Evidence level: B
上一轮检验”三传统是否共享可迁移签名”,这一轮换一把更细的尺子:即便三个传统都能被同一个模型认出”入定”,它们各自靠的究竟是同一组脑电特征(θ前额中线、α拓扑),还是各自拿着不同特征的”拼图”凑出了相似的分类结果——用 SHAP 贡献度排名与重叠率裁决。
正文
“万法归一”的检验走到第二步,会遇到一个更隐蔽的陷阱:分类器说”是同一类”,不等于特征说”是同一套”。 用一个模型同时识别气功、站桩、冥想三种修炼态,准确率达标,只能证明”这三个状态的脑电都能被区分开”;至于三者是否真的共享同一把钥匙,还要看模型是”靠哪几个特征”做出判断的。这就是本期要裁决的问题:三传统的”共享签名”,是同一组特征,还是不同特征拼出的相似结论?
一、为什么准确率会说谎:分类正确 ≠ 特征相同
机器学习圈近年反复踩同一个坑:模型准确率很高,但决定它判断的”关键特征”却极不稳定。2025年一篇方法学评论把这种现象命名为”稳定性悖论”:一份611人样本上,逻辑回归拿到了最高准确率(66.2%),但一旦移除排名第一的特征,特征排名立刻大幅重排;而一些无监督方法虽然准确率平平,特征排名却纹丝不动(PMID 41202777)。准确率衡量的是”猜得对不对”,特征贡献度衡量的是”凭什么猜”——两件事可以完全脱节。
这正是”共享签名”验证必须补上的一刀:上一轮用留一传统交叉验证(LOT)证明”模型能跨传统迁移”,但如果三个传统各自依赖的特征根本不同——气功靠θ、冥想靠α、站桩靠某个特定电极的功率——迁移成功只是”三个分类器恰好长得像”,而不是”三个传统共享同一神经签名”。翻译成传统语境:三个人都考上了清华,不能证明他们用的是同一套方法。
二、怎么解剖模型:SHAP 贡献度排名的原理
解剖”模型凭什么猜”的标准工具是 SHAP(SHapley Additive exPlanations)——它把模型的每个判断按博弈论夏普利值分摊给每个特征,输出”每个特征贡献了多少”的排名。2025年一项冥想研究把这个工具直接用在脑电上:26名有经验的冥想者做专注冥想与走神对照,XGBoost 分类准确率83%,随后用SHAP排出贡献度前十的特征,都是高频振荡的功率与相干性特征(PMID 39911539)。同期,一项大型多中心静息态脑电的抑郁严重度分级研究也验证了SHAP的临床可用性:模型74.5%的十折准确率之外,SHAP稳定指出左侧顶枕区是核心脑区,贡献最大的特征是左顶叶相对β功率、顶中线时域特征与全局α能量(PMID 40414904)。SHAP 的价值在于:它把”黑箱模型”翻译成一张可以跨样本、跨传统比较的”特征账单”。
三、裁决方案:贡献度重叠率怎么算
有了账单,裁决就有了依据。方案分三步:
第一步,同协议采集。 三传统各自用同一套脑电协议采集静息与修炼双态数据,保证特征字典一致——否则”排名重叠”没有可比基础。这一条与母文的LOT设计共用同一批数据(PMID 28663069、29526764)。
第二步,逐传统解释。 域对齐后,对每个传统单独训练分类器,分别用SHAP排出该传统Top-K个关键特征(K建议取10或15),得到三张排名表。
第三步,算重叠。 用两个指标量化三张表的相似度:Top-K重叠率(Jaccard系数)——三张表共有的特征数占并集的比例;Kendall W 和谐系数——三张排名在顺序上的一致程度。预设裁决线:
- 若 θ前额中线功率、α网络整合度等”经典入定特征”同时出现在三张表的前三,且 Jaccard ≥ 0.5、Kendall W ≥ 0.7 → 三传统共享同一组特征的假设获支持,”万法归一”有了第一级神经证据;
- 若三张表的前三各说各话(气功=θ、冥想=α、站桩=γ),但每张表单独都能解释本传统的分类 → “共享签名”降级为”不同特征拼出的相似结论”,共性存在于分类学层面而非特征层面,需回到更底层的网络拓扑找共性。
四、为什么这个裁决对易学研究不可绕过
跳过贡献度重叠检验,”共享签名”永远只是”准确率意义上的一厢情愿”。2021年一项对比冥想者与新手脑电谱特性的研究发现,冥想与走神的脑电差异在经验组之间系统性不同(PMID 34688899)——不做特征层面的对齐,三传统的”共享”很可能只是”都区别于走神”的共享,而不是”彼此同一”的共享。2018年NIMHANS的三档分层研究(新手/资深者/教师)更提示:共享签名的强度可能随功夫深浅分层(PMID 29526764)。把 SHAP 贡献度排名按经验分层重算,还能顺带回答”共享签名是先天共有还是功夫练出来的”——这正是下一篇(#673)的剂量-反应检验。
对传统文化而言,这个裁决给出的是一个可以接受的结局:若重叠率达标,三传统在神经层面共享同一把钥匙,”殊途同归”得证;若不达标,三传统各自成路,但”都是通往入定的有效路径”依然成立——共性退到更深的网络层面等待下一把尺子。
证据与局限
已验证: SHAP 是脑电分类标准可解释工具,能稳定指出生理可解释的关键脑区与特征(PMID 40414904);冥想态脑电分类可被 XGBoost+SHAP 解剖出前十贡献特征(PMID 39911539);高准确率不等于特征排名稳定,特征重要性存在稳定性悖论(PMID 41202777);冥想与走神脑电差异随经验水平分层(PMID 34688899)。
部分验证: 冥想者α频段网络整合度随经验增强(PMID 28663069);熟练度与练习时长可被脑电区分(PMID 29526764);SHAP 特征选择在小样本上优于 LIME 且更紧凑(来自癫痫检测研究的比较证据)。
待验证: 气功/站桩/冥想三传统间 SHAP 贡献度排名是否跨传统重叠(Jaccard/Kendall W)尚无直接实验;”共享签名=同一组特征”与”不同特征拼出相似结论”两种假说尚无裁决数据;Top-K 取值与域对齐强度对重叠率指标的影响未标定。
相关阅读
证据等级: B(SHAP 可解释机器学习与冥想脑电分类均有 PubMed 直接支撑,”跨传统贡献度重叠”本身尚无直接实证,属可检验假说)
参考文献: PMID 39911539, PMID 40414904, PMID 41202777, PMID 34688899, PMID 28663069, PMID 29526764
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。
Shared Signatures Across Three Traditions of Practice: Are They Identical EEG Patterns or Convergent Conclusions from Distinct Features?
Evidence level: B
The previous round of testing whether “the three traditions share a transferable signature” now switches to a finer ruler: even if all three traditions can be recognized by the same model as “in trance,” do they rely on the same set of EEG features (medial frontal theta, alpha topology), or do they each assemble similar classification results using different feature “puzzles”—to be adjudicated via SHAP contribution rankings and overlap rates.
Main Text
