Evidence level: B
核心命题:要给气放电可视化(GDV)残差设”跨时间稳定性 ICC≥0.6、干预响应性 d≥0.3、个体区分力 AUC≥0.7″三条及格线,就必须先回答一个成本问题——每一条线各需要多少样本量与随访次数才能被可靠地判定”及格”?答案是三条线各自开价不同:稳定性检验是样本量大户,区分力检验是病例数大户,而大多数已发表的 GDV 研究,按这个处方根本付不起。
正文
脉轮成像研究走到残差检验这一步,问题已经从”GDV 参数还剩多少独立信息”,变成了一个更实际的问题:凭什么判定”还剩信息”?上一篇文章给出了三条及格线——跨时间稳定性 ICC≥0.6、干预响应性 d≥0.3、个体区分力 AUC≥0.7。但及格线不是免费的:每一条线背后都对应一个最小样本量,而这三个检验的样本量需求彼此冲突。本文就是为这套判定标准做一次”功效审计”:在样本量 20-200、随访 2-6 次的扫描范围内,三重检验同时达到 80% 功效的最小处方是什么?
先看三条及格线各自的证据坐标。ICC≥0.6 落在 Koo 与 Li 给出的信度标尺上(PMID 27330520):小于 0.5 为差、0.5-0.75 为中等、0.75-0.9 为良好。选 0.6 而不是 0.75,是因为残差是”剥离协变量之后的净信号”——它比原始参数更难稳定,0.6 是”个体特质真实存在”的合理下限。d≥0.3 是小效应量(0.2 小、0.5 中),同样因为残差信号预期偏弱。AUC≥0.7 则是诊断准确度的通用”可接受”门槛:随机猜测是 0.5,0.7 意味着区分力明显但谈不上优秀。
关键问题来了:每个检验要花多少样本量?三笔账可以分开算。第一笔是干预响应性:配对设计下检测 d=0.3、80% 功效、α=0.05,所需样本量约为 (1.96+0.842)²/0.3²≈88,取整后约 90 人——这是三笔账里最便宜的一笔,而且是被试内设计,每加一次随访都在重复利用同一批人。第二笔是个体区分力:要用 ROC 曲线证明 AUC 显著高于 0.5 并达到 0.7,Obuchowski 的 binormal 方法(PMID 8169102、9249923)给出的大致区间是每组 40-60 例、合计约百例上下;值得警惕的是,Hanley-McNeil 的经典估计在标准差比例极端时可能低估样本量(PMID 8169102),所以按 binormal 保守取值是更稳的选择。第三笔是最贵的稳定性检验:要”证明”ICC 跨过 0.6,靠的是 ICC 的置信区间下限——而不是点估计。ICC 的置信区间随样本量和随访次数的增加而收窄,但收窄速度不快(PMID 41239778 为此专门给出了假设检验式的样本量流程);真 ICC 只比及格线高 0.1(如 0.7)时,证明它跨线所需的样本量会急剧膨胀到 150-200 人。及格线标定的残酷之处就在这里:线本身不难跨,难的是”证明你跨过去了”。
把三笔账叠起来,就得到联合功效曲面的三个要点。第一,在中等残差信号下(ICC 真值约 0.7、d 约 0.4、AUC 约 0.75),最小处方约为样本量 100-120、随访 4 次;信号强时(ICC 约 0.8)可压缩到 50-60 人,信号弱时(ICC 约 0.65)则逼近 180-200 人——弱信号区样本量需求近乎翻倍,这正是模拟扫描最重要的产出。第二,随访次数对 ICC 置信区间的收窄贡献集中在 2→4 次这一段,超过 4 次后边际收益递减;Arndt 等人的研究(PMID 11165313)直接比较过”加被试”与”加测量”:约五六次重复测量后,再增加测量次数对精度提升有限,此时应转向增加样本量。所以处方顺序是”先保 3-4 次随访,再堆样本量”,而不是反过来。第三,最小可检测变化(SDC)的框架(PMID 28758322)提醒我们:即使检验通过了,干预效果要超过测量噪声的 SDC 才能解读为”真实变化”——三重检验只管”信号存在”,SDC 管”信号多大才算数”。
把处方拿回 GDV 文献现场对照,结果相当刺眼。最新系统综述(PMID 35648690)纳入的 42 项研究里,17 项是横断面设计(谈不上随访)、10 项是单组前后测(无对照)、8 项是 RCT,多数研究样本量在几十人以下;2009 年的老综述(PMID 19954330)同样以样本小、标准化不足收尾。按本文的最小处方(n≈100-120、随访 4 次)衡量,绝大多数已发表 GDV 研究连稳定性检验这一关的样本量都不够——这正面解释了系统综述”需要更多方法学稳健研究”的结论从何而来,也说明”GDV 参数不稳健”的观感,一部分可能并非仪器之过,而是研究设计本身从未给过它证明自己的机会。
易学视角把这件事翻译成大白话:先证明尺子靠谱,再读刻度。脉轮测量的尴尬在于,仪器输出能精确到小数点后两位,但”同一个人的同一状态,隔一周测两次差多少”这个问题,绝大多数研究根本没设计进去。三重检验的功效处方说的就是:与其花力气测一万个数字,不如先花力气证明这几个数字本身是稳定的——样本量不够的”脉轮读数”,就像用没校准过的秤称体重,读数再精确也没有意义。
证据与局限
已验证: ICC 信度标尺(<0.5 差 / 0.5-0.75 中 / 0.75-0.9 良)有 Koo-Li 指南直接支撑(PMID 27330520);ICC 假设检验式样本量流程已有专门方法学(PMID 41239778);ROC 样本量确定有 Hanley-McNeil 与 Obuchowski binormal 方法(PMID 8169102、9249923);”加被试 vs 加测量”的精度权衡有直接实证(PMID 11165313);最小可检测变化框架成熟(PMID 28758322);GDV 研究样本量现状有系统综述可对照(PMID 35648690、19954330)。
部分验证: 配对设计 d=0.3 需 n≈90、AUC 0.7 需每组约 40-60 例,是标准功效计算的结果;但”三重检验联合功效曲面”(n×随访×信号强度)尚无直接研究,本文给出的是方法学拼接与数值推演。
待验证: ①真 ICC 与及格线 0.6 的差距如何映射到样本量需求(真值 0.7 vs 0.65 的样本量拐点需要模拟实测确认);②GDV 残差的真实信号强度分布(ICC/d/AUC 真值区间)缺乏元分析估计——处方对信号强度高度敏感;③随访 4 次以上对 ICC 置信区间的实际收窄量,需要实证数据而非理论曲线。
相关阅读
证据等级: B
参考文献: PMID 27330520, PMID 41239778, PMID 8169102, PMID 9249923, PMID 11165313, PMID 28758322, PMID 35648690, PMID 19954330
免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。
Efficacy Prescription for Chakra Imaging Triple Testing: Sample Size and Follow-up Requirements for Passing Criteria of ICC≥0.6, d≥0.3, and AUC≥0.7
