把心率剥掉,浮沉弦滑还剩几分

Written by

in

Language

Evidence level: B

数脉和迟脉本来就是心率本身,它们分得开是废话;真正要问的是——浮沉弦滑这些”形状”里,有多少也只是心跳快慢投下来的影子。

正文

上一篇把六类脉象拆成五个要素、各自去找频谱上的轴,顺手留下一根刺:45 例高血压患者的自动评级里,浮脉的判定阈值落在了心率 ≥ 58.25 次/分(PMID 26988431)。一个讲”按得深不深”的脉象,判定条件竟然是心跳快慢。这不是中医的问题,是测量协议的问题,但它足以让整条分类管线的可信度打折。

这一篇专门处理这根刺。

先看一个不太舒服的事实

增强指数(AIx)是脉波分析里最主流的输出之一,长期被当作动脉僵硬度的指标。2000 年一项 22 名起搏器植入者的研究给出了它的真面目:起搏频率从 60 递増到 110 次/分,AIx 与心率呈显著负线性相关(r = −0.76,P < 0.001),心率每快 10 次/分,AIx 掉约 4 个百分点;射血时长与心率同样负相关(r = −0.51)(PMID 10811742)。机制说得很清楚:心率快了,收缩期绝对时长变短,反射波返回时落在收缩期的位置就变了,波峰形状跟着变。不是血管变软了,是节拍变了。

2009 年那篇更狠。50 例患者做多巴酚丁胺负荷,同时测桡动脉波形与组织多普勒:AIx 与射血时长 r = 0.88、与心率 r = −0.81、与心肌收缩应变率 r = 0.72(全 P < 0.001)。但把收缩应变率做心率校正后,它与 AIx 的关系消失了(r = 0.18,P = 0.11)。射血时长单独解释了 AIx 78% 的方差(β = 0.88,模型 R² = 0.77)。作者的结论一句话:AIx 由变时性(chronotropic)而非变力性(inotropic)决定(PMID 19720955)。

一个被写进无数论文当”血管结构指标”的量,近八成变化来自射血时长——而射血时长基本就是心率的函数。这不是孤例:40 名健康青年的压力脉波研究显示,时间类参数的主控因素是心率,振幅类的主控因素是收缩压(PMID 26406041)。二阶导数那套指标(b/a、c/a、d/a、老化指数 AGI,正是”弦””滑”最常借用的形状量)也一样:848 人样本里,D:A 比与 AGI 独立地负相关于心率,B:A 比正相关,校正年龄与血压后仍成立(PMID 15831356);294 例高血压患者重复了同一方向(PMID 12473866)。

所以”控制心率”不是一句客套话。形状类脉象的判别力里,很可能有一块是纯心率租来的。

三步分离,加一个必须崩的对照

第一步:心率匹配配对。 按 1:1 最近邻匹配,卡钳 ±3 次/分,配对年龄、性别、平均压。匹配后仍保留的判别力才可能是形状信号。

第二步:协变量调整。 心率作为连续协变量进模型,同时放年龄、平均压、接触压力档位。这里有个关键的方法学分歧:不要用”校正到 75 次/分”这种一步到位的做法。 2014 年一篇专文指出,把 AIx 统计校正到 75 bpm 在生理和统计上都站不住:一是心率与动脉僵硬度存在慢性交互,二是这种校正隐含假设”心率-AIx 关系在全人群一致”,而这个斜率本身随年龄、族群、病理状态变化。更稳的做法是把心率作为独立预测因子或协变量放进模型,而不是先折算再忽略(PMID 24257465)。

第三步:窄心率带内判别。 在 60-70、70-80、80-90 三个窄带内分别做判别,看 AUC 是否还在。这一层最能说明问题——如果判别力只在全样本里存在、一进窄带就蒸发,那它一直是靠心率撑着的。

第四步:阳性对照。 这是本设计最要紧的一步:数脉/迟脉的判别 AUC,在控制心率之后必须崩到 0.5。 因为它们本来就是按心率定义的,控制心率等于把定义变量拿走。如果它们没崩,就说明管线存在信息泄漏——那么浮沉弦滑的所有结果一律作废。 一个必然归零的对照,用途是给整条管线做体检。这比任何 AUC 都值钱。

两个容易翻车的地方

一是碰撞偏倚。 如果某种病理状态同时升高心率和改变波形,那么”匹配心率”等于在这条通路上做了条件化,反而会造出假关联。补救办法是报告匹配前后的疾病构成,并在有疾病标签的子集里做敏感性分析。