The second step of the “all methods return to one” verification encounters a more subtle trap: a classifier saying “same category” does not mean the features are “the same set.” Using a single model to identify Qi Gong, standing桩 (Zhan Zhuang), and meditation simultaneously with acceptable accuracy only proves that “EEG from these three states can be distinguished.” Whether they truly share the same key depends on which features the model relies upon for judgment. This is the question this issue adjudicates: Are the three traditions’ “shared signatures” the same set of features, or similar conclusions assembled from different features?
I. Why Accuracy Lies: Classification Correctness ≠ Feature Identity
The machine learning community has repeatedly stumbled into the same pitfall recently: models achieve high accuracy, yet the “key features” driving their judgments are highly unstable. A 2025 methodological review named this phenomenon the “Stability Paradox”: on a sample of 611 individuals, logistic regression achieved the highest accuracy (66.2%), yet removing the top-ranked feature immediately caused a significant reshuffling of the ranking; conversely, some unsupervised methods had mediocre accuracy but stable feature rankings (PMID 41202777). Accuracy measures “whether the guess is correct,” while feature contribution measures “why the guess was made”—these two can be completely decoupled.
This is exactly what must be added to the “shared signature” validation: the previous round used Leave-One-Traditional-Out (LOTO) cross-validation to prove “model transferability across traditions.” However, if the three traditions rely fundamentally on different features—Qi Gong on theta, meditation on alpha, and Zhan Zhuang on power from a specific electrode—successful transfer merely means “three classifiers happen to look alike,” not that “the three traditions share the same neural signature.” Translated into traditional context: Just because three people all gained admission to Tsinghua University does not prove they used the same methodology.
II. How to Dissect the Model: Principles of SHAP Contribution Rankings
The standard tool for dissecting “why the model guesses” is SHAP (SHapley Additive exPlanations)—it distributes each model’s judgment according to Shapley values from game theory to every feature, outputting a ranking of “how much each feature contributed.” A 2025 meditation study applied this tool directly to EEG: with 26 experienced meditators performing focused meditation versus mind-wandering comparisons, an XGBoost classifier achieved 83% accuracy; subsequent SHAP analysis ranked the top ten contributing features, all of which were power and coherence features in high-frequency oscillations (PMID 39911539). Concurrently, a large multi-center resting-state EEG study on depression severity grading validated SHAP’s clinical utility: beyond the model’s 74.5% ten-fold accuracy, SHAP consistently identified the left parietal-occipital region as the core brain area, with the largest contributions coming from relative beta power in the left superior parietal lobule, temporal domain features in the midline, and global alpha energy (PMID 40414904). The value of SHAP lies in translating the “black-box model” into a comparable “feature bill” across samples and traditions.
