5D Library › 指—脑皮层放大、远端微针传入与卒中后认知网络重建的计算神经学 · English

第一章 腧穴共现网络与脑网络的牵制控制

论文做了什么

卢顺康论文的研究内容一,是对针刺治疗卒中后轻度认知障碍的临床处方作数据挖掘 (Lu 2024)。

证据

论文的复杂网络分析(据第 1 章)。 检索中国知网、维普、万方与 PubMed 至 2024 年 5 月 31 日,纳入随机对照试验 172 篇(中文 99 篇、英文 73 篇),文献报告的病例合计 14,075 例(第 1.1.6 节)。每篇文献按「一组主穴 + 一组配穴 = 一条处方」提取处方,共 172 条。高频腧穴前五位是百会(127 次,73.83%)、四神聪(68 次)、神庭(63 次)、足三里(51 次)、风池(47 次)(表 1-3)。高权重穴对中,百会—四神聪 65、百会—神庭 59、百会—风池 44(表 1-5)。用 Gephi 对「172 个腧穴节点、466 条边」作网络分析,平均加权度 15.278,平均聚类系数 0.822,网络直径 3(第 1.2.4 节)。综合各项分析,论文把百会、四神聪、神庭、风池、印堂、水沟、内关、足三里、丰隆列为核心穴(结论第 1 条)。

这里有几处文本细节需要先说明。论文中网络的节点数有三个说法:构建词频表时「标注 152 个腧穴名称」(第 1.1.7.1 节),关联规则分析时「将 179 个腧穴进行关联规则分析」(第 1.2.2 节),Gephi 网络则是「172 个腧穴节点」(第 1.2.4 节),而 172 恰好也是文献与处方的条数。第 1.2.1 节说表 1-5 中的 8 对是「权重值>200 的穴对」,但它们的权重是 31 到 65。表 1-3 中内关 43 次对应的频率印为 40.03%,按 43/172 应为 25.00%。这些问题不影响本章的主要论点,但说明从论文中只能取用明确印出的计数,不能反推其他量。

本章要回答两个问题。第一,这个网络描述的是什么?第二,作者框架提出「在头部督脉枢纽之外加上远端井穴的边界驱动,可以把最小控制能量降低一个数量级」,这一说法在什么意义上能够成立?

共现权重从哪里来

先看一个最简单的事实。百会出现在 172 条处方中的 127 条,占 73.8%。如果医生在每条处方中独立地决定用不用每个穴位(完全没有配伍规则),那么任何一个出现 fjf_j 次的穴位,与百会同时出现的期望次数约为 127fj/172≈0.74fj127 f_j / 172 \approx 0.74 f_j。一般地,两穴 ii、jj 在独立选择下的期望共现次数是

Eij=fifjN,N=172,E_{ij} = \frac{f_i f_j}{N}, \qquad N = 172,

观测值与期望值之比 Oij/EijO_{ij}/E_{ij} 称为提升度(lift)。在给定两穴频次的条件下,共现次数服从超几何分布,可以计算精确的检验概率。下面的程序对论文表 1-5 中两穴都列在表 1-3 里的 32 对逐一计算。

推导

# Chapter 1. What an acupoint co-occurrence network can say. Inputs from Lu (2024):
# N = 172 prescriptions, point frequencies (Table 1-3), pair weights (Table 1-5).
# A: expected co-occurrence under independent choice E = f_i f_j / N, lift O/E and an
#    exact hypergeometric test. B: a NULL database in which every point enters every
#    prescription independently with its own frequency, summarised like the thesis.
import math, random

N = 172
F = {"Baihui": 127, "Sishencong": 68, "Shenting": 63, "Zusanli": 51, "Fengchi": 47,
     "Neiguan": 43, "Sanyinjiao": 40, "Taixi": 40, "Shenmen": 38, "Fenglong": 36,
     "Shuigou": 29, "Taichong": 29, "Xuanzhong": 24, "Yintang": 23, "Fengfu": 22}
T15 = """Baihui Sishencong 65|Baihui Shenting 59|Baihui Fengchi 44|Baihui Neiguan 37|
Baihui Zusanli 35|Sishencong Shenting 33|Sishencong Fengchi 32|Baihui Sanyinjiao 31|
Baihui Taixi 30|Baihui Shenmen 30|Baihui Shuigou 27|Baihui Fenglong 26|
Zusanli Sanyinjiao 26|Shenting Fengchi 23|Baihui Yintang 22|Baihui Taichong 22|
Zusanli Fenglong 22|Shenting Neiguan 21|Sishencong Fenglong 21|Zusanli Neiguan 21|
Sishencong Neiguan 21|Shenting Taixi 20|Sishencong Shenmen 20|Sishencong Sanyinjiao 20|
Sishencong Zusanli 20|Fengchi Shenmen 19|Sanyinjiao Taixi 19|Shenting Zusanli 19|
Fengchi Neiguan 19|Baihui Fengfu 18|Baihui Xuanzhong 18|Sishencong Taixi 18"""
PAIRS = [(s.split()[0], s.split()[1], int(s.split()[2]))
         for s in T15.replace("\n", "").split("|")]
HEAD = {"Baihui", "Sishencong", "Shenting", "Fengchi", "Yintang", "Shuigou", "Fengfu"}