二是匹配了数字,没匹配生理。 48 例起搏器植入者研究显示,心房、房室、心室三种起搏方式对”心率-射血时长”关系有显著影响(P = 0.01);即便固定心率,不同起搏方式的 AIx 与脉压放大率仍有显著差异(P < 0.0001)(PMID 27471239)。所以心率匹配必须同时锁定搏动来源。

还有一个反向陷阱

归一化能剥掉心率,但会连真信息一起扔。2023 年一项方法学研究比对波形时发现:不归一化的情况下,两个波形尾端实际形态差异很小,但因心率不同,尾部反而成了累积距离的主要来源,形态差异被心率彻底淹没;双重归一化后才真正聚焦形态——代价是心率与脉压被移出主特征,只能降级另算(PMID 36767962)。“控制心率”和”丢掉心率信息”是两件事,后者会顺手毁掉数脉迟脉的判别基础。

心率本身也不干净:52 例起搏器研究剥离血压效应后,颈股脉波速度每 10 次/分仍升 0.16-0.20 m/s(PMID 27245180)。

怎么读结果

判据提前写死,避免事后解释:

  • 匹配+调整后保留 ≥ 80% 判别力 → 该类主要是真形状信号,心率不是主因;
  • 保留 50-80% → 混合型,必须在报告里标明”含心率成分”,并给出心率贡献的比例;
  • 保留 < 50% → 主要是心率伪装,该类的现有文献判别力需要整体下调;
  • 阳性对照(数/迟)未崩 → 管线有泄漏,全部结论作废重跑。

这不是又一篇”证明脉诊有用”的文章。它做的是审计:在把脉象交给任何分类器之前,先确认分类器学的是脉,不是心跳的快慢。

证据与局限

已验证: 起搏器递增起搏下 AIx 与心率显著负线性相关、心率每 10 次/分 AIx 降约 4%、射血时长与心率 r = −0.51(PMID 10811742);AIx 与射血时长 r = 0.88、射血时长解释 78% 方差、收缩应变率经心率校正后与 AIx 无关(PMID 19720955);AIx 校正到 75 bpm 在生理与统计上均不合适(PMID 24257465);压力脉波时间类参数主因为心率、振幅类主因为收缩压(PMID 26406041);SDPTG 的 D:A、AGI 负相关于心率而 B:A 正相关,校年龄血压后仍成立(PMID 15831356、PMID 12473866);DTW 不归一化时心率主导波形距离、归一化会移除心率与脉压信息(PMID 36767962);不同起搏方式在固定心率下 AIx 仍有显著差异(PMID 27471239);剥离血压效应后脉波速度仍随心率上升 0.16-0.20 m/s/10bpm(PMID 27245180);45 例高血压中浮脉自动阈值落在心率 ≥ 58.25(PMID 26988431)。

部分验证: “形状类脉象判别力中有相当比例来自心率”目前是从上述推广到脉象分类的合理推断,尚无研究在同一批受试者身上直接做过心率匹配的脉象判别对照;”保留 ≥ 80% 即真形状信号”这一阈值是本文设定的判读口径,非既有共识。

待验证: ①心率匹配前后浮沉弦滑四类判别 AUC 的实际保留比例;②数/迟作为阳性对照是否真能崩到 0.5(即管线是否存在泄漏);③心率匹配引入的碰撞偏倚在本议题下是否会制造假关联。

相关阅读

证据等级: B(心率对波形时相与形状的主导作用、AIx 的变时性依赖、SDPTG 的心率依赖、归一化的双刃性均有直接文献;”心率匹配分离检验”用于脉象分类本身无先例,属研究提案)
参考文献: PMID 10811742, PMID 19720955, PMID 24257465, PMID 26406041, PMID 15831356, PMID 12473866, PMID 36767962, PMID 27471239, PMID 27245180, PMID 26988431


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。

Stripping away the heart rate, how much remains of the floating and sliding strings

Evidence level: B

Pulse counting and late pulse are inherently heart rate itself; separating them is redundant. The real question is—within these “shapes,” how much is merely a shadow cast by heart rate variability.

Main Text

The previous article decomposed six pulse categories into five elements and identified their spectral axes, inadvertently leaving a thorn: in an automated rating of 45 hypertensive patients, the threshold for identifying the Floating pulse fell at a heart rate ≥ 58.25 bpm (PMID 26988431). A pulse quality defined by “how deep it feels” relies on a condition determined by heart rate. This is not a flaw in Traditional Chinese Medicine (TCM); it is a flaw in the measurement protocol, yet it is sufficient to undermine the credibility of the entire classification pipeline.

