Evidence level: C
残差压到噪声底 ≠ 模型修好了:如果参数值跟着瞬时血压一起漂,那不是修好了,是把结构缺失的方差记到了假参数名下。
正文
中医脉象的现代建模里,有一个非常诱人的里程碑:波形残差终于被压到了噪声底。残差小了,模型”拟合好了”,于是顺理成章地,从模型里反推出来的血管参数(硬度、阻力、几何)就被当作”这个人的真实状态”拿来个体化、拿来诊断、拿来随访。这套逻辑在集总参数模型的框架里尤其流行——把动脉树压缩成几个电阻-电容元件,波形一拟合,参数就出来了。
但 #1179 那一刀切在了最痛的地方:残差压平了,可能只是把账做平了,账本身却记错了。当模型少了一个真实的物理维度(比如血压依赖的硬度变化),拟合算法会把缺掉的那部分方差”吸收”进现有的参数里——参数值因此不再是它本来的生理量,而是”残差吸收器”:一个把结构缺失伪装成个体差异的假变量。
这一格目前完全靠间接推断。我们知道幅值依赖这个维度真实存在:血压波形的谐波失真(harmonic distortion, HD)随收缩压线性下降(PMID 35433650),交感神经可以独立于血流动力学调节大动脉硬度(PMID 33035441)——这两条证据都指向同一个事实:血管系统里有一个”随血压变化的旋钮”。但没有任何一条证据告诉我们:这个旋钮转动的时候,模型里那 8-10 个参数会跟着漂多少、漂几个、占多大比例。
而这件事的分量,恰恰藏在”占几成”这三个字里。如果第三格的发生率很高——比如超过三成参数呈现跨个体同号的血压斜率——那么既往所有”单参数个体化成功”的案例都要重判:它们可能集体落在第三格里,报出来的”个体化参数”其实是”当时那一下血压的快照”。
为什么参数会漂?三条生理路径,层层递进。第一条:硬度本身就是血压的函数。动脉壁的应力-应变关系是非线性的,血压一高,管壁被撑紧,硬度上升,模型里的弹性参数必须跟着变才能继续拟合波形——这在小鼠主动脉上被直接测过:谐波失真与收缩压的线性回归斜率在年龄和饮食组间显著不同(PMID 35433650)。第二条:阻力不是常数。小动脉里的血液是剪切变稀的非牛顿流体,灌注压下降时局部血管阻力反而上升——阻力这个”旋钮”本身就随灌注压变(PMID 26011432)。第三条:交感张力是独立的开关。肌肉交感神经活动(MSNA)可以独立于血压调节大动脉硬度,这意味着即使血压没变,参数也可能在漂(PMID 33035441)。
更麻烦的是第四条,方法学层面的:当参数之间存在相关结构时,”多解”会让漂移更难识别。贝叶斯校准的全局敏感性分析已经证明,在相关参数结构下,同一组波形可以对应多组同样”拟合良好”的参数组合——你看到的”这个参数变了”,可能只是算法在相关结构里换了一条等效路径(PMID 41838759)。三条生理路径负责”为什么漂”,这一条负责”为什么我们看不出它在漂”。
那么怎么测?设计其实不复杂,可以和 #1179 主实验共用受试者与波形:≥30 人,每人采集 ≥20 分钟连续波形并同步逐拍血压(指套或连续无创),按 30 秒窗滑动重估参数,得到每个参数的”时间序列”。然后做三件事:① 对每个参数做”参数值 vs 瞬时收缩压”的混合效应回归,斜率显著且跨个体同号者判为”系统性漂移”;② 只在残差已被压到噪声底的那些阶次上统计漂移发生率——这一条最重要,它保证我们测的不是”拟合不好所以参数乱跳”,而是”拟合好了参数还在跟着血压跑”;③ 用隔日重测的稳定性作对照,区分”真稳定的个体化值”与”看起来稳定其实在漂”的假稳定。
判据分三档。若在残差已达噪声底的阶次中,有超过 30% 的参数呈现跨个体同号的显著血压斜率——即漂移是系统性的而非噪声——那么”残差压到噪声底”就不足以支撑”小修成功”的结论,第二格判据必须附加”参数稳定性”条件,既往所有单参数个体化结论整体降级。若在 10%-30% 之间,则须在报告个体化参数时强制附带”该参数在本受试者上的血压斜率及其置信区间”。若低于 10% 且斜率符号随机,则第三格降为次要情形,小修判据维持原样。
这件事对易学-科学桥接的意义在于:脉象是”窗口”,但窗口的玻璃会随气压变形。当我们说”这个人的脉象参数反映了某条经络/某个脏腑的状态”时,如果参数本身是血压快照而非稳定个体特征,那么跨时辰、跨日、跨状态的比较就全都不成立——时辰对应关系(十二时辰与五脏)恰恰建立在”同一个人不同时间的脉象差异”之上。参数漂移率实测,就是给这张”时辰-脏腑对应网”做地基勘察:如果地基在随血压晃动,上面的对应关系就要重新标定。
L1 网络状态:CORRESPONDENCE_NETWORK 边表止于 E-050(纳音↔频率,open),无任何以”参数值随瞬时血压漂移 / 残差吸收 / 血压依赖硬度 / 假参数”为源或目标的边;EDGE_TABLE 概念计数中”谐波 147 篇/A 级 20″仅覆盖波形谐波失真与硬度(PMID 35433650 等),无”参数漂移发生率”概念条目。该主题尚未进入 L1 对应网络,已登记 gap(谐波残差簇第 19 条)。
证据与局限
已验证: 血压波形谐波失真随收缩压线性变化,幅值依赖维度真实存在(PMID 35433650);交感神经可独立于血流动力学调节大动脉硬度(PMID 33035441);微血管阻力随灌注压变化、非恒定(PMID 26011432)。
部分验证: 相关参数结构下多解使”参数漂移”更难识别(PMID 41838759,方法学仿真层面)。
待验证: ① ≥30 人连续波形+逐拍血压下,残差压到噪声底的阶次中参数随血压漂移的实际发生率(第三格占比);② 隔日重测能否区分”真稳定个体化值”与”假稳定”;③ 漂移率>30% 时,既往单参数个体化结论的实际降级比例。
相关阅读
证据等级: C(传统文化共识/初步探索——核心问题”漂移发生率”无直接实证,间接证据为动物实验与方法学仿真)
参考文献: PMID 35433650, PMID 33035441, PMID 26011432, PMID 41838759
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。
How much of \”Fixed but Recorded Wrong\” is accounted for: the incidence rate of parameter drift following residual absorption in instantaneous blood pressure
Evidence level: C
Residuals compressed to the noise floor ≠ model fixed: if parameter values drift with instantaneous blood pressure, it is not fixed; the missing structural variance has been recorded under false parameter names.
Body
