那个 56% 的心率污染只在「整段只算一次」时成立:每搏都测,它自己就缩到 7%

Written by

in

Language

Evidence level: C

上一轮担心那个 1.43 小时的相位散差里混着心率波动。这一轮把混入项算清楚,发现三处毛病:56% 只在「整段只做一次变换」时成立,判据把差值和分量当成了一回事,而心率还有第二条符号相反的通道。

正文

一、先说结论

上一轮把 1.43 小时的逐阶相位散差当成生理量,担心它混着心率在晃——固定基频做变换而心搏间期本身在变,第 k 阶相位会随基频一起漂,且这一项在所有阶次上同比例出现、精确折叠进散差,所以任何「散差除以阶次是否恒定」的检验都看不见它。

算到底,结论要改三处。第一,那个 0.80 小时(占 56%)只在整段记录只产出一个相位值时成立,逐搏处理时缩到 0.10 小时(7%)。第二,判据用的「两套变换的散差之差」不是分量本身,判据线 0.3 小时对应的其实是 61%,不是括号里写的 21%;拿它自己举的 0.80 去套自己的判据,过不了线。第三,心率影响相位有两条符号相反的通道,第二条抵掉两成,抵多少取决于血管硬度。

二、0.80 小时是一个「只测一次」的数

算法是基频标准差乘传导时间再折算二十四小时:0.167 赫兹乘 0.20 秒乘 24,得 0.80 小时。这个式子隐含一个前提——整段记录只出一个相位值,它带着一个按基频标准差分布的随机偏移。

若分析是逐搏的,第 i 搏的偏移等于该搏基频偏离乘传导时间,跨 N 搏平均后方差按 N 分之一衰减、标准差按根号 N 衰减。N 取 60 搏,0.104 小时,占 7.2%;取一天全量八万余搏,0.003 小时,占 0.2%。

所以 56% 不算错,但它属于「整段傅里叶变换一次」——而这恰是逐阶谐波谱的标准做法:已发表的径向脉搏波谐波协议就是整段取前十一阶、用组内相关系数评重测信度,未做逐搏归一化。问题不在数字,在报数字时没交代它属于哪种流程;换个流程,同一个量差一个量级。

三、差值不是分量

判据原本写:两套变换的散差之差超过 0.3 小时,即心率贡献占 21% 以上。那个 21% 是拿差值除以总散差得来的,但差值不是分量。

设分量方差可加,差值等于总散差减去「总散差平方减分量平方」的平方根。分量取 0.80 时差值 0.245;要让差值达到 0.30,分量得到 0.876 小时,占 61%。也就是说 0.3 这条线守的不是 21%,是 61%。而 0.80(56%)代进去只有 0.245——示例与判据互相否定。

压缩不是巧合:分量越接近总散差,差值被压得越狠。分量 0.30 时差值 0.032,只占一成一;分量 1.20 时 0.652,占五成四。用一个被非线性压缩的量当线性判据,任何固定阈值都会在某一区间失效,且失效方向是「越严重越测不出」。

四、第二条通道符号相反

心率升高不只抬高基频,还让波传得更快。在体模拟显示,血压固定时颈股脉搏波速度每十次每分升 0.21 米每秒,血压自由变动时升 0.66;实测侧,起搏与药物升心率同样抬高该速度,且不是计时算法造成的。传导时间因此缩短,相位往反方向走。

基频通道给正 0.0333 周期,折算 0.800 小时;传导通道给负 0.007 周期,折算负 0.168 小时;净 0.632 小时,打七九折。折扣因子等于一减去「速度相对变化除以基频相对变化」,本例是百分之三点五除百分之十六点七,得 0.21。

麻烦在于它不是常数:同一模拟里低顺应性组每十次每分升 0.26、高顺应性组升 0.16。而公式里只有基频标准差与传导时间,没有硬度。

五、该比斜率,不该比两套变换

既然差值被压缩、又被观测次数与血管硬度双重污染,不如直接测分量。相位偏移等于传导时间乘基频偏移,而基频偏移等于负的基频乘「间期偏差除间期」,代进去得:相位散差(小时)等于负二十四乘传导时间乘基频乘「间期偏差除间期」。代入 0.20 秒与 1.0 赫兹,得负 4.8 小时每秒。

