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

第三章 皮层—纹状体—丘脑环路、随机共振与可塑性

病灶分布

证据

入组患者的病灶部位(论文表 2-2 与第 2.3.1 节)。 68 例中,单部位 43 例:左基底节 14、右基底节 12、「左侧胼底体」8、右半卵圆中心 6、脑干 2、左丘脑 1;多部位 25 例:包含基底节 18、包含额叶 5、包含颞叶 2。正文概括为「基底节梗死的概率最多,占 64%,其次为胼底体、半卵圆中心、额,分别占 11%、8% 和 7%」,并据此推测卒中后轻度认知障碍「可能与基底节、胼底体、半卵圆中心、额病变密切相关」(Lu 2024)。

下面的程序核对这些比例,并给出威尔逊 95% 置信区间。

推导

# Chapter 3. (A) Lesion sites of the 68 trial participants, Lu (2024) Table 2-2, with
# Wilson 95 % intervals. (B) Stochastic resonance in its simplest form: a unit with a
# FIXED threshold 1 receives either nothing or a weak input a < 1, plus Gaussian noise
# of standard deviation s. Hit rate H = P(a + noise > 1), false-alarm rate
# F = P(noise > 1); the detection index H - F and the mutual information between
# input and output (equal priors) are computed exactly. (C) The same index for a unit
# whose threshold may be placed optimally: then noise never helps.
import math, statistics

ND = statistics.NormalDist()

def wilson(k, n, z=1.959964):
    p = k / n; c = (p + z * z / (2 * n)) / (1 + z * z / n)
    h = z * math.sqrt(p * (1 - p) / n + z * z / (4 * n * n)) / (1 + z * z / n)
    return c - h, c + h

def h2(p):
    return 0.0 if p <= 0 or p >= 1 else -p * math.log2(p) - (1 - p) * math.log2(1 - p)

def detect(a, s, th=1.0):
    H = 1 - ND.cdf((th - a) / s) if s > 0 else float(a > th)
    F = 1 - ND.cdf(th / s) if s > 0 else float(0 > th)
    mi = h2((H + F) / 2) - (h2(H) + h2(F)) / 2
    return H - F, mi

if __name__ == "__main__":
    n = 68
    print("A. Lesion sites, Table 2-2 (n = 68)")
    for lab, k in (("basal ganglia: left 14 + right 12 + multi-site 18", 44),
                   ("left 'callosal body' (see text)", 8),
                   ("right centrum semiovale", 6), ("multi-site including frontal", 5),
                   ("brainstem", 2), ("multi-site including temporal", 2),
                   ("left thalamus", 1)):
        lo, hi = wilson(k, n)
        print(f"   {lab:50s} {k:2d} {k / n:6.1%}  ({lo:.1%}-{hi:.1%})")
    print("   single-site 14+12+8+6+2+1 =", 14 + 12 + 8 + 6 + 2 + 1,
          "; multi-site 18+5+2 =", 18 + 5 + 2)
    print("B. Weak input a = 0.6, fixed threshold 1: detection vs noise s")
    SS = (0.05, 0.2, 0.3, 0.4, 0.6, 1.0, 2.0)
    print("   s        " + "".join(f"{s:7.2f}" for s in SS))
    for a in (0.6, 0.8):
        row = [detect(a, s) for s in SS]
        print(f"   H-F a={a} " + "".join(f"{r[0]:7.3f}" for r in row))
        print(f"   MI  a={a} " + "".join(f"{r[1]:7.3f}" for r in row))
    grid = [0.01 * k for k in range(5, 301)]
    best = max(grid, key=lambda s: detect(0.6, s)[1])
    print(f"   information peaks at s = {best:.2f} (a = 0.6):"
          f" MI {detect(0.6, best)[1]:.3f} bits of a possible 1")
    print("C. Threshold free to move (placed at a/2 = 0.3): MI by noise")
    print("   " + "  ".join(f"s={s:g}: {detect(0.6, s, 0.3)[1]:.3f}"
                          for s in (0.05, 0.2, 0.4, 1.0)))
# A. Lesion sites, Table 2-2 (n = 68)
#    basal ganglia: left 14 + right 12 + multi-site 18  44  64.7%  (52.8%-75.0%)
#    left 'callosal body' (see text)                     8  11.8%  (6.1%-21.5%)
#    right centrum semiovale                             6   8.8%  (4.1%-17.9%)
#    multi-site including frontal                        5   7.4%  (3.2%-16.1%)
#    brainstem                                           2   2.9%  (0.8%-10.1%)
#    multi-site including temporal                       2   2.9%  (0.8%-10.1%)
#    left thalamus                                       1   1.5%  (0.3%-7.9%)
#    single-site 14+12+8+6+2+1 = 43 ; multi-site 18+5+2 = 25
# B. Weak input a = 0.6, fixed threshold 1: detection vs noise s
#    s           0.05   0.20   0.30   0.40   0.60   1.00   2.00
#    H-F a=0.6   0.000  0.023  0.091  0.152  0.205  0.186  0.112
#    MI  a=0.6   0.000  0.011  0.045  0.068  0.064  0.034  0.010
#    H-F a=0.8   0.000  0.159  0.252  0.302  0.322  0.262  0.152
#    MI  a=0.8   0.000  0.084  0.137  0.155  0.125  0.062  0.018
#    information peaks at s = 0.47 (a = 0.6): MI 0.071 bits of a possible 1
# C. Threshold free to move (placed at a/2 = 0.3): MI by noise
#    s=0.05: 1.000  s=0.2: 0.646  s=0.4: 0.228  s=1: 0.040

