Evidence level: B
机器学习能从脑电里认出”冥想状态”,但把模型扔给一个从没见过的新人,它还能认出来吗?这个问题,正是检验修炼状态是否具有”普遍生理标记”的关键。
正文
气功、禅修、瑜伽都讲”入静”——一种可以被师傅一眼看出、却很难用仪器说清的状态。过去十年,脑电(EEG)研究者尝试让机器学习替人类”看”:能不能用算法把冥想状态和普通静坐状态自动分开?答案是能,而且准确率不低。但这引出了一个更扎心的问题:模型是在”认状态”,还是在”认人”?
个体内的漂亮成绩,跨个体就掉链子
如果让模型分别对每个受试者单独训练、单独测试——这叫”个体内”分类——成绩通常很亮眼。一项 2025 年的研究对比了慈心禅(loving-kindness meditation)与非冥想状态:在个体内设定下,平均分类准确率可达 72.1%(PMID 41305083)。听起来,”入静”确实有可测的电生理特征。
但把 12 个人的数据混在一起,用其中一部分人的数据训练、拿从没参与训练的新人做测试——即”跨个体”或”受试者无关”分类——同样这批数据,平均准确率掉到 62.3%,且不同训练会话组合之间波动巨大(最好的半数组合约 78%,最差的半数仅 46%)(PMID 41305083)。这不是个例:跨个体设定下性能系统性下滑,是冥想脑电领域公认的”认人难题”——模型很容易学会识别”某个人的脑电长相”,而不是”冥想状态的普遍特征”。
破局:域对齐与迁移学习
怎么把”认人”的模型掰回”认状态”?2025 年一项来自鲁汶大学的研究给出了方法学路径(PMID 41370855)。研究者用 53 名(含新手与资深冥想者)的公开冥想脑电数据,采用留一受试者交叉验证(LOSO)——每次拿 52 人训练、1 个全新的人测试。基线模型(不做任何对齐)跨个体准确率仅 55.7%(接近随机猜测 50%)。随后引入两类无监督域适应技术:黎曼空间数据对齐(RSDA,对齐电极空间模式)与卷积 Monge 映射归一化(CMMN,对齐频谱分布),组合后跨个体准确率提升到 66.6%。换句话说,不重新训练、不做个体标定,仅靠数学对齐就能显著提升”认出陌生人”的能力。他们还发现,对分类贡献最稳定的是 θ(4-8Hz)、α(8-14Hz)、β(14-30Hz)三个频段,以及额极区与颞区的电极——这些正是冥想研究中反复出现的”经典频段”。
多模态:不止脑电,还有全身
单一脑电通道信息有限,近年研究开始叠加外周生理信号。2024 年一项研究用 64 通道脑电加生物外周监测(心率等),对 34 名资深内观禅修者解码其自评的”冥想深度”(1-5 级),在从未见过的会话上实现了显著的连续与二分类解码,最佳模型融合了空间、频谱与连接信息,而单一通道方法明显逊色(PMID 39660274)。这说明”入静深浅”这种更细腻的维度,恰恰需要多模态协同才抓得住。
对修炼科学的意义:跨个体泛化是”普遍标记”的试金石
这套方法学对易学修炼体系有直接的桥梁价值。气功”得气”、入静”深度”,长期以来靠师徒口传心授、靠主观体验,缺乏客观标尺。机器学习给出的路径是:先证明存在可分类的生理信号(已验证),再证明这种分类对陌生人成立(正在验证)。如果某天一个”入静检测模型”能在没见过的受试者上稳定泛化,那它支持的就不再是”某人特有的脑电习惯”,而是”人类共有的入静生理标记”——这正是把”功夫”从主观经验推向可重复科学的关键一步。反过来,如果模型始终只能认熟人、认不了生人,那”入静”的生理特征就更可能是高度个体化的模式,需要重新思考普遍性主张。
证据与局限
已验证: 冥想与非冥想状态可被机器学习从脑电中分类(系统综述 PMID 33618287);个体内分类准确率高(约72%,PMID 41305083);域对齐技术能显著提升跨个体泛化(LOSO 55.7%→66.6%,PMID 41370855);多模态融合可解码冥想深度(PMID 39660274)。
部分验证: 跨个体泛化虽可行,但准确率仍明显低于个体内(62% vs 72%),且跨个体方法多为单中心、样本量小(12-53人),缺乏跨文化、跨修炼传统的验证。
待验证: “入静/得气”的普遍生理标记是否真实存在且跨个体稳定;对齐与迁移学习方法能否推广到气功、站桩等非冥想类修炼;多模态(脑电+心率+呼吸)在跨个体场景的增量价值尚未量化。
相关阅读
证据等级: B
参考文献: PMID 41370855, PMID 41305083, PMID 39660274, PMID 33618287, PMID 38213699
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。
Cross-Individual Generalization in Meditation State Classification: Can the Model Recognize \”Entering Stillness\” but Not \”You\”?
Evidence level: B
Machine learning can identify a “meditation state” from EEG data, but if the model is applied to a novel individual never seen during training, can it still recognize it? This question serves as the critical test for whether cultivated states possess “universal physiological markers.”
