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

第四章 六经作为隐马尔可夫状态空间

把「传经」写成转移矩阵

《伤寒论》以太阳、阳明、少阳、太阴、少阴、厥阴六经概括外感病的阶段,并区分依次相传(顺传)、跳过一经(越经)与两经同病(合病)等情形(第七章引原文)。把它写成隐马尔可夫模型 (Rabiner 1989):隐藏状态是六经之一,每天按转移概率停留、进入下一经或跳过一经;观测是四项可以量化的生理指标,即第三章的体温节律振幅与核心至皮肤温差,加上胃肠(迷走)蠕动频率与皮质醇觉醒反应(CAR)。每一经在四项指标上有自己的平均值(示意设定),用 Viterbi 算法从十四天的观测反推每天所在的阶段。

推导

# Chapter 4. p07: The six stages as a hidden Markov model. Hidden state:
# taiyang, yangming, shaoyang, taiyin, shaoyin, jueyin (illustrative
# order of ordinary progression). Daily transitions: stay 0.70, move to
# the next stage 0.20 ("sequential transmission"), skip one stage 0.05
# ("skipping"), any other stage 0.05 shared. Each day four measurements
# are seen, in standard units: temperature-rhythm amplitude, core-to-skin
# gradient, gut (vagal) motility rate, cortisol awakening response; each
# stage has its own mean (illustrative) and noise sd 0.8. Viterbi decoding
# with the true model, 3,000 courses of 14 days, seed 86.
import random
from math import log, pi

STAGES = ("taiyang", "yangming", "shaoyang", "taiyin", "shaoyin", "jueyin")
MEAN = ((1.0, 1.2, 0.0, 0.5), (1.5, -0.5, -0.8, 1.0),
        (0.5, 0.3, -0.3, 0.3), (-0.3, 0.0, 1.0, -0.3),
        (-1.0, 0.8, 0.3, -1.0), (-0.6, 1.2, -0.6, -0.7))
SD = 0.8

def trans():
    T = []
    for i in range(6):
        row = [0.05 / 3] * 6
        row[i] = 0.70
        if i + 1 < 6:
            row[i + 1] = 0.20
        if i + 2 < 6:
            row[i + 2] = 0.05
        z = sum(row)
        T.append([x / z for x in row])
    return T

T = trans()

def logpdf(x, m, sd=SD):
    return sum(-0.5 * ((a - b) / sd) ** 2 - log(sd) - 0.5 * log(2 * pi)
               for a, b in zip(x, m))

def course(rng, days=14):
    s, xs, ss = 0, [], []
    for _ in range(days):
        ss.append(s)
        xs.append([m + rng.gauss(0, SD) for m in MEAN[s]])
        u, acc = rng.random(), 0.0
        for j in range(6):
            acc += T[s][j]
            if u < acc:
                s = j
                break
    return ss, xs

def viterbi(xs, start=(1.0, 0, 0, 0, 0, 0)):
    lt = [[log(p) for p in row] for row in T]
    v = [(log(start[j]) if start[j] > 0 else -1e9) + logpdf(xs[0], MEAN[j])
         for j in range(6)]
    back = []
    for x in xs[1:]:
        nv, bk = [], []
        for j in range(6):
            i = max(range(6), key=lambda i: v[i] + lt[i][j])
            nv.append(v[i] + lt[i][j] + logpdf(x, MEAN[j]))
            bk.append(i)
        v, back = nv, back + [bk]
    s = max(range(6), key=lambda j: v[j])
    path = [s]
    for bk in reversed(back):
        s = bk[s]
        path.append(s)
    return path[::-1]

if __name__ == "__main__":
    rng = random.Random(86)
    hit = tot = 0
    by = [[0, 0] for _ in range(6)]
    conf = {}
    for _ in range(3000):
        ss, xs = course(rng)
        est = viterbi(xs)
        for a, b in zip(ss, est):
            hit += a == b
            tot += 1
            by[a][0] += a == b
            by[a][1] += 1
            if a != b:
                conf[(a, b)] = conf.get((a, b), 0) + 1
    print("   days decoded correctly: {:.3f}".format(hit / tot))
    for i, n in enumerate(STAGES):
        print("   {:<9} {:.3f}  ({} days)".format(n, by[i][0] / by[i][1],
                                                 by[i][1]))
    a, b = max(conf, key=conf.get)
    print("   most common mix-up: {} read as {}".format(STAGES[a],
                                                     STAGES[b]))
#    days decoded correctly: 0.830
#    taiyang   0.916  (11342 days)
#    yangming  0.871  (7267 days)
#    shaoyang  0.727  (6638 days)
#    taiyin    0.812  (5770 days)
#    shaoyin   0.700  (4695 days)
#    jueyin    0.853  (6288 days)
#    most common mix-up: shaoyin read as jueyin

