5D Library › 胰岛素胰高血糖素双稳态、内脏脂解动力学与线性规划配餐 · English
第四章 两项便宜的测量
一把皮尺对一次血检
前几章讲的都是胰岛素已经在闹脾气的那些人,于是有一个很实际的问题:你怎么把他们找出来?昂贵的答案是钳夹试验。便宜的答案是一把皮尺与一次空腹血检,而关于两者各自表现如何,是有文献的。Ashwell、Gunn 与 Gibson 的荟萃分析发现,就心血管代谢风险的筛查而言,腰高比优于腰围与体质指数,并支持那条熟悉的建议:把腰围保持在身高的一半以下 (Ashwell et al. 2012)。血液那一侧,Simental-Mendia 等人提出用空腹甘油三酯与空腹葡萄糖之积作为胰岛素抵抗的替代指标 (Simental-Mendia et al. 2008),随后 Guerrero-Romero 等人把这个指数直接与正常血糖高胰岛素钳夹作了比较,而这才是要紧的那次比较 (Guerrero-Romero et al. 2010)。
证据
已发表。 Ashwell、Gunn 与 Gibson 的系统综述与荟萃分析发现,在筛查成年人心血管代谢风险上,腰高比优于腰围与体质指数 (Ashwell et al. 2012)。Simental-Mendia 等人引入空腹甘油三酯与葡萄糖之积作为替代测量 (Simental-Mendia et al. 2008)。Guerrero-Romero 等人以钳夹验证了该指数 (Guerrero-Romero et al. 2010)。
注
这份样本是什么。 本章里的人并不存在。两组等量的人由一个带共享潜在严重度的模型生成,使腰高比与血液指标像现实中那样相关,组均值则取自已发表的成年人区间。这里没有任何一项是对两个指标在任何真实人群中表现的估计。算术能显示的是:在已经握有第一个指标之后,第二个指标最多还能添加多少,而这是一个关于相关测量之间几何关系的问题,不是关于某个队列的问题。
推导
# Chapter 4. p07: How well waist-to-height ratio separates the insulin
# resistant from the insulin sensitive. The sample is synthetic, not a
# cohort: two groups of equal size, a shared latent severity that makes
# the markers correlate, and group means taken from the ranges reported
# for adults. The curve of sensitivity against one minus specificity is
# summarised by the area under it, computed from ranks, and the cutoff
# is the one that maximises sensitivity plus specificity minus one.
import random
SEED, N = 95, 4000
WHTR0, WHTR_G, WHTR_SD = 0.485, 0.100, 0.050
LAT = 0.6 # weight of the shared severity term
def sample():
rng = random.Random(SEED)
rows = []
for g in (0, 1):
for _ in range(N):
z = rng.gauss(0.0, 1.0)
ew = rng.gauss(0.0, 1.0)
et = rng.gauss(0.0, 1.0)
ef = rng.gauss(0.0, 1.0)
mix = LAT * z + (1.0 - LAT ** 2) ** 0.5 * ew
whtr = WHTR0 + WHTR_G * g + WHTR_SD * mix
rows.append((g, whtr, z, et, ef))
return rows
def auc(pos, neg):
both = sorted([(v, 1) for v in pos] + [(v, 0) for v in neg])
ranks, i = [0.0] * len(both), 0
while i < len(both):
j = i
while j + 1 < len(both) and both[j + 1][0] == both[i][0]:
j += 1
r = (i + j) / 2.0 + 1.0
for k in range(i, j + 1):
ranks[k] = r
i = j + 1
s = sum(r for r, (_, lab) in zip(ranks, both) if lab == 1)
n1, n0 = len(pos), len(neg)
return (s - n1 * (n1 + 1) / 2.0) / (n1 * n0)
def roc_point(pos, neg, cut):
sens = sum(1 for v in pos if v >= cut) / len(pos)
spec = sum(1 for v in neg if v < cut) / len(neg)
return sens, spec
if __name__ == "__main__":
rows = sample()
pos = [w for g, w, _, _, _ in rows if g == 1]
neg = [w for g, w, _, _, _ in rows if g == 0]
print(" {} resistant and {} sensitive, synthetic".format(len(pos),
len(neg)))
print(" mean waist-to-height ratio: resistant {:.4f},"
" sensitive {:.4f}".format(sum(pos) / len(pos),
sum(neg) / len(neg)))
a = auc(pos, neg)
print(" area under the curve: {:.4f}".format(a))
best, best_cut = -1.0, 0.0
for i in range(2001):
cut = 0.35 + 0.00025 * i
sens, spec = roc_point(pos, neg, cut)
j = sens + spec - 1.0
if j > best:
best, best_cut = j, cut
sens, spec = roc_point(pos, neg, best_cut)
print(" best cutoff {:.4f}: sensitivity {:.4f}, specificity {:.4f}"
.format(best_cut, sens, spec))