A 部分确认,涉及基底节的共 44 例,占 64.7%(95% 置信区间 52.8%–75.0%);单部位与多部位的分项之和分别为 43 与 25,与表中一致。正文的 11%、8%、7% 是 11.8%、8.8%、7.4% 的截尾写法。「胼底体」不是解剖学名词,疑为「胼胝体」之误;若确为胼胝体梗死,8 例(11.8%)在一般卒中人群中偏多,值得在原始影像中复核。

这些比例能说明什么,取决于比较对象。论文只纳入了卒中后轻度认知障碍患者,没有卒中后认知正常的对照组;而基底节区的腔隙性梗死在缺血性卒中中本来就很常见。没有对照,就无法判断基底节病灶在认知障碍患者中是否多于在认知正常的卒中患者中,也就无法说两者「密切相关」。此外,排除标准中有「左额叶完全损伤」(第 2.1.4 节),这会人为地降低额叶病灶的比例。

皮层—纹状体—丘脑环路

Alexander、DeLong 与 Strick 1986 年把基底节与皮层的联系概括为若干条并行、功能分隔的环路:运动环路、眼动环路、背外侧前额叶环路、外侧眶额环路与前扣带环路,每条环路都经过纹状体、苍白球或黑质与丘脑,再回到额叶 (Alexander et al. 1986)。按照这一框架,基底节区的梗死可能打断背外侧前额叶环路,表现为执行功能与处理速度下降;卒中后认知障碍的综述也把皮层下缺血性病变列为重要的病因类型之一 (Rost et al. 2022)。

作者框架由此提出「绕过纹状体的去传入」:指端的触觉信号经内侧丘系到达丘脑腹后外侧核,再到初级体感皮层,不经过纹状体;如果这一通路完好,它或许可以为受损环路所服务的皮层区域提供另一条输入。这一想法在解剖上有一定道理,但有一个明显的缺口:到达初级体感皮层不等于到达背外侧前额叶。从体感皮层到前额叶要经过多级皮层—皮层连接,而环路受损后前额叶需要的是什么样的输入,目前并不清楚。下面两节用两个常被引用的机制——随机共振与脉冲时序依赖可塑性——检验这一想法在最简单模型中的条件。两者都是作者的模型。

随机共振:噪声什么时候有用

随机共振指的是:一个有阈值的系统收到低于阈值的弱信号时,加入适量的噪声可以让信号「越过」阈值,从而提高检测;噪声太小信号过不去,噪声太大信号被淹没,所以检测表现随噪声先升后降。Collins、Imhoff 与 Grigg 1996 年在人体中证明,在皮肤上施加适量的机械噪声,可以提高对阈下触觉刺激的检测 (Collins et al. 1996)。

B 部分用最简单的形式重现这一点。阈值固定为 1,弱输入为 0.6,噪声标准差从 0.05 增加到 2:命中率减去虚报率先从 0 升到噪声 0.6 时的 0.205,再降到噪声 2 时的 0.112;输入与输出之间的互信息在噪声约 0.47 时达到峰值 0.071 比特(满分 1 比特)。C 部分给出关键的限定:如果系统可以把阈值移到 0.3,即弱输入的一半,那么不加噪声时互信息就是 1 比特,任何噪声都只会降低它。McDonnell 与 Abbott 2009 年在综述中强调了这一点:随机共振只在阈值固定且不够理想时才有好处,噪声不能胜过一个可以调整阈值的最优检测器 (McDonnell and Abbott 2009)。