def hyper_tail(k, K, n, upper=True):           # P(X >= k) or P(X <= k), exact
    tot = math.comb(N, n)
    ks = range(k, min(K, n) + 1) if upper else range(0, k + 1)
    return sum(math.comb(K, x) * math.comb(N - K, n - x) for x in ks) / tot

def part_a():
    print("A. Reported pair weights (Table 1-5) against independent choice")
    print("   pair                      obs   exp  lift  two-sided p (hypergeometric)")
    rows = []
    for a, b, o in PAIRS:
        e = F[a] * F[b] / N
        p = 2 * min(hyper_tail(o, F[a], F[b]), hyper_tail(o, F[a], F[b], upper=False))
        rows.append((a, b, o, e, o / e, min(1.0, p)))
    for a, b, o, e, l, p in rows[:8] + sorted(rows[8:], key=lambda r: -r[4])[:3]:
        print(f"   {a + '-' + b:24s} {o:5d} {e:5.1f} {l:5.2f}  {p:.4f}")
    lifts = [r[4] for r in rows]
    alpha = 0.05 / len(rows)
    hh = [r[4] for r in rows if r[0] in HEAD and r[1] in HEAD]
    bb = [r[4] for r in rows if r[0] not in HEAD and r[1] not in HEAD]
    print(f"   {len(rows)} pairs: lift < 1 in {sum(l < 1 for l in lifts)}; median lift"
          f" {sorted(lifts)[len(lifts) // 2]:.2f}; Bonferroni-significant"
          f" {sum(r[5] < alpha for r in rows)}")
    print(f"   head-head pairs: mean lift {math.fsum(hh) / len(hh):.2f} (n = {len(hh)});"
          f" body-body pairs: {math.fsum(bb) / len(bb):.2f} (n = {len(bb)})")
    print(f"   Baihui is in {F['Baihui']}/{N} = {F['Baihui'] / N:.1%} of prescriptions,")
    print("   so its weight with any point j is about 0.74 f_j by chance alone")

def tail_freqs(m=137, total=527):
    f = [max(1, round(19 * 0.965 ** k)) for k in range(m)]
    i = 0
    while sum(f) != total:                       # adjust the largest entries by one
        f[i % 20] += 1 if sum(f) < total else -1
        i += 1
    return f

def build(rng):
    names = list(F) + [f"t{k:03d}" for k in range(137)]
    freq = list(F.values()) + tail_freqs()
    db = [[j for j, fj in enumerate(freq) if rng.random() < fj / N] for _ in range(N)]
    w = {}
    for rx in db:
        for x in range(len(rx)):
            for y in range(x + 1, len(rx)):
                w[(rx[x], rx[y])] = w.get((rx[x], rx[y]), 0) + 1
    return names, freq, db, w

def summary(w, thr):
    adj = {}
    for (a, b), c in w.items():
        if c >= thr:
            adj.setdefault(a, set()).add(b); adj.setdefault(b, set()).add(a)
    cc = []
    for v, nb in adj.items():
        k = len(nb)
        if k >= 2:
            links = sum(1 for x in nb for y in nb if x < y and y in adj[x])
            cc.append(2 * links / (k * (k - 1)))
        else:
            cc.append(0.0)
    diam = 0
    for s in adj:                                # BFS from every node
        dist, front = {s: 0}, [s]
        while front:
            nxt = []
            for u in front:
                for v in adj[u]:
                    if v not in dist:
                        dist[v] = dist[u] + 1; nxt.append(v)
            front = nxt
        diam = max(diam, max(dist.values()))
    m = sum(len(nb) for nb in adj.values()) // 2
    return len(adj), m, math.fsum(cc) / len(cc), diam