print(" that cutoff scores sensitivity plus specificity minus one"
" at {:.4f}".format(best))
print(" cutoff sensitivity specificity sum minus one")
for cut in (0.4800, 0.5000, 0.5200, 0.5400, 0.5600, 0.6000):
sens, spec = roc_point(pos, neg, cut)
print(" {:<9.4f}{:<14.4f}{:<14.4f}{:.4f}"
.format(cut, sens, spec, sens + spec - 1.0))
# 4000 resistant and 4000 sensitive, synthetic
# mean waist-to-height ratio: resistant 0.5857, sensitive 0.4841
# area under the curve: 0.9244
# best cutoff 0.5317: sensitivity 0.8568, specificity 0.8403
# that cutoff scores sensitivity plus specificity minus one at 0.6970
# cutoff sensitivity specificity sum minus one
# 0.4800 0.9778 0.4675 0.4453
# 0.5000 0.9533 0.6252 0.5785
# 0.5200 0.9008 0.7702 0.6710
# 0.5400 0.8177 0.8715 0.6892
# 0.5600 0.6980 0.9413 0.6393
# 0.6000 0.3910 0.9888 0.3798在这份样本里,抵抗组的平均腰高比是 0.5857,敏感组是 0.4841,曲线下面积为 0.9244。使敏感度加特异度减一取最大的界值是 0.5317,给出敏感度 0.8568 与特异度 0.8403,得分 0.6970。那条熟悉的「腰围不过身高之半」的建议,对应表中的界值 0.5000,它在这里抓住抵抗组的 0.9533,代价是把敏感组里除 0.6252 之外的人全部标出来。筛查阈值存在的意义正是这桩交换:0.5000 是一个偏敏感的界值,而特异度 0.9413 的 0.5600 是一个偏特异的界值。
那个指数,以及它添加了什么
这个指数是空腹甘油三酯与空腹葡萄糖之积除以二,再取自然对数。p08 用完全相同的生成器与完全相同的种子抽样,所以它看的是同一批人。
推导
# Chapter 4. p08: The triglyceride-glucose index on the same synthetic
# sample as p07. The index is the natural logarithm of fasting
# triglycerides times fasting glucose divided by two, with both in
# mg/dL, so it is the sum of the two logarithms less the logarithm of
# two. The generator, seed and draw order are identical to p07, so the
# two programs see the same people. Reported: the area under the curve,
# the best cutoff, and what a two-step rule gains over either marker.
import math
import random
SEED, N = 95, 4000
WHTR0, WHTR_G, WHTR_SD = 0.485, 0.100, 0.050
TG0, TG_G, TG_SD = 90.0, 175.0, 0.30
FPG0, FPG_G, FPG_SD = 90.0, 108.0, 0.08
LAT = 0.6
def sample():
rng = random.Random(SEED)
rows = []
for g in (0, 1):
for _ in range(N):
z = rng.gauss(0.0, 1.0)
ew = rng.gauss(0.0, 1.0)
et = rng.gauss(0.0, 1.0)
ef = rng.gauss(0.0, 1.0)
mix = LAT * z + (1.0 - LAT ** 2) ** 0.5 * ew
whtr = WHTR0 + WHTR_G * g + WHTR_SD * mix
ltg = (math.log(TG0) + math.log(TG_G / TG0) * g
+ TG_SD * (LAT * z + (1.0 - LAT ** 2) ** 0.5 * et))
lfpg = (math.log(FPG0) + math.log(FPG_G / FPG0) * g
+ FPG_SD * (LAT * z + (1.0 - LAT ** 2) ** 0.5 * ef))
rows.append((g, whtr, ltg + lfpg - math.log(2.0)))
return rows
def auc(pos, neg):
both = sorted([(v, 1) for v in pos] + [(v, 0) for v in neg])
ranks, i = [0.0] * len(both), 0
while i < len(both):
j = i
while j + 1 < len(both) and both[j + 1][0] == both[i][0]:
j += 1
r = (i + j) / 2.0 + 1.0
for k in range(i, j + 1):
ranks[k] = r
i = j + 1
s = sum(r for r, (_, lab) in zip(ranks, both) if lab == 1)
n1, n0 = len(pos), len(neg)
return (s - n1 * (n1 + 1) / 2.0) / (n1 * n0)
def point(pos, neg, cut):
sens = sum(1 for v in pos if v >= cut) / len(pos)
spec = sum(1 for v in neg if v < cut) / len(neg)
return sens, spec
if __name__ == "__main__":
rows = sample()
pos = [t for g, _, t in rows if g == 1]
neg = [t for g, _, t in rows if g == 0]
print(" mean index: resistant {:.4f}, sensitive {:.4f}"