所以,把指端揿针的传入看作给受损环路加入「有益噪声」,至少需要三个条件:环路中确实存在一个固定而偏高的阈值;揿针的传入以合适的强度到达这一阈值所在的位置;噪声的强度落在峰值附近,而不是两侧。这三个条件在目前的证据中都没有被测量过。

脉冲时序依赖可塑性:多给输入不等于加强连接

作者框架的另一个机制是:持续的指端传入通过可塑性「重建」认知网络。Bi 与 Poo 1998 年在培养的海马神经元中发现,突触前脉冲在突触后脉冲之前几十毫秒内到达时连接增强,在其后到达时连接减弱 (Bi and Poo 1998);Song、Miller 与 Abbott 2000 年把这一时间窗写成成对规则,并指出当减弱一侧的面积略大时,学习会自动偏向那些在时间上预测突触后放电的输入 (Song et al. 2000)。

推导

# Chapter 3. Pair-based spike-timing-dependent plasticity (STDP): does extra afferent
# input strengthen a synapse by itself? Window of Song, Miller & Abbott (2000):
# dw = +A_plus exp(-d / tau) if the presynaptic spike precedes the postsynaptic one by
# d > 0, dw = -A_minus exp(d / tau) if it follows (d < 0); tau = 20 ms,
# A_minus / A_plus = 1.05 (slightly depression-dominated). Spikes are Poisson; all
# pre-post pairs within 100 ms are summed. Scenarios (the author's model):
#   independent: pre and post fire at their own rates (input unrelated to activity)
#   causal:      a share of post spikes is triggered 4-8 ms after a pre spike
#   anti-causal: the same share fires 4-8 ms BEFORE a pre spike
import math, random

AP, AM, TAU, WIN = 0.005, 0.00525, 0.020, 0.100

def poisson(rng, rate, T):
    t, out = 0.0, []
    while True:
        t += rng.expovariate(rate)
        if t > T:
            return out
        out.append(t)

def dw(pre, post):
    total, j0 = 0.0, 0
    for tp in pre:
        while j0 < len(post) and post[j0] < tp - WIN:
            j0 += 1
        j = j0
        while j < len(post) and post[j] <= tp + WIN:
            d = post[j] - tp
            total += AP * math.exp(-d / TAU) if d > 0 else -AM * math.exp(d / TAU)
            j += 1
    return total

def scenario(rng, kind, rate_pre=10.0, rate_post=10.0, share=0.3, T=100.0):
    pre = poisson(rng, rate_pre, T)
    post = poisson(rng, rate_post * (1 - share if kind != "independent" else 1.0), T)
    if kind != "independent":
        sign = 1 if kind == "causal" else -1
        post += [tp + sign * rng.uniform(0.004, 0.008) for tp in pre
                 if rng.random() < share * rate_post / rate_pre]
    return dw(pre, sorted(post))

if __name__ == "__main__":
    rng = random.Random(1998)
    pred = 10 * 10 * (AP - AM) * TAU * 100
    print("independent firing, theory: r_pre r_post (A+ - A-) tau T ="
          f" {pred:+.4f} per 100 s")
    for kind in ("independent", "causal", "anti-causal"):
        v = [scenario(rng, kind) for _ in range(20)]
        m = math.fsum(v) / len(v)
        sd = math.sqrt(math.fsum((x - m) ** 2 for x in v) / (len(v) - 1))
        print(f"   {kind:12s} mean dw per 100 s {m:+.4f}  (sd over 20 runs {sd:.4f})")
    print("doubling the presynaptic rate (more afferent input), still independent:")
    v = [scenario(rng, "independent", rate_pre=20.0) for _ in range(20)]
    print(f"   20 Hz input  mean dw per 100 s {math.fsum(v) / 20:+.4f}")
    gain = AP * (math.exp(-0.004 / TAU) - math.exp(-0.008 / TAU)) * TAU / 0.004
    s0 = -pred / (10 * 100 * gain - pred)
    print(f"causal share that offsets the depression bias (theory: {s0:.3f}):")
    for share in (0.0, 0.005, 0.01, 0.02, 0.05):
        v = [scenario(rng, "causal", share=share) for _ in range(20)]
        print(f"   share {share:5.3f}: {math.fsum(v) / 20:+.4f}")
# independent firing, theory: r_pre r_post (A+ - A-) tau T = -0.0500 per 100 s
#    independent  mean dw per 100 s -0.0604  (sd over 20 runs 0.0628)
#    causal       mean dw per 100 s +1.0630  (sd over 20 runs 0.0962)
#    anti-causal  mean dw per 100 s -1.2146  (sd over 20 runs 0.0939)
# doubling the presynaptic rate (more afferent input), still independent:
#    20 Hz input  mean dw per 100 s -0.0764
# causal share that offsets the depression bias (theory: 0.013):
#    share 0.000: -0.0535
#    share 0.005: -0.0511
#    share 0.010: -0.0227
#    share 0.020: +0.0100
#    share 0.050: +0.1533