def spearman(x, y):
    def rank(v):                                 # average ranks for ties
        srt = sorted(v); first = {}
        for i, a in enumerate(srt):
            first.setdefault(a, i)
        return [first[a] + (srt.count(a) - 1) / 2 for a in v]
    rx, ry = rank(x), rank(y); mx, my = math.fsum(rx) / len(rx), math.fsum(ry) / len(ry)
    sxy = math.fsum((a - mx) * (b - my) for a, b in zip(rx, ry))
    sxx = math.fsum((a - mx) ** 2 for a in rx); syy = math.fsum((b - my) ** 2 for b in ry)
    return sxy / math.sqrt(sxx * syy)

def part_b():
    names, freq, db, w = build(random.Random(2024))
    occ = sum(len(r) for r in db)
    print(f"B. Null database: {len(names)} points, {N} prescriptions, {occ:,} point uses")
    print("   (Table 1-4 implies about 1,207 uses: 335 Du-vessel uses = 27.75 %)")
    print("   threshold  nodes  edges  mean degree  avg clustering  diameter")
    for thr in (1, 3, 5, 8):
        n, m, c, d = summary(w, thr)
        print(f"   weight>={thr:<3d} {n:6d} {m:6d} {2 * m / n:12.1f} {c:15.3f} {d:9d}")
    print("   thesis (Gephi): 172 nodes, 466 edges, mean degree 5.4, clustering 0.822,"
          " diameter 3")
    wd = [0] * len(names)
    for (a, b), c in w.items():
        wd[a] += c; wd[b] += c
    top = sorted(range(len(names)), key=lambda i: -wd[i])[:5]
    print("   top 5 by weighted degree:", ", ".join(names[i] for i in top))
    print(f"   Spearman(frequency, weighted degree) = {spearman(freq, wd):.3f}")

if __name__ == "__main__":
    part_a()
    part_b()
# A. Reported pair weights (Table 1-5) against independent choice
#    pair                      obs   exp  lift  two-sided p (hypergeometric)
#    Baihui-Sishencong           65  50.2  1.29  0.0000
#    Baihui-Shenting             59  46.5  1.27  0.0000
#    Baihui-Fengchi              44  34.7  1.27  0.0002
#    Baihui-Neiguan              37  31.8  1.17  0.0499
#    Baihui-Zusanli              35  37.7  0.93  0.4104
#    Sishencong-Shenting         33  24.9  1.32  0.0142
#    Sishencong-Fengchi          32  18.6  1.72  0.0000
#    Baihui-Sanyinjiao           31  29.5  1.05  0.7036
#    Zusanli-Sanyinjiao          26  11.9  2.19  0.0000
#    Zusanli-Fenglong            22  10.7  2.06  0.0000
#    Sanyinjiao-Taixi            19   9.3  2.04  0.0002
#    32 pairs: lift < 1 in 3; median lift 1.27; Bonferroni-significant 8
#    head-head pairs: mean lift 1.32 (n = 9); body-body pairs: 1.99 (n = 4)
#    Baihui is in 127/172 = 73.8% of prescriptions,
#    so its weight with any point j is about 0.74 f_j by chance alone
# B. Null database: 152 points, 172 prescriptions, 1,148 point uses
#    (Table 1-4 implies about 1,207 uses: 335 Du-vessel uses = 27.75 %)
#    threshold  nodes  edges  mean degree  avg clustering  diameter
#    weight>=1      125   1582         25.3           0.752         3
#    weight>=3       62    343         11.1           0.722         3
#    weight>=5       42    172          8.2           0.729         3
#    weight>=8       27     76          5.6           0.463         2
#    thesis (Gephi): 172 nodes, 466 edges, mean degree 5.4, clustering 0.822, diameter 3
#    top 5 by weighted degree: Baihui, Sishencong, Shenting, Zusanli, Fengchi
#    Spearman(frequency, weighted degree) = 0.843

A 部分的结果有两层意思。第一,高权重主要来自高频率:百会—四神聪的 65 次中,独立选择就会产生约 50.2 次,提升度 1.29;百会—足三里的 35 次甚至低于期望的 37.7 次(提升度 0.93,p = 0.41),也就是说,医生用了百会之后并没有更倾向于再用足三里。第二,配伍规则确实存在,但它的形态是流派的形态:32 对中有 8 对在邦费罗尼校正后显著;头部穴之间的平均提升度为 1.32,而足三里、三阴交、太溪、丰隆这几个躯干四肢穴之间的平均提升度高达 1.99(例如足三里—三阴交 26 次,期望 11.9 次)。一种合理的读法是,文献中有「以头部督脉穴为主」与「以辨证取四肢穴为主」两类处方传统,同类穴位倾向于一起出现。

