers: simulation first, then real measurement

Unpacking the above two layers yields an executable decision path in three tiers:

Tier 1: If cross-individual Kendall τ median <0.6 (widespread rank swapping), “last-place exemption” cannot be applied across people; each person must build their own table. Criterion #1179’s second tier must change to “individualized ranking + individualized exemption list”;

Tier 2: If τ median falls between 0.6–0.85, exemption can only be grouped by parameter class (geometric / elastic / boundary), not precisely to single parameters—because swaps within classes remain frequent, while classes themselves are relatively stable;

Tier 3: If τ >0.85 and retest reliability ICC >0.75, a unified last-place list applies, and the original criterion can be fully retained.

Two tracks run in parallel: the simulation track uses §504’s 1-D model on 60 synthetic individuals (parameters sampled from measured distributions) to compute Sobol total effect indices one by one, generating 60 ranking tables, then calculates the cross-individual distribution of Kendall τ and rank intervals for each parameter; the real measurement track uses ≥30 individuals × 3 retests on consecutive days to estimate each person’s sensitivity ranking, checking the retest reliability of the ranking. Inverting “how many repeated measurements are needed to robustly assign a parameter to the last-place group” transforms the exemption logic from “shot in the dark” into “calibrated.”

V. Why this edge must be pulled out before all experiments

The significance of this criterion lies not in the method itself, but in how it determines the accounting of all prior conclusions. The previous article’s three-tier criterion already assumed “last-place exemption” exists; but if rankings swap across individuals, then past cases of “single-parameter individualized success” may have mixed in false successes where the parameter merely fell into someone else’s exemption list. Prior ranking stability equals an audit of the statistical ledger of the entire resonance cavity hypothesis—before the books are balanced, any conclusion that “this parameter succeeded individually” must be questioned.

L1 network status: This edge has no corresponding edge in L1; it is registered in science-yixue-l1-gaps.md. The parent article #1179 (parameter absorbability of harmonic residuals / structural missing dimension discrimination) was registered on 2026-10-22 as the 17th item in the “Measurement and External Validity” cluster (original excerpt: “The node table’s three domains only record ‘what to observe,’ lacking ‘which mathematical space this quantity belongs to,’ ‘whether this quantity can be uniquely inverted given the observation,’ and ‘how to recognize when the model itself is wrong’… adding a core edge ‘Dantian parameter ↔ Measurement and External Validity dimension’ can once converge all 17 items in the cluster”); this item is the 18th deepening of the same cluster: even after adding “identifiability layer” registration, the network still cannot express “whether sensitivity ranking itself is stable across individuals, whether the last-place exemption list can be unified”—”last place = exemption” requires the prerequisite chain “ranking stability → unified list → universal criterion,” but L1 has no position to bear “ranking stability,” which is not itself measured.

Evidence and Limitations

Verified: Lumped-parameter models cannot reliably estimate the least sensitive parameters under 5% noise, validating the magnitude reference for last-place exemption logic (PMID 34758345); sensitivity ranking depends on parameter space location, LSA and GSA rankings can differ, and QOI selection alters ranking (PMID 34921743); profile likelihood distinguishes structural from practical non-identifiability (PMID 19505944); structural non-identifiability requires reparameterization first (PMID 22343490); waveform metrics (PWV) achieve high cross-day retest reliability (PMID 33496965).

Partially verified: “Pulse model sensitivity ranking swaps across individuals” is indirectly inferred from large variation coefficients among geometric/elastic/boundary parameters and the ranking-space position dependency (PMID 34921743’s ranking swap demonstration), but not directly tested on pulse models.

To be verified: ① cross-individual Kendall τ distribution of Sobol rankings for 60 synthetic individuals—where τ median falls determines whether exemption lists can be unified; ② retest reliability (ICC) of ≥30 individuals’ rankings measured on consecutive days—decides whether “list stability” is measured or guessed; ③ inversion of “number of repeats needed to robustly assign a parameter to the last-place group”—transforms exemption from guesswork into calibration.

Related Reading

Evidence level: C
References: PMID 34758345, PMID 34921743, PMID 19505944, PMID 22343490, PMID 33496965


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 *