In modern modeling of Traditional Chinese Medicine pulse waveforms, there is a particularly alluring milestone: waveform residuals have finally been compressed to the noise floor. Residuals are small, so the model “fits well”; consequently, vascular parameters (stiffness, resistance, geometry) inferred from the model are treated as “the individual’s true state” for personalization, diagnosis, and follow-up. This logic is especially prevalent within the lumped-parameter model framework—where the arterial tree is compressed into a few resistor-capacitor elements, and once the waveform is fitted, parameters emerge.
But #1179 cuts at the most painful point: residuals flattened may only balance the account, not record it correctly. When the model lacks a true physical dimension (e.g., stiffness variation dependent on blood pressure), the fitting algorithm absorbs the missing variance into existing parameters—parameter values are no longer their original physiological quantities but become “residual absorbers”: false variables that disguise structural deficits as individual differences.
This aspect currently relies entirely on indirect inference. We know amplitude dependence is a real dimension: harmonic distortion (HD) of the blood pressure waveform decreases linearly with systolic pressure (PMID 35433650), and the sympathetic nervous system can independently regulate large artery stiffness via hemodynamic regulation (PMID 33035441)—both pieces of evidence point to the same fact: there is a “blood-pressure-dependent knob” in the vascular system. Yet no evidence tells us how much, how many, or what proportion of the model’s 8–10 parameters drift when this knob turns.
The weight of this matter lies precisely in the phrase “what proportion.” If the incidence rate of the third scenario is high—for example, if more than 30% of parameters exhibit cross-individual consistent blood pressure slopes—then all previous cases of “successful single-parameter personalization” must be re-evaluated: they may collectively fall into the third scenario, where reported “personalized parameters” are actually “snapshots of that moment’s blood pressure.”
Why do parameters drift? Three physiological pathways, progressing stepwise. First: stiffness itself is a function of blood pressure. The stress-strain relationship of the arterial wall is nonlinear; as blood pressure rises, the vessel wall stretches tighter, increasing stiffness, so the elastic parameter in the model must adjust to continue fitting the waveform—this has been directly measured on mouse aortas: the linear regression slope of harmonic distortion versus systolic pressure differs significantly between age and diet groups (PMID 35433650). Second: resistance is not constant. Blood in arterioles behaves as a shear-thinning non-Newtonian fluid; when perfusion pressure drops, local vascular resistance actually increases—the “resistance knob” itself varies with perfusion pressure (PMID 26011432). Third: sympathetic tone is an independent switch. Muscle sympathetic nerve activity (MSNA) can regulate large artery stiffness independently of blood pressure, meaning even if blood pressure remains unchanged, parameters may still drift (PMID 33035441).
