关于「同一餐,不同的曲线:个体差异的方差分解与可预测上限」,《连续血糖曲线与最小模型》给出的定量结论是什么?

结论:第二章以泽维等人对 800 人、46,898 餐的记录为出发点,把个体差异分解为人、人与食物交互和重复噪声:示意模型中单次比较只有 0.610 是真实差异;平均差很小的两种食物,只有 0.611 的人个人排序与平均一致,平均差大时为 0.921。

数据口径:本书不是医疗建议,不宣称任何功效,不提供任何饮食、禁食或用药安排;任何自我试验须先咨询执业医师。 书中参数为示意;全部计算由书中程序生成。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.611, 0.921

(a) share of people whose ranking agrees with the average
d ratio true1 test 2 tests 4 tests
0.101.105 0.611 0.587 0.597 0.604
0.301.350 0.802 0.746 0.770 0.785
0.501.649 0.921 0.865 0.891 0.905
(b) correlation ceiling sqrt(rel) and reported R as its share
rel ceiling0.596/ceil 0.395/ceil
0.600.775 0.769 0.510
0.700.837 0.712 0.472
0.800.894 0.666 0.442
R^2: personal 0.355, carbohydrate 0.156
check: perfect predictor, rel 0.7, simulated r0.839
查看完整验算程序
# Chapter 2. p22: two ceilings on "predicting one person's curve".
# (a) Ranking: food 2 raises log iAUC by d on average. A person's own
#     gap is d + (b2 - b1), sd sqrt(2) * SB; with R repeated tests the
#     observed gap adds noise of sd sqrt(2 / R) * SE (sds as in p21).
# (b) Correlation: if one meal response has test-retest reliability
#     rel, no predictor can correlate with it above sqrt(rel). R values
#     0.596 (personal model) and 0.395 (carbohydrate) are the ones
#     reported by Mendes-Soares et al. 2019 for 327 people; the model
#     came from Zeevi et al. 2015 (800 people, 46,898 logged meals).
#     The rel values are illustrative.
import math
import random
from statistics import NormalDist

PHI = NormalDist().cdf
SB, SE = 0.25, 0.20
print("(a) share of people whose ranking agrees with the average")
print("    d     ratio   true    1 test  2 tests  4 tests")
for d in (0.10, 0.30, 0.50):
    true = PHI(d / (math.sqrt(2.0) * SB))
    obs = [PHI(d / math.sqrt(2.0 * SB ** 2 + 2.0 * SE ** 2 / r))
           for r in (1, 2, 4)]
    print("  %.2f  %6.3f   %.3f   %.3f   %.3f    %.3f"
          % (d, math.exp(d), true, obs[0], obs[1], obs[2]))
print("(b) correlation ceiling sqrt(rel) and reported R as its share")
print("   rel   ceiling   0.596/ceil  0.395/ceil")
for rel in (0.6, 0.7, 0.8):
    c = math.sqrt(rel)
    print("  %.2f   %.3f     %.3f       %.3f"
          % (rel, c, 0.596 / c, 0.395 / c))
print("  R^2: personal %.3f, carbohydrate %.3f"
      % (0.596 ** 2, 0.395 ** 2))
# 蒙特卡洛核对:完美预测真值,测量信度 0.7,相关系数应接近 0.837
rng = random.Random(20261115)
n = 20000
noise = math.sqrt(1.0 / 0.7 - 1.0)
xs = [rng.gauss(0.0, 1.0) for _ in range(n)]
ys = [x + rng.gauss(0.0, noise) for x in xs]
mx, my = math.fsum(xs) / n, math.fsum(ys) / n
sxy = math.fsum((x - mx) * (y - my) for x, y in zip(xs, ys))
sxx = math.fsum((x - mx) ** 2 for x in xs)
syy = math.fsum((y - my) ** 2 for y in ys)
print("check: perfect predictor, rel 0.7, simulated r = %.3f"
      % (sxy / math.sqrt(sxx * syy)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《连续血糖曲线下面积、双室胰岛素最小模型与生酮代谢切换动力学母典》第 2 章「同一餐,不同的曲线:个体差异的方差分解与可预测上限」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-286 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026286,
  author    = {Fan, Yafu},
  title     = {Areas under Continuous Glucose Curves, the Two-Compartment Minimal Model of Insulin Action and the Dynamics of the Switch to Ketone Metabolism},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论