.format(sum(pos) / len(pos), sum(neg) / len(neg)))
print(" area under the curve: {:.4f}".format(auc(pos, neg)))
best, best_cut = -1.0, 0.0
for i in range(3001):
cut = 7.00 + 0.001 * i
sens, spec = point(pos, neg, cut)
if sens + spec - 1.0 > best:
best, best_cut = sens + spec - 1.0, cut
sens, spec = point(pos, neg, best_cut)
print(" best cutoff {:.3f}: sensitivity {:.4f}, specificity {:.4f}"
.format(best_cut, sens, spec))
print(" cutoff sensitivity specificity sum minus one")
for cut in (8.300, 8.500, 8.600, 8.700, 8.800, 9.000):
sens, spec = point(pos, neg, cut)
print(" {:<9.3f}{:<14.4f}{:<14.4f}{:.4f}"
.format(cut, sens, spec, sens + spec - 1.0))
wp = [w for g, w, _ in rows if g == 1]
wn = [w for g, w, _ in rows if g == 0]
print(" waist-to-height ratio on the same people: area {:.4f}"
.format(auc(wp, wn)))
gate = 0.5200
both_pos = sum(1 for g, w, t in rows
if g == 1 and w >= gate and t >= best_cut)
both_neg = sum(1 for g, w, t in rows
if g == 0 and not (w >= gate and t >= best_cut))
print(" two-step rule, waist ratio at or above {:.2f} and then the"
" index:".format(gate))
print(" sensitivity {:.4f}, specificity {:.4f}, sum minus one {:.4f}"
.format(both_pos / len(pos), both_neg / len(neg),
both_pos / len(pos) + both_neg / len(neg) - 1.0))
print(" so the blood test adds {:.4f} to the waist measurement alone"
.format(auc(pos, neg) - auc(wp, wn)))
# mean index: resistant 9.1594, sensitive 8.3078
# area under the curve: 0.9645
# best cutoff 8.705: sensitivity 0.9165, specificity 0.8812
# cutoff sensitivity specificity sum minus one
# 8.300 0.9958 0.4883 0.4840
# 8.500 0.9770 0.7183 0.6953
# 8.600 0.9525 0.8115 0.7640
# 8.700 0.9190 0.8780 0.7970
# 8.800 0.8558 0.9290 0.7848
# 9.000 0.6785 0.9825 0.6610
# waist-to-height ratio on the same people: area 0.9244
# two-step rule, waist ratio at or above 0.52 and then the index:
# sensitivity 0.8423, specificity 0.9435, sum minus one 0.7858
# so the blood test adds 0.0400 to the waist measurement alone在同一批人身上,指数的曲线下面积达到 0.9645,而腰高比是 0.9244,指数的最佳界值是 8.705,敏感度 0.9165,特异度 0.8812。这个界值值得一提,因为它落在验证文献所报告的区间之内,这是一道虽弱但真实的核查,说明生成器不是为了讨好这个指数而设的。
最后几行才是有实际后果的。血检给皮尺添加了 0.0400 的面积。一条两步规则,先量腰,只把达到或超过 0.52 的人送去抽血,达到敏感度 0.8423 与特异度 0.9435,得分 0.7858,而对所有人单用指数的得分是 0.7970。所以两步规则在区分能力上只差一点点,却便宜得多,因为它只对少数人开血检。对一个筛查项目来说,这通常是更好的交换;对一个已经坐在诊室里的人来说,就不是。
假说 · 尚未验证
作者的诠释(尚未验证): 因为这两个指标共享了大部分信息,任何已经在量腰的筛查项目都应把空腹血检当作确认性检查而非筛查性检查,而在任何指标相关性与此处一样强的人群里,加上它所带来的增益都会停在此处所得 0.0400 的两倍以内。什么会证明它错: 在一个同时具备两项测量与钳夹的队列里,如果把指数加到腰高比上使面积提高超过该数的两倍,或者两个指标之间的相关弱到它们携带的信息基本独立,那么这条两步建议就不成立。
小结
在一份合成样本上,腰高比的曲线下面积达到 0.9244,指数达到 0.9645,所以血检添加了 0.0400。一条两步规则得分 0.7858,单用指数得分 0.7970,而前者只对少数人开血检。指数的最佳界值 8.705 正落在验证文献给出的位置。第五章从量人转向喂人,先从低碳水的一周所拿走的那种矿物质开始。
参考文献
Fan, Y. (2026). Insulin-Glucagon Bistability, Visceral Lipolysis Kinetics and Linear-Programming Diet Optimization: Hepatic Glycogen Depletion Times, Triglyceride-Glucose Indices and Micronutrient-Constrained Menus [胰岛素胰高血糖素双稳态、内脏脂解动力学与线性规划配餐] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090248 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).