一个没有任何配伍规则的处方库

论文报告的聚类系数 0.822 与直径 3 常被读作「小世界」特性,并进一步被读作腧穴之间存在某种协同。B 部分构造了一个对照:152 个穴位,前 15 个用论文表 1-3 的频次,其余 137 个用递减的小频次补足到总计约 1,207 次使用(这是表 1-4 中督脉 335 次占 27.75% 所隐含的总数),然后让每个穴位按自己的频率独立地进入 172 条处方。这个处方库中没有任何配伍规则。

按共现次数至少为 1、3、5 建网,节点数为 125、62、42,平均聚类系数为 0.752、0.722、0.729,直径都是 3;阈值提高到 8 时网络才变得稀疏。按加权度排名的前五位是百会、四神聪、神庭、足三里、风池——与论文按频次排出的前五位完全相同。原因很简单:每一条处方在共现网络中都是一个「团」(处方中的每两个穴位都连一条边),许多团相互重叠,就自然产生高聚类与短路径;高频穴出现在多数团中,就自然成为高加权度的「枢纽」 (Watts and Strogatz 1998)。网络统计量本身不区分「配伍规律」与「频率」。

左:论文表 1-5 中 32 个穴对的报告权重与独立选择下的期望次数(虚线为 y = x;蓝色为头部穴之间,金色为头部与躯干四肢之间,紫色为躯干四肢穴之间)。右:没有任何配伍规则的模拟处方库中,各穴的频次与加权度。

图 1 中,绝大多数穴对落在 y = x 附近或略高;没有配伍规则时,加权度几乎是频次的函数(斯皮尔曼相关 0.843,尾部大量相同频次拉低了秩相关)。

注

共现网络能回答什么。 它能回答「文献中的医生通常怎样开处方」,并在扣除频率之后找出哪些穴对的共现超出偶然,这需要像提升度、超几何检验或「骨干抽取」之类以零模型为基准的方法 (Serrano et al. 2009)。它不能回答两穴之间是否存在生理上的相互作用,也不能回答哪个穴位更有效:一个穴位在处方中频繁出现,可能因为它有效,也可能因为它是某一流派的标准配穴、操作方便或被教科书推荐。第 25 本讨论过聚类系数与中心性的一般性质,这里不重复。

牵制控制与格拉姆能量

作者框架的第二步,是把控制论用到这个网络上。复杂网络控制的研究主要有三条线索。刘洋宇、斯洛廷与巴拉巴西 2011 年证明,有向网络的结构可控性可以用最大匹配求出所需的最少驱动节点数,而这些驱动节点往往不是高度数的枢纽 (Liu et al. 2011)。汪小帆与陈关荣 2002 年的牵制控制研究的是:在一个耦合网络中,只对少数节点施加反馈 −k(xi−x*)-k(x_i – x^*),能否把整个网络拉到目标状态,以及应当牵制哪些节点 (Wang and Chen 2002)。Yan 等人 2012 年指出,「可控」不等于「容易控制」:把线性系统 ẋ=Ax+Bu\dot{x} = Ax + Bu 从原点推到目标 xfx_f 所需的最小能量是

Emin(T)=xf⊤W(T)−1xf,W(T)=∫0TeAtBB⊤eA⊤tdt,E_{\min}(T) = x_f^{\top}\, W(T)^{-1} x_f, \qquad W(T) = \int_0^T e^{At} B B^{\top} e^{A^{\top} t}\, dt,

其中 W(T)W(T) 是可控性格拉姆矩阵;驱动节点很少时,能量随目标方向可以相差许多个数量级 (Yan et al. 2012; Pasqualetti et al. 2014)。

Gu 等人 2015 年把这套方法用于弥散成像得到的结构脑网络,提出「平均可控性」与「模态可控性」,并认为单个脑区在理论上即可控制整个脑网络 (Gu et al. 2015)。Tu 等人 2018 年用四套数据重做,发现所需节点数总是大于一,平均可控性与加权度的关系在随机化网络中同样存在,并认为用这类指标选择神经刺激靶点的设想应当重新审视 (Tu et al. 2018)。可见即使在真实的脑网络上,「哪里驱动最省能量」也高度依赖建模选择。