这是测之前就能算出来的数。把每搏相位残差对每搏间期偏差做回归,斜率应落在负 4.8(只有基频通道)或负 3.8(两条都在,含 0.79 折扣)。两者差 1.0,样本量把斜率标准误压到 0.5 以下就能分开。这条路不需要两套变换,也不必先假定哪套是「真值」。

六、这个判据能分辨什么、不能分辨什么

斜率法也有天花板。间期精度受采样率限制:32 赫兹下量化步 31 毫秒、等效噪声约 9 毫秒,占 50 毫秒标准差的百分之十八;已发表的模拟比对显示 32 赫兹只够均值与一项庞加莱指标,128 赫兹才撑得住均方根差与相邻间期差比例。若噪声占一成八,负 4.8 与负 3.8 之间那 1.0 就要求真实间期散布足够大。

更麻烦的是心率本身非平稳:长程相关只在自主平衡的清醒日常状态出现,交感或副交感单方面主导时退化成近白噪。躺着与活动时同一个斜率含义不同,所以报斜率必须同报体位、时段与状态。

七、结论与处方

三条。第一,报心率分量必须同时报「每条记录产出几次独立观测」;56% 与 7% 是同一件事的两个观测次数,不写清等于没报。第二,停用「两套变换的散差之差大于 0.3」这条判据,改用斜率负 4.8 与负 3.8 的两点判别——它把「有没有第二条通道」变成可分辨的二分问题,而不是被压缩过的连续量。第三,心率项系数须带血管硬度修正,跨人群不可直接搬;没有硬度分层的队列,它只能作内部自检。

证据与局限

已验证:心率的机械影响确实分成两条符号相反的通道——传导速度随心率升高而升高(在体模拟每十次每分 0.21 至 0.66 米每秒,实测起搏与药物一致),传导时间随速度升高而缩短;逐阶谐波谱可以整段一次变换取前十一阶并达到良好重测信度;间期检测精度对脉搏率变异性指标的影响已有模拟量化。

部分验证:方差可加假设下的非线性格差是纯代数推演,不依赖任何生理假设,但「实测分量是否落在预测区间」尚无数据集;斜率负 4.8 与负 3.8 的可分辨性依赖间期噪声,本库无同批双通道数据。

待验证:血管硬度修正项在人群中的取值范围;不同体位与睡眠状态下的斜率稳定性;逐搏归一化后组间方差是否真降四成(本轮无数据)。三项都需同批受试者同步记录心搏间期序列与脉搏波、并分层报体位与时段。

相关阅读

证据等级: C
参考文献: PMID 27170904, PMID 32740405, PMID 15642772, PMID 35255373, PMID 17347751
L1 网络状态:本主题尚未进入 L1 对应网络。全表检索「相位 / 心率 / 基频 / 归一化 / 散差 / 折叠」无对应边;仅节点表两行登记相关对象而无层级无状态——物理侧「相位 / 节律 — 周期性耦合的关键」,生理/体验侧「心率变异性 HRV — 可穿戴 — 自主神经窗口」。边表止于 E-061(呼吸节律 0.1 Hz 谱峰—Mayer 波,状态 open),无任何以「相位散差」为源或目标的边。


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。

That 56% heart rate contamination only applies when counting once per segment: measuring every beat reduces it to 7%.

Evidence level: C

In the previous round, we worried that the 1.43-hour phase dispersion error contained heart rate fluctuations. This round clarifies the mixed-in terms and reveals three flaws: the 56% figure only holds when “the entire segment undergoes a single transformation,” because the criterion conflates differences with components, while heart rate possesses a second channel with an opposite sign.