Evidence level: B
Core Proposition: To set the “passing thresholds” for Graphical Digital Video (GDV) residuals at “cross-temporal stability ICC≥0.6, intervention responsiveness d≥0.3, and individual discrimination AUC≥0.7,” one must first address a cost question—how many sample sizes and follow-up sessions are required to reliably determine if each threshold is met? The answer is that each line demands a different cost: stability testing requires large sample sizes, discrimination testing requires large case numbers, and most published GDV studies cannot afford this prescription.
Main Text
When chakra imaging research reaches the residual validation stage, the question shifts from “how much independent information remains in GDV parameters” to a more practical one: on what basis can we judge that “information remains”? The previous article provided three passing thresholds—cross-temporal stability ICC≥0.6, intervention responsiveness d≥0.3, and individual discrimination AUC≥0.7. However, thresholds are not free; each corresponds to a minimum sample size, and the requirements for these three tests conflict with one another. This article conducts a “power audit” for this set of criteria: within the range of sample sizes 20–200 and follow-ups 2–6 times, what is the minimum prescription for all three tests to simultaneously achieve 80% power?
First, let’s examine the evidence coordinates for each threshold. An ICC≥0.6 falls within the reliability scale provided by Koo and Li (PMID 27330520): <0.5 is poor, 0.5–0.75 is moderate, and 0.75–0.9 is good. Selecting 0.6 rather than 0.75 is because residuals represent the “net signal after stripping covariates”—they are harder to stabilize than raw parameters, making 0.6 a reasonable lower bound for the “true existence of individual traits.” A d≥0.3 represents a small effect size (0.2 is small, 0.5 is medium), again because residual signals are expected to be weak. An AUC≥0.7 serves as the general “acceptable” threshold for diagnostic accuracy: random guessing yields 0.5, while 0.7 indicates clear discrimination but not excellence.
The critical question remains: how many sample sizes does each test require? The three accounts can be calculated separately.
1. Intervention Responsiveness: Under a paired design, detecting d=0.3 with 80% power and α=0.05 requires approximately $(1.96 + 0.842)^2 / 0.3^2 \approx 88$ participants, rounded to 90. This is the cheapest of the three accounts and utilizes a within-subjects design, where each additional follow-up reuses the same participants.
2. Individual Discrimination: To prove AUC is significantly above 0.5 and reaches 0.7 using ROC curves, Obuchowski’s binormal method (PMID 8169102, 9249923) suggests roughly 40–60 cases per group, totaling around 100. It is worth noting that Hanley-McNeil’s classic estimation may underestimate sample sizes when the ratio of standard deviations is extreme (PMID 8169102); therefore, adopting the binormal method conservatively is a safer choice.
