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

第四章 α 振荡:Welch 谱估计、Jansen—Rit 模型与六例脑电

论文的脑电研究

证据

研究内容三(据论文第 3 章)。 2023 年 9 月至 12 月公开招募卒中后轻度认知障碍患者 6 例,采用自身前后对照设计,没有对照组。设备为 Neuroscan 脑电系统,64 导 Quik-Cap 电极帽(10-10 排列),采样率 1,000 Hz。流程:闭眼静息 5 分钟、睁眼静息 5 分钟;依次用揿针刺激双侧少商、商阳、中冲、关冲、少冲、少泽各 60 秒;再记录闭眼 5 分钟、睁眼 5 分钟。预处理包括平均参考、按「刺激点前 0.5 秒至后 1 秒」分段、独立成分分析去伪迹与基线校正;主要观察指标是 Welch 法功率谱(第 3.2.4–3.2.6 节)。结果部分给出一位受试者的频谱图,文字描述为:刺激前 α 波「呈现下降状态」,刺激后「α 波能量在睁眼状态下有明显提升」,前额 α「在 10 Hz 左右达到峰值」,其他波段能量下降;没有报告组水平的数值或统计检验(第 3.3 节)(Lu 2024)。

论文在不足中也写明「脑电研究样本量有待扩充」(第 4.4 节)。文本本身还有几个问题:数据采集一节说受试者「保持闭眼状态」,而流程中有两段睁眼记录;「刺激点前 0.5 秒至后 1 秒」的分段与基线校正是事件相关电位的做法,用于 5 分钟静息记录时「刺激点」指什么没有说明;「刺激前 α 下降、刺激后睁眼 α 提升」若比较的是刺激前闭眼与刺激后睁眼,就混入了最强的已知效应——睁眼使 α 大幅下降的伯杰效应;这 6 位患者是否也参加了临床试验,也没有说明。

丘脑—皮层的 α 节律

Berger 1929 年首次记录人的脑电时就描述了约 10 赫兹的节律在闭眼时出现、睁眼时减弱 (Berger 1929)。今天一般认为,这一节律由丘脑—皮层回路与皮层内回路共同产生,Lopes da Silva 等人 1974 年用丘脑神经元群体的模型重现了 α 节律 (Lopes da Silva et al. 1974)。α 功率与峰频率与认知表现有关,但关系复杂:静息时较高的 α 功率与较好的认知表现相关,而任务中的 α 下降(去同步)又是皮层投入工作的标志 (Klimesch 1999)。在缺血性卒中中,定量脑电的「变慢」(δ、θ 增加,α 减少)与预后有关,δ/α 比值等指标在急性与亚急性期具有预后价值 (Finnigan and Putten 2013)。Barry 等人 2007 年系统比较了闭眼与睁眼静息,发现睁眼时 α 功率在全头皮显著下降,并伴有其他频段的变化 (Barry et al. 2007)。

所以,「卒中后 α 减少、恢复时 α 回升」是有根据的整体趋势;但在一个 20 分钟的记录中比较 α,必须区分睁眼与闭眼、考虑警觉水平的变化,并有一个不接受刺激的对照。

精确的 Welch 谱估计

论文第 3.2.6.1 节写道:Welch 法把 N 个采样分成 k 段,段间可以重叠(50% 或 0%),加窗后计算周期图并平均,「Welch 法得到频谱估计的方差是周期图频谱估计方差的 1/k」。Welch 1967 年的原文正是讨论了重叠分段的方差 (Welch 1967)。对于白噪声,第 jj 个与第 j+mj+m 个分段周期图在同一频点(远离零频与奈奎斯特频率)的相关系数近似为 ρ(m)2\rho(m)^2,其中

ρ(m)=∑nw[n]w[n+mD]∑nw[n]2,\rho(m) = \frac{\sum_n w[n]\, w[n + mD]}{\sum_n w[n]^2},

DD 是相邻分段的间隔。于是

Var⁡ŜS2=1K[1+2∑m=1K−1(1−mK)ρ(m)2].\frac{\operatorname{Var}\hat{S}}{S^2} = \frac{1}{K}\Big[1 + 2\sum_{m=1}^{K-1}\Big(1 – \frac{m}{K}\Big)\rho(m)^2\Big].