一个示意性的脑区网络

把这些方法用到腧穴共现网络上,第一步就不成立:共现网络的边是处方中的共同出现,不是信号传递,网络上没有任何状态变量在按 ẋ=Ax\dot{x} = Ax 演化。所以本章改在一个示意性的脑区网络上讨论作者的问题。这是作者的模型:8 个节点(初级体感皮层 S1、初级运动皮层 M1、丘脑腹后外侧核 VPL、丘脑背内侧核 MD、纹状体 STR、前额叶 PFC、前扣带 ACC、后顶叶 PPC),13 条对称边,权重是示意值,不是由任何影像数据估计的。「远端输入」从 VPL 进入(手部传入的丘脑中继),「头部输入」从 ACC 进入(代表内侧额叶),目标状态是 PFC、ACC、PPC 等量激活;「基底节受损」把所有与纹状体相连的边乘以 0.3。

推导

# Chapter 1. Control energy on a SCHEMATIC 8-region network (the author's model; weights
# illustrative). dx/dt = A x + B u, A = -I + 0.9 W / lambda_max(W). Energy to reach x_f
# (PFC = ACC = PPC) is x_f' G^-1 x_f with G the controllability Gramian, also solved
# EXACTLY in rational arithmetic (Lyapunov equation) since floats fail for one driver.
# Inputs: "distal" at VPL, "head" at ACC; lesion: STR weights x 0.3. B: pinning rate.
import math, random
from fractions import Fraction

NODES = ["S1", "M1", "VPL", "MD", "STR", "PFC", "ACC", "PPC"]
EDGES = {("VPL", "S1"): 1.0, ("S1", "M1"): 0.8, ("S1", "PPC"): 0.7, ("M1", "PPC"): 0.5,
         ("PPC", "PFC"): 0.6, ("PFC", "ACC"): 0.7, ("PFC", "MD"): 0.6, ("ACC", "MD"): 0.4,
         ("PFC", "STR"): 0.7, ("ACC", "STR"): 0.5, ("M1", "STR"): 0.5, ("STR", "MD"): 0.6,
         ("MD", "VPL"): 0.2}
IX = {v: i for i, v in enumerate(NODES)}
n = len(NODES)

def jacobi(S):                                   # eigen-decomposition, symmetric S
    a = [row[:] for row in S]; V = [[float(i == j) for j in range(n)] for i in range(n)]
    for _ in range(100):
        off = math.fsum(a[i][j] ** 2 for i in range(n) for j in range(n) if i != j)
        if off < 1e-22:
            break
        for p in range(n - 1):
            for q in range(p + 1, n):
                if abs(a[p][q]) < 1e-30:
                    continue
                th = 0.5 * math.atan2(2 * a[p][q], a[q][q] - a[p][p])
                c, s = math.cos(th), math.sin(th)
                for k in range(n):
                    akp, akq = a[k][p], a[k][q]
                    a[k][p], a[k][q] = c * akp - s * akq, s * akp + c * akq
                for k in range(n):
                    apk, aqk = a[p][k], a[q][k]
                    a[p][k], a[q][k] = c * apk - s * aqk, s * apk + c * aqk
                for k in range(n):
                    vkp, vkq = V[k][p], V[k][q]
                    V[k][p], V[k][q] = c * vkp - s * vkq, s * vkp + c * vkq
    return [a[i][i] for i in range(n)], V

def weights(scale=None, lesion=1.0):
    W = [[0.0] * n for _ in range(n)]
    for (u, v), w in EDGES.items():
        f = (scale or {}).get((u, v), 1.0) * (lesion if "STR" in (u, v) else 1.0)
        W[IX[u]][IX[v]] = W[IX[v]][IX[u]] = w * f
    return W

def system(W, norm):
    return [[(-1.0 if i == j else 0.0) + 0.9 * W[i][j] / norm for j in range(n)]
            for i in range(n)]

def gramian(A, drivers, T):
    lam, V = jacobi(A)
    b = [[V[IX[dd]][i] for dd in drivers] for i in range(n)]
    M = [[0.0] * n for _ in range(n)]
    for i in range(n):
        for j in range(n):
            s = -(lam[i] + lam[j])
            g = (1 - math.exp(-s * T)) / s if T != math.inf else 1 / s
            M[i][j] = math.fsum(b[i][k] * b[j][k] for k in range(len(drivers))) * g
    return [[math.fsum(V[r][i] * M[i][j] * V[c][j] for i in range(n) for j in range(n))
             for c in range(n)] for r in range(n)]