This article addresses that thorn directly.

Start with an Uncomfortable Fact

The Augmentation Index (AIx) is one of the most mainstream outputs in pulse wave analysis, long regarded as an indicator of arterial stiffness. A 2000 study of 22 pacemaker recipients revealed its true nature: as pacing frequency increased from 60 to 110 bpm, AIx showed a significant negative linear correlation with heart rate ($r = -0.76$, $P < 0.001$); AIx dropped by approximately 4 percentage points for every 10 bpm increase in heart rate; ejection duration also showed a negative correlation with heart rate ($r = -0.51$) (PMID 10811742). The mechanism is clear: faster heart rates shorten absolute systolic duration, shifting the position where the reflected wave returns during systole, thereby altering the waveform peak shape. It is not that the vessels have softened; it is that the rhythm has changed.

The 2009 study was even more definitive. In 50 patients undergoing dobutamine stress testing, simultaneous measurements of radial artery waveforms and tissue Doppler showed AIx correlated strongly with ejection duration ($r = 0.88$), heart rate ($r = -0.81$), and myocardial contraction strain rate ($r = 0.72$; all $P < 0.001$). However, after adjusting for heart rate, the relationship between strain rate and AIx disappeared ($r = 0.18$, $P = 0.11$). Ejection duration alone explained 78% of the variance in AIx ($\beta = 0.88$, model $R^2 = 0.77$). The authors concluded in one sentence: AIx is determined by chronotropy (heart rate variability) rather than inotropy (contractility) (PMID 19720955).

A metric written into countless papers as a “vascular structural indicator” changes nearly 80% due to ejection duration—which is essentially a function of heart rate. This is not an isolated case: a study of 40 healthy young adults under stress showed that time-based parameters were primarily driven by heart rate, while amplitude-based parameters were driven by systolic blood pressure (PMID 26406041). The same holds for second-derivative metrics (b/a, c/a, d/a, and the Aging Index AGI, which are commonly used to quantify “Wiry” and “Slippery” shapes): in a sample of 848 individuals, D:A ratio and AGI independently correlated negatively with heart rate, while B:A ratio correlated positively; these associations held even after adjusting for age and blood pressure (PMID 15831356); the same direction was replicated in 294 hypertensive patients (PMID 12473866).

Therefore, “controlling heart rate” is not just polite rhetoric. A significant portion of the discriminative power in shape-based pulse qualities may be rented from heart rate.

Three Steps to Separate, Plus One Critical Control

Step 1: Heart Rate Matching. Match 1:1 using nearest neighbors within ±3 bpm, pairing for age, sex, and mean blood pressure. Only discriminative power remaining after matching can be considered a true shape signal.

Step 2: Covariate Adjustment. Enter heart rate as a continuous covariate in the model, alongside age, mean blood pressure, and contact pressure levels. There is a critical methodological divergence here: do not use a “one-step correction to 75 bpm”. A 2014 article pointed out that statistically adjusting AIx to 75 bpm is physiologically and statistically unsound: first, heart rate interacts chronically with arterial stiffness; second, such correction implicitly assumes the “heart rate-AIx relationship is consistent across all populations,” a slope that itself varies by age, ethnicity, and pathological state. A more robust approach is to include heart rate as an independent predictor or covariate in the model rather than pre-calculating and ignoring it (PMID 24257465).

Step 3: Discrimination within Narrow Heart Rate Bands. Perform discrimination separately within three narrow bands (60-70, 70-80, 80-90 bpm) to see if AUC remains. This layer reveals the truth—if discriminative power exists only in the full sample but evaporates within narrow bands, it has always been propped up by heart rate.

Step 4: Positive Control. This is the most critical step in this design: the AUC for Pulse Counting/Late Pulse discrimination must collapse to 0.5 after controlling for heart rate. Since they are defined by heart rate, controlling for heart rate removes the defining variable. If they do not collapse to 0.5, it indicates information leakage in the pipeline—rendering all results for Floating/Sunken/Wiry/Slippery invalid. A control that must theoretically reach zero serves as a health check for the entire pipeline. It is more valuable than any AUC metric.

Two Places Prone to Failure

First: Collision Bias. If a pathological state simultaneously elevates heart rate and alters waveform morphology, “matching heart rate” effectively conditions on this pathway, potentially creating spurious associations. The remedy is to report disease composition before and after matching and conduct sensitivity analyses within subsets labeled with specific diseases.