Main Text
Qigong, Zen meditation, and yoga all describe “entering stillness” (入静)—a state discernible by a master at a glance yet difficult to define with instrumentation. Over the past decade, EEG researchers have attempted to let machine learning “see” for humans: Can algorithms automatically distinguish meditation states from ordinary quiet sitting? The answer is yes, with decent accuracy. However, this raises a more poignant question: Is the model recognizing “the state,” or is it recognizing “the person”?
Striking Individual Performance Collapses Across Individuals
If a model is trained and tested separately for each subject—termed “intra-individual” classification—the performance is typically impressive. A 2025 study compared loving-kindness meditation against non-meditation states: under an intra-individual setting, average classification accuracy reached 72.1% (PMID 41305083). This suggests that “entering stillness” indeed possesses measurable electrophysiological characteristics.
However, when data from 12 individuals were pooled, using a subset for training and testing on novel individuals never exposed to the training set—termed “cross-individual” or “subject-agnostic” classification—the same dataset yielded an average accuracy of 62.3%, with massive fluctuations across different training-session combinations (the best half of combinations reached ~78%, while the worst half dropped to only 46%) (PMID 41305083). This is not an isolated case; systematic performance degradation under cross-individual settings is a recognized “person-recognition problem” in the meditation EEG field. Models readily learn to identify “the EEG fingerprint of a specific individual” rather than “universal features of the meditation state.”
Breaking the Impasse: Domain Alignment and Transfer Learning
How can we steer a model from “recognizing people” back to “recognizing states”? A 2025 study from KU Leuven provided a methodological pathway (PMID 41370855). Researchers utilized public meditation EEG data from 53 participants (including novices and experienced meditators), employing Leave-One-Subject-Out Cross-Validation (LOSO)—training on 52 subjects and testing on a single novel individual. The baseline model (with no alignment) achieved only 55.7% cross-individual accuracy (close to random guessing at 50%). Subsequently, two unsupervised domain adaptation techniques were introduced: Riemannian Space Data Alignment (RSDA) for aligning electrode spatial patterns and Convolutional Monge Mapping Normalization (CMMN) for aligning spectral distributions. Combined, cross-individual accuracy rose to 66.6%. In other words, without retraining or individual calibration, mathematical alignment alone can significantly enhance the ability to recognize strangers. They also found that the frequency bands contributing most stably to classification were θ (4-8Hz), α (8-14Hz), and β (14-30Hz), along with frontal and temporal electrodes—precisely the “classic bands” repeatedly emerging in meditation research.
Multimodal: Beyond EEG, Including Whole-Body Signals
Single-channel EEG information is limited; recent studies begin to overlay peripheral physiological signals. A 2024 study combined 64-channel EEG with bio-peripheral monitoring (heart rate, etc.) to decode self-rated “meditation depth” (levels 1-5) in 34 experienced Vipassana practitioners on novel sessions. Significant continuous and binary decoding was achieved using a best model that fused spatial, spectral, and connectivity information, whereas single-channel methods were notably inferior (PMID 39660274). This indicates that finer dimensions such as “depth of stillness” require multimodal synergy to be captured effectively.
Implications for Cultivation Science: Cross-Individual Generalization as the Litmus Test for “Universal Markers”
This methodological framework holds direct bridging value for Yijing cultivation systems. Concepts like Qigong “receiving Qi” (得气) and the “depth of stillness” have historically relied on oral transmission between master and disciple and subjective experience, lacking objective scales. Machine learning offers a pathway: first prove the existence of classifiable physiological signals (verified), then prove this classification holds for strangers (currently under verification). If a “stillness detection model” stabilizes in generalization to unseen subjects, it supports not “person-specific EEG habits,” but “human-shared physiological markers of stillness”—a crucial step in moving “cultivation” from subjective experience toward repeatable science. Conversely, if a model consistently recognizes acquaintances but fails to recognize strangers, the physiological characteristics of “stillness” are likely highly individualized patterns, requiring a re-evaluation of claims regarding universality.
Evidence and Limitations
Verified: Meditation and non-meditation states can be classified from EEG by machine learning (systematic review PMID 33618287); intra-individual classification accuracy is high (~72%, PMID 41305083); domain alignment techniques significantly enhance cross-individual generalization (LOSO 55.7%→66.6%, PMID 41370855); multimodal fusion can decode meditation depth (PMID 39660274).
Partially Verified: Cross-individual generalization is feasible but accuracy remains significantly lower than intra-individual (62% vs 72%), and cross-individual methods are predominantly single-center with small sample sizes (12-53 participants), lacking validation across cultures and cultivation traditions.
To Be Verified: Whether universal physiological markers for “stillness/Qi reception” truly exist and remain stable across individuals; whether alignment and transfer learning methods can be extended to non-meditation practices such as Qigong and standing meditation; the incremental value of multimodal (EEG + heart rate + respiration) approaches in cross-individual scenarios remains unquantified.
Related Reading
- Qigong = Bioelectromagnetic Field
- Multimodal Physiological Features and Synchronous Correlation Matrix Validation of GDV Chakra Indicators
- Chakra = Bioelectricity
- Individual Frequency Signature as a Biometric Feature
Evidence Level: B
References: PMID 41370855, PMID 41305083, PMID 39660274, PMID 33618287, PMID 38213699
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