这是谱分析教科书中的标准结果 (Percival and Walden 1993)。只有分段互不重叠、ρ=0\rho = 0 时,方差才恰好是 1/K1/K。

推导

# Chapter 4. Variance of the Welch estimate. Lu (2024, 3.2.6.1) says it is 1/k of the
# periodogram's; that holds for NON-overlapping segments. For white noise with overlap,
#   Var / E^2 = (1/K) [1 + 2 sum_{j=1}^{K-1} (1 - j/K) rho(j)^2],
#   rho(j) = sum_n w[n] w[n + jD] / sum_n w[n]^2   (Welch 1967; Percival & Walden 1993),
# checked by simulation (N = 4,096, segments of 256, bin 20, 500 replicates each).
import math, random

M, N, BIN, REPS = 256, 4096, 20, 500
W = [0.5 * (1 - math.cos(2 * math.pi * n / M)) for n in range(M)]       # Hann
U = math.fsum(w * w for w in W)
C = [W[n] * math.cos(2 * math.pi * BIN * n / M) for n in range(M)]
S = [W[n] * math.sin(2 * math.pi * BIN * n / M) for n in range(M)]

def theory(D):
    K = (N - M) // D + 1
    rho = [math.fsum(W[n] * W[n + j * D] for n in range(M - j * D)) / U
           for j in range(1, K) if j * D < M]
    return K, (1 + 2 * math.fsum((1 - j / K) * r * r for j, r in enumerate(rho, 1))) / K

def welch_once(x, D):
    vals = []
    for s0 in range(0, N - M + 1, D):
        re = math.fsum(C[n] * x[s0 + n] for n in range(M))
        im = math.fsum(S[n] * x[s0 + n] for n in range(M))
        vals.append((re * re + im * im) / U)
    return math.fsum(vals) / len(vals)

def simulate(D, seed):
    rng = random.Random(seed); est = []
    for _ in range(REPS):
        x = [rng.gauss(0.0, 1.0) for _ in range(N)]
        est.append(welch_once(x, D))
    m = math.fsum(est) / REPS
    return math.fsum((e - m) ** 2 for e in est) / (REPS - 1) / (m * m)

if __name__ == "__main__":
    print(f"(simulation s.e. about {math.sqrt(2 / (REPS - 1)):.0%} of the value)")
    print("overlap   K    1/K      theory   simulated   theory x K")
    for lab, D in (("0 %", M), ("50 %", M // 2), ("75 %", M // 4)):
        K, th = theory(D)
        print(f"{lab:6s} {K:4d} {1 / K:8.4f} {th:9.4f} {simulate(D, 1967 + D):10.4f}"
              f" {th * K:10.3f}")
    rho1 = math.fsum(W[n] * W[n + M // 2] for n in range(M // 2)) / U
    print(f"Hann, 50 % overlap: rho(1) = {rho1:.4f} (exactly 1/6),"
          f" rho(1)^2 = {rho1 ** 2:.4f}")
    K = (300_000 - 2000) // 1000 + 1
    K2, th2 = K, (1 + 2 * (1 - 1 / K) * rho1 ** 2) / K
    print("EEG case, 5 min at 1,000 Hz, 2-s Hann windows, 50 % overlap:"
          f" K = {K2},"
          f"\n   relative s.d. of each spectral value sqrt({th2:.5f}) ="
          f" {math.sqrt(th2):.3f}")
# (simulation s.e. about 6% of the value)
# overlap   K    1/K      theory   simulated   theory x K
# 0 %      16   0.0625    0.0625     0.0552      1.000
# 50 %     31   0.0323    0.0340     0.0345      1.054
# 75 %     61   0.0164    0.0313     0.0319      1.909
# Hann, 50 % overlap: rho(1) = 0.1667 (exactly 1/6), rho(1)^2 = 0.0278
# EEG case, 5 min at 1,000 Hz, 2-s Hann windows, 50 % overlap: K = 299,
#    relative s.d. of each spectral value sqrt(0.00353) = 0.059