Main Text

I. The Conclusion First

The previous round treated the 1.43-hour stepwise phase dispersion error as a physiological quantity, worrying it was swaying due to heart rate—using a fixed base frequency for transformation while inter-beat intervals themselves vary. Consequently, the phase of the $k$-th order drifts with the base frequency, and this term appears proportionally across all orders, folding precisely into the dispersion error. Thus, any test of “whether dispersion divided by order is constant” fails to see it.

After calculation, three corrections are required. First, that 0.80-hour figure (representing 56%) holds only when the entire record yields a single phase value; under beat-by-beat processing, it shrinks to 0.10 hours (7%). Second, the criterion’s “difference between two transformations” is not the component itself; the criterion line of 0.3 hours actually corresponds to 61%, not the 21% written in parentheses; applying its own 0.80 to this criterion fails the threshold. Third, heart rate affects phase via two channels of opposite sign, with the second canceling 20%, depending on vascular stiffness.

II. The 0.80-Hour Figure Is a “Single-Measurement” Number

The algorithm is: base frequency standard deviation multiplied by conduction time, converted to 24 hours: $0.167 \text{ Hz} \times 0.20 \text{ s} \times 24 = 0.80 \text{ hours}$. This formula implies a premise—the entire record yields only one phase value, carrying a random offset distributed by the base frequency standard deviation.

If analysis is beat-by-beat, the offset for beat $i$ equals that beat’s base frequency deviation multiplied by conduction time. Averaging across $N$ beats reduces variance by $1/N$ and standard deviation by $\sqrt{N}$. For $N=60$ beats, this is 0.104 hours (7.2%); for a full day of ~80,000 beats, it is 0.003 hours (0.2%).

Thus, 56% is not wrong, but it belongs to “a single Fourier transform of the entire segment”—which is precisely the standard procedure for stepwise harmonic spectra: published radial pulse wave harmonic protocols take the first eleven orders from the whole segment, evaluate retest reliability using intraclass correlation coefficients, and do not perform beat-by-beat normalization. The problem lies not in the numbers, but in reporting them without specifying which process they belong to; changing the process shifts the same quantity by an order of magnitude.

III. Differences Are Not Components

The criterion originally states: the difference between dispersions of two transformations exceeds 0.3 hours, implying heart rate contribution is over 21%. That 21% comes from dividing the difference by total dispersion, but the difference is not the component.

Assuming component variances are additive, the difference equals total dispersion minus the square root of (total dispersion squared minus component squared). When the component is 0.80, the difference is 0.245; to reach a difference of 0.30, the component must be 0.876 hours (61%). Thus, the 0.3 line guards against 61%, not 21%. Plugging in 0.80 (56%) yields only 0.245—the example contradicts the criterion.

Compression is not coincidental: as the component approaches total dispersion, the difference is compressed more severely. At component 0.30, the difference is 0.032 (11%); at component 1.20, it is 0.652 (54%). Using a nonlinearly compressed quantity as a linear criterion means any fixed threshold fails in some interval, and the failure direction is “the more severe, the harder to detect.”

IV. The Second Channel Has an Opposite Sign

Rising heart rate not only elevates base frequency but also accelerates wave conduction. In vivo simulations show that, with fixed blood pressure, carotid-femoral pulse wave velocity increases by 0.21 m/s per 10 bpm; with free blood pressure variation, it increases by 0.66 m/s. Measured side, pacing and drugs raise heart rate similarly, elevating this velocity, not due to timing algorithms. Consequently, conduction time shortens, and phase shifts in the opposite direction.

The base frequency channel contributes +0.0333 cycles, converting to 0.800 hours; the conduction channel contributes -0.007 cycles, converting to -0.168 hours; net 0.632 hours, a 79% discount. The discount factor equals one minus (relative velocity change divided by relative base frequency change); in this case, 3.5% divided by 16.7%, yielding 0.21.

The complication is that it is not constant: in the same simulation, low-compliance groups increase by 0.26 m/s per 10 bpm, while high-compliance groups increase by only 0.16 m/s. Yet the formula includes only base frequency standard deviation and conduction time, lacking stiffness.

