5D Library › 自主神经交感迷走平衡、六经辨证状态空间与多模态标志物反演 · English
第六章 测量的极限:摄像头抖动与筛查的阳性预测值
摄像头会把 RMSSD 放大
RMSSD 衡量的是相邻心跳间期之差,正是最容易被计时误差放大的量。摄像头只能把每一次心跳定位在帧格之内,再加上肤色、光线与动作带来的误差。如果每一个心跳时刻都带有标准差为 的独立误差,相邻间期之差的误差方差是 ,于是观测到的 RMSSD 是 。每秒 30 帧时,仅取整误差就有 ,约 9.6 毫秒。
推导
# Chapter 6. p11: What a camera does to RMSSD. A camera places each beat
# only to within its frame grid and its own noise. If every beat time
# carries an independent error of sd e, each interval error is
# e(n+1) - e(n), and each successive difference of intervals carries
# e(n+2) - 2 e(n+1) + e(n), of variance 6 e^2. So
# RMSSD_seen = sqrt(RMSSD^2 + 6 e^2).
# At 30 frames a second rounding alone gives e = 33.3 / sqrt(12) ms; skin
# tone, light and motion add more. Checked on 20 x 300 beats per case,
# seed 86.
import random
from math import sqrt
def seen(true_rmssd, e, rng, n=300):
rr = [900.0 + rng.gauss(0, true_rmssd / sqrt(2)) for _ in range(n)]
t, times = 0.0, []
for r in rr:
t += r
times.append(t + rng.gauss(0, e))
obs = [b - a for a, b in zip(times, times[1:])]
return sqrt(sum((b - a) ** 2 for a, b in zip(obs, obs[1:]))
/ (len(obs) - 1))
if __name__ == "__main__":
rng = random.Random(86)
print(" true RMSSD timing sd formula simulated inflation")
for true in (15.0, 40.0):
for e in (33.3 / sqrt(12), 15.0, 25.0):
f = sqrt(true ** 2 + 6 * e * e)
s = sum(seen(true, e, rng) for _ in range(20)) / 20
print(" {:10.0f} {:9.1f} {:7.1f} {:9.1f} {:+8.0%}"
.format(true, e, f, s, f / true - 1))
# true RMSSD timing sd formula simulated inflation
# 15 9.6 27.9 28.3 +86%
# 15 15.0 39.7 39.4 +165%
# 15 25.0 63.0 63.9 +320%
# 40 9.6 46.4 46.2 +16%
# 40 15.0 54.3 54.0 +36%
# 40 25.0 73.1 72.9 +83%真实 RMSSD 为 15 毫秒的人(迷走张力偏低,正是筛查最关心的人),在 9.6 毫秒的计时误差下会被读成 28.3 毫秒左右,放大 86%;计时误差到 25 毫秒时读成约 63.9,放大 320%。真实 RMSSD 为 40 的人放大比例小得多。换句话说,摄像头的误差会把最需要被看见的低 RMSSD 者「抬」成看似正常,公式与模拟的结果彼此吻合,这是一个系统性偏差,不是随机噪声 (Shaffer and Ginsberg 2017)。
筛查不是诊断
即使筛查本身相当准确,阳性结果的意义仍取决于受筛人群中真正有问题的人有多少。p12 假定灵敏度 0.80、特异度 0.85(示意),并把第五章三模态的标签正确率 1 减 0.298 也算进去。
推导
# Chapter 6. p12: A screen is not a diagnosis. Suppose the three-minute
# screen flags "needs a clinical look" with sensitivity 0.80 and
# specificity 0.85 (illustrative). Positive predictive value depends on
# how common the condition is among those screened. Among true positives
# the pattern label is right with probability 1 - 0.298 (p09, three
# modalities), so a flag with the right label is PPV * 0.702.
SENS, SPEC, LABEL = 0.80, 0.85, 1 - 0.298
if __name__ == "__main__":
print(" prevalence PPV NPV false flags per 1,000"
" right label")
for prev in (0.02, 0.05, 0.10, 0.20, 0.40):
tp = SENS * prev
fp = (1 - SPEC) * (1 - prev)
tn = SPEC * (1 - prev)
fn = (1 - SENS) * prev
ppv, npv = tp / (tp + fp), tn / (tn + fn)
print(" {:10.0%} {:5.3f} {:5.3f} {:21.0f} {:11.3f}".format(
prev, ppv, npv, 1000 * fp, ppv * LABEL))
# prevalence PPV NPV false flags per 1,000 right label
# 2% 0.098 0.995 147 0.069
# 5% 0.219 0.988 143 0.154
# 10% 0.372 0.975 135 0.261
# 20% 0.571 0.944 120 0.401
# 40% 0.780 0.864 90 0.548在患病率 2% 的一般人群里,阳性预测值只有 0.098:被标记的人里十个只有一个真正需要进一步评估,而每一千人中有 147 人被误标;既被标记、又被贴对模式标签的比例只有 0.069。患病率升到 40%(例如已经转介到专科的人群)时,阳性预测值才到 0.780。同一个工具,用在不同人群,意义完全不同 (Buderer 1996)。
利益披露
医学声明: 本书讨论的是测量与推断的数学,不是医疗建议,也不是诊断工具。书中提到的任何筛查方案都没有经过临床验证。任何身体或情绪上的困扰,请向合格的医疗专业人员求助。
小结
摄像头计时误差对低 RMSSD 者的放大最大,恰好掩盖了筛查最想找出的人;而在一般人群中,即使准确的筛查,阳性结果多数也是误报。任何把「三分钟手机检测」当作诊断来宣传的做法,都忽略了这两项限制。
参考文献
Fan, Y. (2026). Autonomic Sympathovagal Balance, Six-Meridian State-Space Models and Multimodal Biomarker Inversion: Heart-Rate Variability Spectral Decomposition, Thermoregulatory Phase Drift and Bayesian Syndrome Classification [自主神经交感迷走平衡、六经辨证状态空间与多模态标志物反演] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090227 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).