对于 Hann 窗,50% 重叠时 ρ(1)\rho(1) 恰为 1/6,ρ(1)2=0.0278\rho(1)^2 = 0.0278,方差是 1/K1/K 的 1.054 倍;75% 重叠时,相邻三个分段都相关,方差是 1/K1/K 的 1.909 倍。模拟(每种 500 次,标准误约为数值的 6%)与公式一致:50% 重叠为 0.0345 对 0.0340,75% 重叠为 0.0319 对 0.0313;不重叠时模拟值 0.0552 比理论值 0.0625 低约两个标准误,属于抽样波动。需要注意的是,在总长度固定时,50% 重叠仍然比不重叠好:方差从 0.0625 降到 0.0340,只是没有降到 1/K=0.03231/K = 0.0323。所以论文的说法应改为:「不重叠时方差为周期图的 1/k;50% 重叠加 Hann 窗时,方差约为 1.05/K,仍低于不重叠的情形。」

左:总长度固定时,Welch 估计的相对方差随分段重叠比例的变化(实线为 Hann 窗的精确值,虚线为 1/K)。右:方差乘以分段数 K;等于 1 时「1/K」才成立。

程序最后一行把这一点放回论文的数据:5 分钟、1,000 赫兹的记录,用 2 秒 Hann 窗、50% 重叠,可得 299 段,每个频点的相对标准差约 5.9%。也就是说,单个记录内部的谱估计噪声并不大;论文脑电部分的问题不在谱估计的方差,而在只有 6 个人、没有对照、没有统计检验。

Jansen—Rit 神经质量模型

作者框架希望用一个模型说明「α 峰在卒中后塌陷、经刺激后恢复」。最常用的工具是 Jansen 与 Rit 1995 年的神经质量模型:一个皮层柱由锥体细胞群、兴奋性中间神经元群与抑制性中间神经元群组成,三者通过 S 形发放率函数与二阶突触核耦合,外部输入是随机的脉冲密度 (Jansen and Rit 1995)。在原文的标准参数下,模型的输出 y1−y2y_1 – y_2 呈现约 10 赫兹的节律。下面用四阶龙格—库塔法求解,并改变总连接强度 CC、兴奋性突触增益 AA、外部输入的均值与抑制性时间常数,看 α 峰怎样变化。这是作者的模型情景,参数变化不对应任何测得的病理。

推导

# Chapter 4. Jansen-Rit (1995) neural mass, RK4 with dt = 1 ms (input fixed per step):
#   y0'' = A a S(y1 - y2) - 2a y0' - a^2 y0
#   y1'' = A a [p + C2 S(C1 y0)] - 2a y1' - a^2 y1
#   y2'' = B b C4 S(C3 y0) - 2b y2' - b^2 y2
#   S(v) = 2 e0 / (1 + exp(r (v0 - v))). Standard: A 3.25, B 22, a 100, b 50, C 135
# (C1..C4 = C, 0.8C, 0.25C, 0.25C), e0 2.5, v0 6, r 0.56, p ~ U[120, 320]. Output y1 - y2;
# Welch spectra (2-s Hann, 50 % overlap) of 10 s after a 2-s transient.
import math, random

STD = dict(A=3.25, B=22.0, a=100.0, b=50.0, C=135.0, pmin=120.0, pmax=320.0)

def S(v):
    return 5.0 / (1 + math.exp(0.56 * (6.0 - v)))

def simulate(par, seed, T=12.0, dt=0.001):
    A, B, a, b, C = par["A"], par["B"], par["a"], par["b"], par["C"]
    C1, C2, C3, C4 = C, 0.8 * C, 0.25 * C, 0.25 * C
    rng = random.Random(seed); y = [0.0] * 6; out = []

    def f(y, p):
        return [y[3], y[4], y[5],
                A * a * S(y[1] - y[2]) - 2 * a * y[3] - a * a * y[0],
                A * a * (p + C2 * S(C1 * y[0])) - 2 * a * y[4] - a * a * y[1],
                B * b * C4 * S(C3 * y[0]) - 2 * b * y[5] - b * b * y[2]]
    for k in range(int(T / dt)):
        p = rng.uniform(par["pmin"], par["pmax"])
        k1 = f(y, p)
        k2 = f([y[i] + dt / 2 * k1[i] for i in range(6)], p)
        k3 = f([y[i] + dt / 2 * k2[i] for i in range(6)], p)
        k4 = f([y[i] + dt * k3[i] for i in range(6)], p)
        y = [y[i] + dt / 6 * (k1[i] + 2 * k2[i] + 2 * k3[i] + k4[i]) for i in range(6)]
        if k * dt >= 2.0:
            out.append(y[1] - y[2])
    return out