V. Compare Slopes, Not Two Transformations

Since the difference is compressed and doubly contaminated by observation count and vascular stiffness, it is better to measure the component directly. Phase shift equals conduction time multiplied by base frequency shift, and base frequency shift equals negative base frequency multiplied by (interval deviation divided by interval), yielding: phase dispersion (hours) = $-24 \times \text{conduction time} \times \text{base frequency} \times (\text{interval deviation} / \text{interval})$. Substituting 0.20 s and 1.0 Hz gives -4.8 hours per second.

This is a number calculable before measurement. Regressing each-beat phase residual against each-beat interval deviation, the slope should fall at -4.8 (only base frequency channel) or -3.8 (both channels, including the 0.79 discount). The difference is 1.0; with sufficient sample size, reducing slope standard error below 0.5 distinguishes them. This approach requires no two transformations and does not presuppose which is the “true value.”

VI. What This Criterion Can and Cannot Distinguish

The slope method has limits. Interval precision is constrained by sampling rate: at 32 Hz, quantization step is 31 ms, equivalent noise ~9 ms, occupying 18% of the 50 ms standard deviation; published simulation comparisons show 32 Hz suffices only for mean and one Poincaré index, while 128 Hz supports RMS difference and adjacent interval ratio. If noise occupies 18%, the 1.0 between -4.8 and -3.8 requires sufficiently large true interval variation.

More problematic is that heart rate itself is non-stationary: long-range correlation appears only in awake daily states with autonomic balance, degrading to near-white noise under sympathetic or parasympathetic dominance alone. Lying and active states have different slope meanings, so reporting slope must accompany posture, time segment, and state.

VII. Conclusion and Prescription

Three points. First, reporting heart rate component must simultaneously state “how many independent observations each record yields”; 56% and 7% are the same quantity observed different times, omitting this equals no report. Second, discard the criterion “difference between two transformations exceeds 0.3”; instead use the two-point discrimination of slopes -4.8 and -3.8—it turns “whether there is a second channel” into a binary question, not a compressed continuous quantity. Third, the heart rate coefficient must include vascular stiffness correction; it cannot be directly applied across populations without stiffness stratification; cohorts lacking stiffness layers can only serve as internal checks.

Evidence and Limitations

Verified: Heart rate’s mechanical influence indeed splits into two channels of opposite sign—conduction velocity rises with heart rate (in vivo simulations: 0.21 to 0.66 m/s per 10 bpm, consistent with pacing and drug measurements), conduction time shortens as velocity rises; stepwise harmonic spectra can take the first eleven orders from a single whole-segment transform and achieve good retest reliability; interval detection precision impact on pulse rate variability indices has been simulated quantified.

Partially Verified: Nonlinear difference under additive variance assumption is pure algebraic derivation, relying on no physiological hypothesis, but “whether measured components fall within predicted intervals” lacks datasets; distinguishability of slopes -4.8 and -3.8 depends on interval noise, which this library lacks dual-channel data for.

To Be Verified: Range of vascular stiffness correction terms across populations; slope stability under different postures and sleep states; whether beat-by-beat normalization truly reduces inter-group variance by 40% (no data this round). All three require simultaneous recording of heart beat interval sequences and pulse waves from the same subjects, stratified by posture and time segment.

Further Reading

Evidence Level: C
References: PMID 27170904, PMID 32740405, PMID 15642772, PMID 35255373, PMID 17347751
L1 Network Status: This topic has not yet entered the L1 corresponding network. A full table search for “phase / heart rate / base frequency / normalization / dispersion / folding” yields no corresponding edge; only the node table registers related objects in two rows without hierarchy or state—physically “phase / rhythm — periodic coupling key,” physiologically/experientially “heart rate variability HRV — wearable — autonomic window.” Edge table stops at E-061 (respiratory rhythm 0.1 Hz spectral peak—Mayer wave, state open), with no edges sourced or targeted by “phase dispersion.”


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 *