XF = [1 / math.sqrt(3) if v in ("PFC", "ACC", "PPC") else 0.0 for v in NODES]

def report(A, d, T=math.inf):
    G = gramian(A, d, T); ev, V = jacobi(G); cond = max(ev) / min(ev)
    tr = math.fsum(G[i][i] for i in range(n))
    if min(ev) <= 0 or cond > 1e10:
        return tr, cond, None                     # numerically singular: no energy
    proj = [math.fsum(V[r][k] * XF[r] for r in range(n)) for k in range(n)]
    return tr, cond, math.fsum(p * p / l for p, l in zip(proj, ev))   # x' G^-1 x

def rref(M, m):                                   # exact Gauss-Jordan, m unknowns
    for c in range(m):
        p = next(r for r in range(c, m) if M[r][c] != 0); M[c], M[p] = M[p], M[c]
        M[c] = [x / M[c][c] for x in M[c]]
        for r in range(m):
            if r != c and M[r][c] != 0:
                f = M[r][c]; M[r] = [x - f * y for x, y in zip(M[r], M[c])]
    return [row[-1] for row in M]

def exact_energy(A, drivers):
    idx = [(i, j) for i in range(n) for j in range(i, n)]
    pos = {p: k for k, p in enumerate(idx)}
    F = [[Fraction(x) for x in row] for row in A]; m = len(idx)
    M = [[Fraction(0)] * (m + 1) for _ in range(m)]
    for r, (i, j) in enumerate(idx):
        for k in range(n):
            M[r][pos[min(k, j), max(k, j)]] += F[i][k]
            M[r][pos[min(i, k), max(i, k)]] += F[j][k]
        M[r][m] = -Fraction(sum(1 for d in drivers if IX[d] == i == j))
    g = rref(M, m)
    e = [Fraction(int(v in ("PFC", "ACC", "PPC"))) for v in NODES]
    G = [[g[pos[min(i, j), max(i, j)]] for j in range(n)] + [e[i]] for i in range(n)]
    y = rref(G, n)
    return float(sum(a * b for a, b in zip(e, y)) / 3)