def spectrum(x, fs=1000, seg=2000, freqs=None):
    freqs = freqs or [0.5 * k for k in range(2, 61)]
    w = [0.5 * (1 - math.cos(2 * math.pi * n / seg)) for n in range(seg)]
    tabs = [([w[n] * math.cos(2 * math.pi * f * n / fs) for n in range(seg)],
             [w[n] * math.sin(2 * math.pi * f * n / fs) for n in range(seg)])
            for f in freqs]
    P = [0.0] * len(freqs); K = 0
    for s0 in range(0, len(x) - seg + 1, seg // 2):
        xs = x[s0:s0 + seg]; m = math.fsum(xs) / seg; xs = [v - m for v in xs]; K += 1
        for i, (c, s) in enumerate(tabs):
            re = math.fsum(c[n] * xs[n] for n in range(seg))
            im = math.fsum(s[n] * xs[n] for n in range(seg))
            P[i] += re * re + im * im
    return freqs, [v / K for v in P]

def describe(par, seed=1995):
    x = simulate(par, seed)
    f, P = spectrum(x)
    tot = math.fsum(P); alpha = math.fsum(p for fr, p in zip(f, P) if 8 <= fr <= 13)
    k = max(range(len(P)), key=lambda i: P[i])
    m = math.fsum(x) / len(x); sd = math.sqrt(math.fsum((v - m) ** 2 for v in x) / len(x))
    return f[k], alpha / tot, sd, f, P

CASES = [("standard (C = 135)", {}), ("C x 0.8 = 108", {"C": 108.0}),
         ("C x 0.7 = 94.5", {"C": 94.5}), ("C x 0.5 = 67.5", {"C": 67.5}),
         ("C x 0.5, input mean + 60", {"C": 67.5, "pmin": 180.0, "pmax": 380.0}),
         ("C x 0.5, A = 4.5 mV (+38 %)", {"C": 67.5, "A": 4.5}),
         ("C x 2 = 270", {"C": 270.0}), ("b = 40/s (slower inhibition)", {"b": 40.0})]

if __name__ == "__main__":
    print("case                              peak Hz   8-13 Hz share"
          "   s.d. of y1-y2 (mV)")
    for lab, ch in CASES:
        par = dict(STD); par.update(ch)
        pk, sh, sd, _, _ = describe(par)
        print(f"{lab:33s} {pk:7.1f} {sh:15.2f} {sd:16.2f}")
# case                              peak Hz   8-13 Hz share   s.d. of y1-y2 (mV)
# standard (C = 135)                   10.5            0.97             1.23
# C x 0.8 = 108                         8.5            0.64             0.36
# C x 0.7 = 94.5                        8.0            0.45             0.31
# C x 0.5 = 67.5                        1.0            0.28             0.31
# C x 0.5, input mean + 60              1.0            0.29             0.31
# C x 0.5, A = 4.5 mV (+38 %)           1.0            0.30             0.42
# C x 2 = 270                           5.0            0.18            11.83
# b = 40/s (slower inhibition)          4.5            0.10             4.80

以下是单次 10 秒的实现,谱峰位置与 α 份额有抽样波动(独立核查用多个种子平均,定性结论相同)。标准参数下,谱峰在 10.5 赫兹,8–13 赫兹的功率占 1–30 赫兹总功率的 97%。把 CC 降到 0.8 倍时谱峰移到 8.5 赫兹,α 份额降到 64%;降到 0.7 倍时为 8.0 赫兹、45%;降到一半时 α 峰消失,功率集中在最低频率(份额 28%)。在 CC 减半的状态下,提高外部输入的均值(+60 脉冲/秒)或把兴奋性突触增益提高 38%,α 都没有恢复;只有把 CC 本身恢复到 0.8 倍以上,α 才回来。CC 加倍时模型进入大幅的棘波样振荡(标准差 11.83 毫伏,谱峰 5 赫兹);把抑制性速率常数从 50 降到 40 每秒,谱峰移到 4.5 赫兹。

Jansen—Rit 模型在四种参数下的功率谱(灰色带为 8–13 赫兹):标准参数、连接强度 0.8 倍、0.5 倍与 2 倍。

这些结果有两点值得记下。第一,在这个模型中,「α 塌陷」与「α 恢复」可以用一个参数的下降与回升来描述,而增加外部输入——最接近「多给一些感觉刺激」的操作——并不能使 α 恢复。第二,同一个谱峰位置可以由不同的参数组合产生(例如 8.5 赫兹附近既可以来自 CC 下降,也可以来自其他参数的变化),所以从头皮脑电的谱形状反推模型参数是一个病态的问题。第 8 本与第 44 本讨论过耦合振子的同步,这里不重复。模型中的「恢复」只说明模型能够写出这样的轨迹,不说明揿针能够产生它。