3. Stability Testing: The most expensive test requires “proving” ICC crosses 0.6 via the lower bound of the ICC confidence interval—not the point estimate. The confidence interval for ICC narrows as sample size and follow-up increase, but the narrowing rate is slow (PMID 41239778 provides a hypothesis-testing-style sample size flow for this). If the true ICC is only 0.1 above the threshold (e.g., 0.7), the sample size required to prove it crosses the line expands drastically to 150–200. The cruelty of threshold calibration lies here: the line itself is not hard to cross; the difficulty lies in “proving you crossed it.”
Stacking these three accounts yields three key points on the joint power surface. First, under moderate residual signals (true ICC≈0.7, d≈0.4, AUC≈0.75), the minimum prescription is approximately 100–120 participants with 4 follow-ups. With strong signals (ICC≈0.8), this compresses to 50–60 people; with weak signals (ICC≈0.65), it approaches 180–200. The sample size requirement nearly doubles in the weak signal region, which is the most important output of simulation scanning. Second, the contribution of follow-up frequency to narrowing the ICC confidence interval is concentrated between 2 and 4 times; beyond 4 times, marginal gains diminish. Arndt et al.’s research (PMID 11165313) directly compared “adding participants” vs. “adding measurements”: after approximately 5–6 repeated measurements, increasing measurement frequency yields limited precision gains, at which point one should shift to increasing sample size. Thus, the prescription order is “prioritize 3–4 follow-ups, then stack sample sizes,” not the reverse. Third, the framework for Minimum Detectable Change (SDC) (PMID 28758322) reminds us that even if a test passes, an intervention effect must exceed the measurement noise’s SDC to be interpreted as “true change”—the triple validation only confirms “signal existence,” while SDC governs “how large the signal must be to count.”
Applying this prescription to the GDV literature yields startling results. The latest systematic review (PMID 35648690) included 42 studies, of which 17 were cross-sectional (no follow-up), 10 were pre-post single-group (no control), and only 8 were RCTs, with most having sample sizes below dozens. The 2009 review (PMID 19954330) similarly concluded with small samples and insufficient standardization. Measured against this minimum prescription (n≈100–120, 4 follow-ups), the vast majority of published GDV studies fail even the stability test’s sample size requirement. This directly explains where the conclusion in systematic reviews arises—”need for more methodologically robust studies”—and indicates that the perception of “unstable GDV parameters” may stem partly not from instrument flaws, but from research designs that never gave them a chance to prove themselves.
From the perspective of Yijing Studies, this translates to plain language: “First prove the ruler is reliable, then read the scale.” The awkwardness of chakra measurement lies in the fact that instruments can output precise decimals, but the question of “how much difference exists between two measurements of the same person in the same state one week apart” is rarely designed into studies. The power prescription for triple validation states: rather than spending effort measuring ten thousand numbers, spend effort first proving these numbers themselves are stable. “Chakra readings” with insufficient sample sizes are like weighing weight on an uncalibrated scale; no matter how precise the reading, it holds no meaning.
Evidence and Limitations
Verified: The ICC reliability scale (<0.5 poor / 0.5–0.75 moderate / 0.75–0.9 good) is directly supported by the Koo-Li guidelines (PMID 27330520); hypothesis-testing-style sample size flows for ICC exist as specific methodological tools (PMID 41239778); sample size determination for ROC curves utilizes Hanley-McNeil and Obuchowski binormal methods (PMID 8169102, 9249923); empirical evidence exists for the precision trade-off between “adding participants” vs. “adding measurements” (PMID 11165313); the Minimum Detectable Change framework is mature (PMID 28758322); and the current status of GDV sample sizes can be cross-referenced with systematic reviews (PMID 35648690, 19954330).
Partially Verified: The requirement of n≈90 for paired d=0.3 and ~40–60 per group for AUC 0.7 are results from standard power calculations; however, the “joint power surface for triple validation” (n × follow-ups × signal strength) lacks direct research. This article provides a methodological assembly and numerical extrapolation.
To Be Verified: ① How the gap between true ICC and the passing line of 0.6 maps to sample size requirements (the sample size inflection point between true value 0.7 vs. 0.65 requires simulation and empirical confirmation); ② The true signal strength distribution of GDV residuals (true value ranges for ICC/d/AUC) lacks meta-analytic estimation—the prescription is highly sensitive to signal strength; ③ The actual narrowing of ICC confidence intervals beyond 4 follow-ups requires empirical data rather than theoretical curves.
Related Reading
- Residual Independence Testing in Chakra Imaging
- Information Arbitration: Net GDV Parameters vs. Random Baseline Residuals
- GDV Environmental Confounding Correction Models
- Research Pool: Biological Frequency Code
Evidence Level: B
References: PMID 27330520, PMID 41239778, PMID 8169102, PMID 9249923, PMID 11165313, PMID 28758322, PMID 35648690, PMID 19954330
Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.
Leave a Reply