在模型完全正确的情况下,按天计算的正确率是 0.830。太阳最容易认出(0.916),少阴最难(0.700),最常见的错误是把少阴读成厥阴。少阳的 0.727 也偏低:它在四项指标上夹在太阳与阳明之间。

可辨识性:哪些阶段本来就分不开

两个阶段能否分开,取决于它们的平均值相距多远,相对于测量噪声有多大。p08 用 Bhattacharyya 距离 DD 给出单日误判率的上界 e−D/2e^{-D}/2,然后让噪声变大,看 Viterbi 的正确率如何下降。

推导

# Chapter 4. p08: How far apart are the stages? For two stages with
# equal-noise Gaussian measurements the Bhattacharyya distance is
# D = |m1 - m2|^2 / (8 sd^2), and a one-day Bayes error (equal priors) is
# at most exp(-D) / 2. Then the Viterbi accuracy of p07 as measurement
# noise grows, and the size of the model if "combined disease" (two
# stages at once) needs its own states: 6 + 15 = 21 states.
import importlib.util
import random
from itertools import combinations
from math import exp, log, pi

def load(name):
    spec = importlib.util.spec_from_file_location(name, name + ".py")
    m = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(m)
    return m

p07 = load("p07_six_stage_hmm")

def bhatt(a, b, sd):
    return sum((x - y) ** 2 for x, y in zip(a, b)) / (8 * sd * sd)

def accuracy(sd, n=1500, seed=86):
    p07.SD = sd
    p07.logpdf = lambda x, m: sum(-0.5 * ((a - b) / sd) ** 2 - log(sd)
                                  - 0.5 * log(2 * pi) for a, b in zip(x, m))
    rng = random.Random(seed)
    hit = tot = 0
    for _ in range(n):
        ss, xs = p07.course(rng)
        est = p07.viterbi(xs)
        hit += sum(a == b for a, b in zip(ss, est))
        tot += len(ss)
    return hit / tot

if __name__ == "__main__":
    pairs = sorted(combinations(range(6), 2),
                   key=lambda p: bhatt(p07.MEAN[p[0]], p07.MEAN[p[1]], 0.8))
    print("   closest pairs at sd 0.8    D      one-day error bound")
    for i, j in pairs[:3]:
        d = bhatt(p07.MEAN[i], p07.MEAN[j], 0.8)
        print("   {:<9} vs {:<9}  {:6.3f}  {:.3f}".format(
            p07.STAGES[i], p07.STAGES[j], d, exp(-d) / 2))
    for sd in (0.6, 0.8, 1.2, 1.6):
        print("   noise sd {:.1f}: Viterbi accuracy {:.3f}".format(
            sd, accuracy(sd)))
    print("   states with every two-stage combination: {} (transitions {})"
          .format(6 + 15, 21 * 20))
#    closest pairs at sd 0.8    D      one-day error bound
#    taiyang   vs shaoyang    0.232  0.396
#    shaoyin   vs jueyin      0.238  0.394
#    taiyin    vs shaoyin     0.412  0.331
#    noise sd 0.6: Viterbi accuracy 0.909
#    noise sd 0.8: Viterbi accuracy 0.829
#    noise sd 1.2: Viterbi accuracy 0.693
#    noise sd 1.6: Viterbi accuracy 0.589
#    states with every two-stage combination: 21 (transitions 420)

最接近的两对是太阳与少阳(DD = 0.232,单日误判上界 0.396)以及少阴与厥阴(0.238,0.394):只看一天,这两对几乎和抛硬币一样。Viterbi 之所以还能做到 0.830,靠的是前后几天的连续性。噪声标准差从 0.6 增加到 1.6 时,正确率从 0.909 降到 0.589。若要让「合病」有自己的状态,六个单经再加上十五种两经组合共 21 个状态,转移概率多达 420 个,在可以收集到的病程数据量下几乎无法估计。

Viterbi 解码正确率随测量噪声增加而下降(p08)。

假说 · 尚未验证

作者的诠释(尚未验证): 六经病程可以用一个以体温节律、核心至皮肤温差、胃肠蠕动与皮质醇觉醒反应为观测的隐马尔可夫模型描述,且模型能重现「顺传多于越经」的转移结构。什么会证明它错: 在由医师逐日独立判定六经的前瞻性病程数据中,以这四项指标拟合的模型对医师判定的预测不优于只用天数的基线模型,或拟合出的转移矩阵不显示顺传多于越经。

小结

把六经写成状态空间,可以明确地问:哪些阶段在可测量的指标上本来就分不开,需要多少噪声以下的测量才能可靠地追踪病程。在示意参数下,太阳与少阳、少阴与厥阴最难分辨;「合病」若单独建模,参数量会超过数据能支持的范围。

参考文献

Rabiner, Lawrence R. 1989. “A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition.” Proceedings of the IEEE 77 (2): 257 to 286.

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).