六个人能说明什么

最后回到论文的设计本身。下面的程序生成没有任何治疗效应的合成数据:6 个人,每人有自己的 α 水平;闭眼比睁眼高约 2.8 倍(伯杰效应);刺激前后两段记录之间有随机变化,并可能有一个所有人共同的时间漂移(例如警觉水平随时间变化)。

推导

# Chapter 4. Six participants, before and after, eyes open and eyes closed: what can
# such a design show? Synthetic log10 alpha power, NO treatment effect anywhere:
#   subject level u_i ~ N(0, 0.25^2); eyes-closed minus eyes-open (Berger effect)
#   b_i ~ N(0.45, 0.15^2), i.e. about 2.8 times; session-to-session change in each
#   state d ~ N(drift, 0.12^2), where drift is a common time effect (0 or +0.06).
# Questions: how often do 5 or 6 of 6 subjects show "more eyes-open alpha after"? How
# large is the Berger contrast compared with such changes? What effect size would a
# paired t test with n = 6 detect with 80 % power? All numbers are illustrative.
import math, random

T975_5 = 2.5706                                  # t quantile, 5 df, two-sided 0.05

def study(rng, drift):
    ups, ratio = 0, []
    for _ in range(6):
        u = rng.gauss(0, 0.25); b = rng.gauss(0.45, 0.15)
        eo_pre = u + rng.gauss(0, 0.05); ec_pre = eo_pre + b
        eo_post = eo_pre + rng.gauss(drift, 0.12)
        ups += eo_post > eo_pre; ratio.append(10 ** (ec_pre - eo_pre))
    return ups, sorted(ratio)[3]

def power(dz, rng, reps=4000):
    hit = 0
    for _ in range(reps):
        d = [rng.gauss(dz, 1.0) for _ in range(6)]
        m = math.fsum(d) / 6; s = math.sqrt(math.fsum((x - m) ** 2 for x in d) / 5)
        hit += abs(m / (s / math.sqrt(6))) > T975_5
    return hit / reps

if __name__ == "__main__":
    rng = random.Random(2023)
    for drift in (0.0, 0.06):
        res = [study(rng, drift) for _ in range(20000)]
        p6 = sum(u == 6 for u, _ in res) / len(res)
        p5 = sum(u >= 5 for u, _ in res) / len(res)
        print(f"common time drift {drift:+.2f} log10 units ({10 ** drift - 1:+.0%}):"
              f" all 6 up {p6:.3f}; at least 5 up {p5:.3f}")
    print(f"   theory without drift: 1/64 = {1 / 64:.3f}; 7/64 = {7 / 64:.3f}")
    rat = [study(rng, 0.0)[1] for _ in range(2000)]
    print("Berger contrast, eyes closed / eyes open alpha power, median of study"
          f" medians: {sorted(rat)[1000]:.2f}")
    print("sign test, n = 6: smallest attainable two-sided p =", f"{2 / 64:.4f}")
    print("paired t test, n = 6, two-sided 0.05: power by standardised effect dz")
    print("   " + "  ".join(f"dz={dz:g}: {power(dz, rng):.2f}"
                          for dz in (0.5, 1.0, 1.5, 2.0)))
# common time drift +0.00 log10 units (+0%): all 6 up 0.017; at least 5 up 0.110
# common time drift +0.06 log10 units (+15%): all 6 up 0.109; at least 5 up 0.402
#    theory without drift: 1/64 = 0.016; 7/64 = 0.109
# Berger contrast, eyes closed / eyes open alpha power, median of study medians: 3.02
# sign test, n = 6: smallest attainable two-sided p = 0.0312
# paired t test, n = 6, two-sided 0.05: power by standardised effect dz
#    dz=0.5: 0.17  dz=1: 0.50  dz=1.5: 0.83  dz=2: 0.97