if __name__ == "__main__":
    W0 = weights(); norm = max(jacobi(W0)[0])
    Ai, Al = system(weights(), norm), system(weights(lesion=0.3), norm)
    configs = [("VPL",), ("ACC",), ("PFC",), ("VPL", "ACC"), tuple(NODES)]
    print("A. Gramian of the schematic network (T = infinity);"
          " STR lesion = weights x 0.3")
    print("   drivers       trace(G) int/les  cond(G)    E intact    E lesion   ratio")
    for d in configs:
        ti, ci, ei = report(Ai, d); tl, cl, el = report(Al, d)
        xi, xl = exact_energy(Ai, d), exact_energy(Al, d)
        lab = "all 8 nodes" if len(d) == n else "+".join(d)
        fi, fl = ("*" if ei is None else " "), ("*" if el is None else " ")
        print(f"   {lab:13s} {ti:5.2f}/{tl:4.2f} {ci:12.1e} {xi:10.3g}{fi} {xl:10.3g}{fl}"
              f" {xl / xi:7.2f}")
    print("   (energies exact, rational arithmetic; * = cond(G) > 1e10, so the double-")
    print("    precision inverse of G is unreliable; float energies cannot be trusted)")
    rng = random.Random(52); bad = 0; ratio = []
    for _ in range(300):                          # weights perturbed by +-30 %
        sc = {e: rng.uniform(0.7, 1.3) for e in EDGES}
        Wr = weights(sc); nr = max(jacobi(Wr)[0])
        bad += report(system(Wr, nr), ("VPL",))[2] is None
        e1 = report(system(Wr, nr), ("VPL", "ACC"))[2]
        e2 = report(system(weights(sc, 0.3), nr), ("VPL", "ACC"))[2]
        if e1 and e2:
            ratio.append(e2 / e1)
    ratio.sort()
    print(f"   300 perturbed networks: VPL alone has cond(G) > 1e10 in {bad};"
          " VPL+ACC\n   lesion/intact"
          f" energy ratio median {ratio[len(ratio) // 2]:.2f}"
          f" (range {ratio[0]:.2g} to {ratio[-1]:.3g}, n = {len(ratio)})")
    print("B. Pinning: convergence rate lambda_min(L + 5 D_pin) of x' = -(L + 5 D_pin) x")
    for lesion in (1.0, 0.3):
        W = weights(lesion=lesion)
        L = [[(math.fsum(W[i]) if i == j else -W[i][j]) for j in range(n)]
             for i in range(n)]
        out = []
        for pins in (("VPL",), ("ACC",), ("PFC",), ("VPL", "ACC")):
            Lp = [[L[i][j] + (5.0 if i == j and NODES[i] in pins else 0.0)
                   for j in range(n)] for i in range(n)]
            out.append(f"{'+'.join(pins)} {min(jacobi(Lp)[0]):.3f}")
        print(f"   STR x {lesion}: " + "; ".join(out))
# A. Gramian of the schematic network (T = infinity); STR lesion = weights x 0.3
#    drivers       trace(G) int/les  cond(G)    E intact    E lesion   ratio
#    VPL            0.78/0.71      8.8e+13   4.33e+12*   9.03e+13*   20.83
#    ACC            1.02/0.63      2.4e+12   3.78e+11*   1.81e+13*   47.72
#    PFC            1.46/0.75      1.9e+13   1.01e+12*   8.98e+13*   89.18
#    VPL+ACC        1.80/1.34      4.4e+06   1.29e+06    1.57e+06     1.22
#    all 8 nodes    8.85/5.81      1.6e+01       1.23        1.23     1.00
#    (energies exact, rational arithmetic; * = cond(G) > 1e10, so the double-
#     precision inverse of G is unreliable; float energies cannot be trusted)
#    300 perturbed networks: VPL alone has cond(G) > 1e10 in 300; VPL+ACC
#    lesion/intact energy ratio median 3.39 (range 0.0045 to 3.46e+03, n = 300)
# B. Pinning: convergence rate lambda_min(L + 5 D_pin) of x' = -(L + 5 D_pin) x
#    STR x 1.0: VPL 0.097; ACC 0.122; PFC 0.177; VPL+ACC 0.320
#    STR x 0.3: VPL 0.091; ACC 0.096; PFC 0.146; VPL+ACC 0.274

结果与作者框架的设想很不一样。单独从 VPL、ACC 或 PFC 驱动时,最小能量有限但极大:完好时为 4.33 × 10¹²、3.78 × 10¹¹、1.01 × 10¹²,「受损」后增加 21–89 倍;但格拉姆矩阵的条件数为 10¹²–10¹⁴,在双精度浮点数下无法可靠求逆——这正是 Tu 等人指出的数值问题,所以程序用有理数精确求解李雅普诺夫方程。VPL 与 ACC 同时驱动时,能量降到约 1.29 × 10⁶,「受损」后为 1.57 × 10⁶,只增加约 22%。多加一个驱动点能降低能量,在线性模型中对任何权重都成立(格拉姆矩阵只会变大),而降低的倍数——这里约 10⁵–10⁶ 倍——完全取决于权重。把 13 条边的权重各自随机扰动 ±30%、重复 300 次,单独从 VPL 驱动的条件数全部超过 10¹⁰;两点驱动下「受损/完好」的能量比中位数为 3.39,但范围从 0.0045 到 3,460,跨越近六个数量级。在这样的模型中,「降低一个数量级」既不是稳定的数字,也不由可测的解剖决定。

相比之下,B 部分的牵制收敛率是稳定的量:牵制 VPL 时为 0.097,牵制 PFC 时为 0.177,同时牵制 VPL 与 ACC 时为 0.320;纹状体受损后分别降为 0.091、0.146、0.274。这说明在本模型中,多一个牵制点会加快收敛,而受损的影响不大——但这同样只是示意性权重下的结论。

左:示意性网络中,不同驱动节点下格拉姆矩阵条件数的常用对数(虚线为 10¹⁰,超过即视为数值病态)。右:300 个权重扰动网络中,VPL 与 ACC 同时驱动时,「基底节受损/完好」能量比的常用对数分布(虚线为 10 倍)。

作者框架用「64% 的人基底节传导受损」作为远端驱动的理由。论文表 2-2 中涉及基底节的梗死共 44/68 例,即 64.7%(第三章核对),这是梗死部位的比例;论文没有测量传导功能,也没有分析病灶部位与认知结局的关系。