More troubling is the fourth aspect, methodological: when parameters exhibit correlated structure, “multiple solutions” make drift harder to identify. Bayesian calibration global sensitivity analysis has shown that under correlated parameter structures, the same waveform can correspond to multiple equally well-fitting parameter combinations—what you see as “this parameter changed” may simply be the algorithm switching to an equivalent path within the correlated structure (PMID 41838759). Three physiological pathways explain “why it drifts”; this one explains “why we cannot see it drifting.”
How to measure? The design is not complex and can share subjects and waveforms with the main #1179 experiment: ≥30 individuals, each collecting ≥20 minutes of continuous waveform with synchronized beat-by-beat blood pressure (finger cuff or continuous non-invasive), sliding re-estimation of parameters in 30-second windows to obtain “time series” for each parameter. Then three steps: ① mixed-effects regression of “parameter value vs instantaneous systolic pressure” for each parameter; significant slopes with cross-individual consistent signs indicate “systematic drift”; ② calculate drift incidence rate only for orders where residuals have been compressed to the noise floor—this is crucial, ensuring we measure “parameters still drifting despite good fit” rather than “poor fit causing erratic parameters”; ③ use cross-day retest stability as control to distinguish “truly stable personalized values” from “apparently stable but actually drifting” false stability.
Criteria are divided into three tiers. If, among orders where residuals have reached the noise floor, more than 30% of parameters show cross-individual consistent significant blood pressure slopes—i.e., drift is systematic rather than noise—then “residuals compressed to the noise floor” is insufficient to support the conclusion of “minor repair success”; the second-tier criterion must add a “parameter stability” condition, and all previous single-parameter personalization conclusions must be downgraded collectively. If between 10%–30%, then reporting personalized parameters must mandatorily include “the blood pressure slope and confidence interval for this parameter in this subject.” If below 10% with random slope signs, the third scenario becomes secondary, and the minor repair criterion remains unchanged.
The significance of this for bridging Yijing Studies and science lies in: pulse waveforms are “windows,” but the glass of the window deforms with air pressure. When we say “this person’s pulse parameter reflects the state of a meridian or organ,” if the parameter itself is a blood pressure snapshot rather than a stable individual characteristic, then comparisons across hours, days, and states all become invalid—the hour correspondence (Twelve Hours and Five Organs) is precisely built upon “pulse differences of the same person at different times.” Measuring the parameter drift rate is essentially surveying the foundation of this “Hour-Organ correspondence network”: if the foundation sways with blood pressure, the corresponding relationships above must be recalibrated.
L1 Network Status: CORRESPONDENCE_NETWORK edge table ends at E-050 (Na Yin ↔ Frequency, open); no edges originate from or target “parameter drift following instantaneous blood pressure / residual absorption / blood pressure-dependent stiffness / false parameters”; EDGE_TABLE concept count includes “harmonic 147 papers / Level A 20” covering only waveform harmonic distortion and stiffness (PMID 35433650, etc.), but lacks the “parameter drift incidence rate” concept entry. This topic has not yet entered the L1 correspondence network; gap registered (Harmonic Residual Cluster #19).
Evidence and Limitations
Verified: Blood pressure waveform harmonic distortion varies linearly with systolic pressure, confirming the amplitude-dependent dimension exists (PMID 35433650); sympathetic nervous system can independently regulate large artery stiffness via hemodynamic regulation (PMID 33035441); microvascular resistance varies with perfusion pressure and is not constant (PMID 26011432).
Partially Verified: Multiple solutions under correlated parameter structures make “parameter drift” harder to identify (PMID 41838759, methodological simulation level).
Pending Verification: ① actual incidence rate of parameter drift following blood pressure among orders where residuals are compressed to the noise floor under ≥30 individuals with continuous waveform and beat-by-beat blood pressure (proportion of the third scenario); ② whether cross-day retest can distinguish “truly stable personalized values” from “false stability”; ③ actual downgrade proportion of previous single-parameter personalization conclusions when drift rate exceeds 30%.
Further Reading
- Balancing the account does not mean recording it correctly: residuals—can adjusting a parameter save, or is the model missing a dimension?
- Is the “last-place list” really the same: cross-individual stability of sensitivity ranking
- Biological Frequency Code
Evidence Level: C (traditional consensus / preliminary exploration—core question “drift incidence rate” lacks direct empirical evidence; indirect evidence comes from animal experiments and methodological simulation)
References: PMID 35433650, PMID 33035441, PMID 26011432, PMID 41838759
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.
Leave a Reply