Second: Matching Numbers but Not Physiology. A study of 48 pacemaker recipients showed that atrial, AV, and ventricular pacing modes significantly affect the “heart rate-ejection duration” relationship ($P = 0.01$); even with fixed heart rates, different pacing modes showed significant differences in AIx and pulse pressure amplification ($P < 0.0001$) (PMID 27471239). Therefore, heart rate matching must simultaneously lock the source of the beat.

Another Reverse Trap

Normalization can strip heart rate but may also discard true information. A 2023 methodological study comparing waveforms found that without normalization, the actual morphological differences at the tail ends of two waveforms were minimal; however, due to different heart rates, the tails became the primary source of cumulative distance, completely drowning out morphological differences; only with double normalization did the focus truly shift to morphology—at the cost of removing heart rate and pulse pressure from main features, relegating them to secondary calculation (PMID 36767962). “Controlling heart rate” and “discarding heart rate information” are two distinct things; the latter would inadvertently destroy the discriminative basis for Pulse Counting and Late Pulse.

Heart rate itself is not clean: a study of 52 pacemaker recipients showed that after stripping blood pressure effects, carotid-femoral pulse wave velocity still increased by 0.16-0.20 m/s per 10 bpm (PMID 27245180).

How to Read the Results

Criteria are fixed in advance to avoid post-hoc interpretation:

  • Retained discriminative power ≥ 80% → The category is primarily a true shape signal; heart rate is not the main driver.
  • Retained 50-80% → Mixed type; must explicitly state in reports that it “contains a heart rate component” and provide the proportion contributed by heart rate.
  • Retained < 50% → Primarily a disguise of heart rate; existing literature discriminative power for this category should be downgraded overall.
  • Positive Control (Pulse Counting/Late Pulse) did not collapse → Pipeline leakage exists; all conclusions are void and must be re-run.

This is not another article “proving pulse diagnosis is useful.” It performs an audit: before handing pulse qualities to any classifier, first confirm the classifier is learning the pulse quality, not the speed of the heartbeat.

Evidence and Limitations

Verified: Significant negative linear correlation between AIx and heart rate under increasing pacing, ~4% drop per 10 bpm increase, ejection duration correlation $r = -0.51$ (PMID 10811742); AIx correlation with ejection duration $r = 0.88$, ejection duration explains 78% variance, strain rate unrelated to AIx after heart rate adjustment (PMID 19720955); adjusting AIx to 75 bpm is inappropriate physiologically and statistically (PMID 24257465); time-based parameters in stress pulse waves driven by heart rate, amplitude-based by systolic pressure (PMID 26406041); SDPTG D:A and AGI negatively correlated with heart rate, B:A positively correlated, associations hold after age/BP adjustment (PMID 15831356, PMID 12473866); DTW without normalization is dominated by heart rate in waveform distance, normalization removes heart rate and pulse pressure info (PMID 36767962); significant AIx differences across pacing modes even at fixed heart rates (PMID 27471239); pulse wave velocity increases 0.16-0.20 m/s per 10 bpm after stripping blood pressure effects (PMID 27245180); floating pulse automatic threshold in 45 hypertensives fell at heart rate ≥ 58.25 (PMID 26988431).

Partially Verified: The inference that “a significant proportion of shape-based pulse discriminative power comes from heart rate” is a reasonable extrapolation from the above to pulse classification, but no study has directly performed heart rate-matched pulse discrimination controls in the same cohort; the threshold of “≥ 80% indicates true shape signal” is an adjudication standard set by this paper, not existing consensus.

To Be Verified: ① Actual retained AUC proportions for Floating/Sunken/Wiry/Slippery after heart rate matching; ② Whether Pulse Counting/Late Pulse truly collapse to 0.5 as positive controls (indicating pipeline leakage); ③ Whether heart rate matching introduces collision bias that creates spurious associations in this specific context.

Related Reading

Evidence Level: B (Direct literature exists for the dominant role of heart rate on waveform timing and shape, AIx’s chronotropic dependence, SDPTG’s heart rate dependence, and the double-edged nature of normalization; “heart rate matching separation test” for pulse classification has no precedent and is a research proposal).
References: PMID 10811742, PMID 19720955, PMID 24257465, PMID 26406041, PMID 15831356, PMID 12473866, PMID 36767962, PMID 27471239, PMID 27245180, PMID 26988431


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 *