假说 · 尚未验证 · 共现提升度反映流派,不反映疗效

作者的假说(尚未确立): 在针刺治疗卒中后认知障碍的随机对照试验中,扣除频率之后共现提升度高的穴对(例如躯干四肢辨证配穴)并不比提升度低的穴对带来更大的组间效应量。什么样的结果能证明它是错的: 在事先登记的元回归中(以试验为单位,控制对照组类型、样本量与偏倚风险),处方中高提升度穴对的数目与标准化效应量的关联的 95% 置信区间不包含零且方向为正。这个检验需要逐篇提取效应量,本书不做。

本章小结

论文的腧穴网络是处方共现网络。它的权重主要由各穴的使用频率决定:独立选择就能产生百会—四神聪 65 次中的约 50 次;一个没有任何配伍规则的模拟处方库,同样产生约 0.72–0.75 的聚类系数、直径 3,以及与论文相同的前五位「核心穴」。扣除频率后仍有可辨认的结构,但其形态是处方流派,而不是生理耦合。可控性与牵制控制只能放在示意性的脑区网络中讨论;在作者的模型中,单点驱动的能量有限但极大、双精度下无法可靠求逆,加驱动点必然降低能量但倍数取决于权重,「受损后能量变化」的倍数在扰动网络中跨越近六个数量级,所以「降低一个数量级」不是一个有根据的结论。64.7% 是梗死部位涉及基底节的比例,不是测得的传导受损比例。

参考文献

Gu, Shi, Fabio Pasqualetti, Matthew Cieslak, et al. 2015. “Controllability of Structural Brain Networks.” Nature Communications 6: 8414. https://doi.org/10.1038/ncomms9414.
Liu, Yang-Yu, Jean-Jacques Slotine, and Albert-László Barabási. 2011. “Controllability of Complex Networks.” Nature 473: 167–73. https://doi.org/10.1038/nature10011.
Lu, Shunkang. 2024. “Shou Shi’er Jingxue Qinzhen Ganyu Zhiliao Cuzhong Hou Qingdu Renzhi Zhang’ai (PSMCI) de Linchuang Liaoxiao Pingjia Ji Naodian Guilü Chutan [Clinical Efficacy Evaluation of Press Needles at the Twelve Jing-Well Points of the Hands for Post-Stroke Mild Cognitive Impairment (PSMCI) and a Preliminary Study of EEG Patterns].” Master’s thesis (defence version; supervisors: Guo Yi 郭义, Chen Bo 陈波, Zhang Jingsha 张静莎; in Chinese), Tianjin University of Traditional Chinese Medicine.
Pasqualetti, Fabio, Sandro Zampieri, and Francesco Bullo. 2014. “Controllability Metrics, Limitations and Algorithms for Complex Networks.” IEEE Transactions on Control of Network Systems 1 (1): 40–52.
Serrano, M. Ángeles, Marián Boguñá, and Alessandro Vespignani. 2009. “Extracting the Multiscale Backbone of Complex Weighted Networks.” Proceedings of the National Academy of Sciences of the USA 106 (16): 6483–88.
Tu, Chengyi, Rodrigo P. Rocha, Maurizio Corbetta, Sandro Zampieri, Marco Zorzi, and Samir Suweis. 2018. “Warnings and Caveats in Brain Controllability.” NeuroImage 176: 83–91.
Wang, Xiao Fan, and Guanrong Chen. 2002. “Pinning Control of Scale-Free Dynamical Networks.” Physica A 310 (3–4): 521–31. https://doi.org/10.1016/S0378-4371(02)00772-0.
Watts, Duncan J., and Steven H. Strogatz. 1998. “Collective Dynamics of ‘Small-World’ Networks.” Nature 393: 440–42. https://doi.org/10.1038/30918.
Yan, Gang, Jie Ren, Ying-Cheng Lai, Choy-Heng Lai, and Baowen Li. 2012. “Controlling Complex Networks: How Much Energy Is Needed?” Physical Review Letters 108: 218703. https://doi.org/10.1103/PhysRevLett.108.218703.

Fan, Y. (2026). Finger–Brain Cortical Magnification, Distal Microneedle Afference and Computational Neurology of Post-Stroke Cognitive Network Reconstruction: From Complex-Network Pinning Control and Thalamocortical 10 Hz Alpha Restoration to Drift-Diffusion Models, Read Beside the Twelve Jing-Well Points of the Lingshu [指—脑皮层放大、远端微针传入与卒中后认知网络重建的计算神经学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23839915 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).