III. Adjudication Scheme: Calculating Contribution Overlap Rates
With the bill in hand, adjudication now has a basis. The scheme proceeds in three steps:
Step 1: Same-Protocol Acquisition. Each tradition collects resting and cultivation dual-state data using the same EEG protocol to ensure consistent feature dictionaries—otherwise, “ranking overlap” lacks a comparative basis. This aligns with the mother text’s LOTO design using the same batch of data (PMID 28663069, 29526764).
Step 2: Per-Tradition Explanation. After domain alignment, train a classifier separately for each tradition and use SHAP to rank the top-K key features for that tradition (suggesting K=10 or 15), resulting in three ranking tables.
Step 3: Calculate Overlap. Quantify the similarity of the three tables using two metrics: Top-K Overlap Rate (Jaccard Coefficient)—the proportion of shared features between the three tables relative to their union; and Kendall W Concordance Coefficient—the degree of consistency in the ordering of the three rankings. The preset adjudication thresholds are:
- If “classic trance features” such as medial frontal theta power and alpha network integration appear in the top three of all three tables, AND Jaccard ≥ 0.5 and Kendall W ≥ 0.7 → The hypothesis that the three traditions share the same set of features is supported, providing first-level neural evidence for “all methods return to one”;
- If the top three of each table differ (Qi Gong=theta, meditation=alpha, Zhan Zhuang=gamma), yet each table individually explains classification for its tradition → The “shared signature” downgrades to “similar conclusions assembled from different features,” with commonality existing at the taxonomic level rather than the feature level, requiring return to deeper network topology to find commonalities.
IV. Why This Adjudication Is Irresignable for Yijing Studies
Skipping the contribution overlap test means “shared signature” remains merely a “one-sided wish in terms of accuracy.” A 2021 study comparing EEG spectral characteristics between meditators and novices found that EEG differences between meditation and mind-wandering were systematically different across experience groups (PMID 34688899)—without alignment at the feature level, the “shared” aspect of the three traditions is likely just “all distinct from mind-wandering,” not “mutually identical.” The 2018 NIMHANS three-tier stratified study (novices/master practitioners/instructors) further suggests that the intensity of shared signatures may stratify with practice depth (PMID 29526764). Recalculating SHAP contribution rankings stratified by experience levels can simultaneously answer whether the “shared signature” is innate or cultivated—a point addressed in the next issue (#673) regarding dose-response testing.
For traditional culture, this adjudication offers an acceptable outcome: If the overlap rate meets standards, the three traditions share the same key at the neural level, validating “different paths, same destination”; if not, each tradition forms its own path, but the assertion that “all are effective paths to trance” remains valid—commonality retreats to a deeper network level awaiting the next ruler.
Evidence and Limitations
Verified: SHAP is a standard interpretable tool for EEG classification, consistently identifying physiologically interpretable key brain regions and features (PMID 40414904); meditation state EEG classification can be dissected via XGBoost+SHAP to reveal top ten contributing features (PMID 39911539); high accuracy does not equal stable feature rankings, with a stability paradox in feature importance (PMID 41202777); EEG differences between meditation and mind-wandering stratify by experience level (PMID 34688899).
Partially Verified: Alpha-band network integration increases with experience in meditators (PMID 28663069); proficiency and practice duration can be distinguished by EEG (PMID 29526764); SHAP feature selection outperforms LIME on small samples and is more compact (based on comparative evidence from epilepsy detection studies).
To Be Verified: Whether SHAP contribution rankings overlap across Qi Gong/Zhan Zhuang/Meditation traditions lacks direct experimental data (Jaccard/Kendall W); no adjudication data exists for the two hypotheses (“shared signature = same set of features” vs. “different features assembled into similar conclusions”); the impact of Top-K selection and domain alignment intensity on overlap metrics remains uncalibrated.
Related Reading
- Qi Gong, Zhan Zhuang, Meditation: Do Three Cultivation Traditions Hide the Same Key in Their EEG?
- Cross-Tradition Cultivation State EEG Recognition: How Domain Alignment Lets “Zhang San’s Model” Recognize “Li Si’s Trance”
- Quantitative Correlation Between Practitioner Neural Oscillation Features and Qi Gong Practice Level
- Research Pool 2: Resonant Consciousness Field
Evidence Level: B (SHAP interpretable machine learning and meditation EEG classification both have direct PubMed support; “cross-tradition contribution overlap” itself lacks direct empirical evidence, belonging to a testable hypothesis)
References: PMID 39911539, PMID 40414904, PMID 41202777, PMID 34688899, PMID 28663069, PMID 29526764
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