没有共同漂移时,6 人全部「刺激后睁眼 α 升高」的概率是 0.017(理论值 1/64),至少 5 人升高的概率是 0.110(理论值 7/64)。只要有一个所有人共同的、约 15% 的时间漂移,这两个概率就升到 0.109 与 0.402。与此相比,在本模拟的设定下(与 Barry 等人报告的量级相近 (Barry et al. 2007)),闭眼与睁眼之间的 α 功率比约为 3.02 倍,比任何合理的刺激效应都大得多,所以任何跨越睁眼与闭眼的比较都会被它淹没。即使做统计检验,6 人的符号检验最小可能的双侧 p 值是 0.031;配对 t 检验要有 80% 的把握度,标准化效应 dzd_z 需要约 1.5(程序中 dz=1.5d_z = 1.5 时把握度为 0.83)。论文只展示了一位受试者的频谱,没有组水平的数字,这些检验都无法进行。

假说 · 尚未验证 · 刺激时的 α 变化是唤醒的变化

作者的假说(尚未确立): 按揉揿针期间及其后数分钟内前额 α 功率的变化,主要反映唤醒与警觉水平的变化;控制同时记录的唤醒指标(皮肤电导、心率)后,真揿针与假揿针之间的 α 差别不再显著。什么样的结果能证明它是错的: 在事先登记的交叉设计中,控制唤醒指标后,真揿针相对假揿针的睁眼前额 α 功率差别的 95% 置信区间仍然不包含零,且在独立样本中重现。

本章小结

论文的脑电部分是 6 例、无对照的自身前后比较,只展示了一位受试者的频谱与定性描述,文中还有闭眼与睁眼记录的矛盾、分段方法不明等问题。「Welch 估计的方差是周期图的 1/k」只对不重叠的分段成立;50% 重叠加 Hann 窗时为 1/K 的 1.054 倍,75% 重叠时为 1.909 倍;但单个 5 分钟记录的谱估计已经足够稳定,问题在于样本与设计。Jansen—Rit 模型可以写出「连接强度下降、α 塌陷、连接恢复、α 回来」的轨迹,但增加外部输入并不能恢复 α,而且谱形状不能唯一确定参数。没有治疗效应时,6 人中至少 5 人「α 升高」的概率在有共同时间漂移时可达约 40%,在本模拟设定下伯杰效应约为 3 倍。「刺激后 α 恢复」在论文中没有得到检验。

参考文献

Barry, Robert J., Adam R. Clarke, Stuart J. Johnstone, Christopher A. Magee, and Jacqueline A. Rushby. 2007. “EEG Differences between Eyes-Closed and Eyes-Open Resting Conditions.” Clinical Neurophysiology 118 (12): 2765–73. https://doi.org/10.1016/j.clinph.2007.07.028.
Berger, Hans. 1929. “Über Das Elektrenkephalogramm Des Menschen.” Archiv für Psychiatrie Und Nervenkrankheiten 87: 527–70.
Finnigan, Simon, and Michel J. A. M. van Putten. 2013. “EEG in Ischaemic Stroke: Quantitative EEG Can Uniquely Inform (Sub-)acute Prognoses and Clinical Management.” Clinical Neurophysiology 124 (1): 10–19. https://doi.org/10.1016/j.clinph.2012.07.003.
Jansen, Ben H., and Vincent G. Rit. 1995. “Electroencephalogram and Visual Evoked Potential Generation in a Mathematical Model of Coupled Cortical Columns.” Biological Cybernetics 73: 357–66. https://doi.org/10.1007/BF00199471.
Klimesch, Wolfgang. 1999. “EEG Alpha and Theta Oscillations Reflect Cognitive and Memory Performance: A Review and Analysis.” Brain Research Reviews 29 (2–3): 169–95.
Lopes da Silva, F. H., A. Hoeks, H. Smits, and L. H. Zetterberg. 1974. “Model of Brain Rhythmic Activity: The Alpha-Rhythm of the Thalamus.” Kybernetik 15: 27–37.
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.
Percival, Donald B., and Andrew T. Walden. 1993. Spectral Analysis for Physical Applications: Multitaper and Conventional Univariate Techniques. Cambridge University Press.
Welch, Peter D. 1967. “The Use of Fast Fourier Transform for the Estimation of Power Spectra: A Method Based on Time Averaging over Short, Modified Periodograms.” IEEE Transactions on Audio and Electroacoustics 15 (2): 70–73. https://doi.org/10.1109/TAU.1967.1161901.

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).