程序采用 Song 等人的参数(时间常数 20 毫秒,减弱幅度是增强幅度的 1.05 倍)。如果突触前(代表指端传入)与突触后(代表皮层神经元)各自以 10 赫兹独立放电,理论上每 100 秒的权重变化是 −0.05,模拟为 −0.060(20 次模拟的标准差 0.063);把突触前放电率提高一倍,变化仍为负(−0.076)。只有当一部分突触后放电由突触前脉冲在 4–8 毫秒后触发(因果配对)时,权重才明显增加(+1.063);反向配对则明显减弱(−1.215)。抵消减弱偏向所需的因果份额,理论值约为 1.3%,模拟中份额 2% 时变化转为正值(+0.010)。

这一模型的含义很直接:在脉冲时序依赖可塑性的框架下,增加传入的数量并不会加强连接,决定方向的是传入与目标神经元活动的时间关系。与任务无关、在下午到睡前某个时间按揉手指所产生的传入,没有理由与前额叶的活动形成因果配对。如果有人希望利用这一机制,更合理的设计是把感觉刺激与认知任务在时间上配对——这本身是一个需要检验的假说,而不是现成的结论。

假说 · 尚未验证 · 与任务同步的指端刺激

作者的假说(尚未确立): 对卒中后轻度认知障碍患者,在执行功能训练任务中与每次正确反应同步给予的指端触觉刺激,比相同总量但与任务不同步的指端刺激,带来更大的训练后执行功能改善。什么样的结果能证明它是错的: 在事先登记、评估者设盲的随机对照试验中,两种刺激方式在主要结局上的差别的 95% 置信区间完全落在事先写定的最小关注效应之内。

本章小结

论文中 64.7%(44/68)的患者梗死部位涉及基底节,但没有卒中后认知正常的对照组,不能说明病灶部位与认知障碍「密切相关」;「胼底体」疑为「胼胝体」之误。皮层—纹状体—丘脑并行环路为「基底节病灶影响执行功能」提供了解剖框架,而「绕过纹状体的去传入」在解剖上只到达初级体感皮层。在作者的模型中,随机共振只在阈值固定而偏高时才有好处,最佳噪声附近的信息量也很有限;脉冲时序依赖可塑性下,增加与活动无关的传入反而使连接略微减弱,只有因果配对的输入才使连接增强。两个机制都没有被测量过,都不能作为揿针疗效的证据。

参考文献

Alexander, Garrett E., Mahlon R. DeLong, and Peter L. Strick. 1986. “Parallel Organization of Functionally Segregated Circuits Linking Basal Ganglia and Cortex.” Annual Review of Neuroscience 9: 357–81. https://doi.org/10.1146/annurev.ne.09.030186.002041.
Bi, Guo-qiang, and Mu-ming Poo. 1998. “Synaptic Modifications in Cultured Hippocampal Neurons: Dependence on Spike Timing, Synaptic Strength, and Postsynaptic Cell Type.” Journal of Neuroscience 18 (24): 10464–72.
Collins, James J., Thomas T. Imhoff, and Peter Grigg. 1996. “Noise-Enhanced Tactile Sensation.” Nature 383: 770. https://doi.org/10.1038/383770a0.
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.
McDonnell, Mark D., and Derek Abbott. 2009. “What Is Stochastic Resonance? Definitions, Misconceptions, Debates, and Its Relevance to Biology.” PLoS Computational Biology 5 (5): e1000348. https://doi.org/10.1371/journal.pcbi.1000348.
Rost, Natalia S., Amy Brodtmann, Matthew P. Pase, et al. 2022. “Post-Stroke Cognitive Impairment and Dementia.” Circulation Research 130 (8): 1252–71. https://doi.org/10.1161/CIRCRESAHA.122.319951.
Song, Sen, Kenneth D. Miller, and L. F. Abbott. 2000. “Competitive Hebbian Learning through Spike-Timing-Dependent Synaptic Plasticity.” Nature Neuroscience 3 (9): 919–26. https://doi.org/10.1038/78829.

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).