六点位置特异模型换台设备还灵吗:跨中心/跨仪器泛化稳定性的检验设计

Written by

in

Language

Evidence level: C

一个在单中心、单设备上训练出的藏医”六点-脏腑”判别模型,换设备、换中心之后判别力还剩多少——这是它从”某实验室的发现”升级为”可复用诊断工具”的唯一门槛。

藏医脉诊把腕部触诊划分为多个固定点位,传统上不同点位承载不同脏腑的信息。上一轮我们已经提出单点判别力的跨脏器异质性检验(见《藏医六点位置特异模型的单点判别力跨脏器异质性检验》)。但那个方案仍有一个前提没被检验:如果”某个点位对某脏腑判别力最强”这一结论只在一家实验室、一台采集仪上成立,它就不是传统宣称的”位置-脏腑结构映射”,而只是数据偶然。测量学上这叫泛化稳定性——换个仪器、换个中心重测,同一模型的判别力(AUC)应当大体不变。

仪器间一致性这件事,现有文献给出的答案好坏两极。差的一端:2026 年一项对青少年脉搏波速度与增强指数的双设备研究显示,PulsePen 与 Arteriograph 两台设备测同一批人的设备间一致性极差,PWV 的组内相关系数(ICC)在 -0.2 到 0.1 之间,作者明确结论”两台设备的数据不应直接比较”(PMID 41612666)。好的一端:2018 年一项 254 人的研究里,新型腕式张力计对金标准 SphygmoCor 的跨设备 ICC 达到 0.78-0.85,一致性良好(PMID 29846328)。而更根本的是:2015 年有研究用人工脉搏波发生器——把同一个周期信号喂给多台设备——证明谐波成分的设备间变异系数可压到 3% 以下、ICC 高于 0.95,且预设的”设备排除标准”能自动剔除缺陷设备(PMID 25693606)。这说明仪器间信度不是天注定,是可以被工程化量化和控制的。

机器学习模型的跨中心泛化同样两极分化。2025 年一项在 45 家医院的外部验证中,手术输血风险模型的 AUROC 中位数达 0.929,跨中心相当稳定(PMID 40577014);但同年另一项跨地域(北美/欧洲)外部验证里,脊柱转移手术大出血预测模型 AUC 只有 0.631、校准斜率 0.28,彻底失败,归因于过拟合与分布漂移(PMID 40157430)。方法学层面,2023 年一项三站点模型迁移评估给出关键设计经验:多站点学习不能提升内部判别力,但能显著提升模型迁移时的稳定性(PMID 37647884)。

于是检验设计可以直接落地:双中心 × 双设备,同步采集同一批受试者的六点脉搏波。逐脏腑建立单点位判别模型后,计算三组指标:① 跨设备 ΔAUC——同一模型换设备后判别力掉多少;② ΔAUC 的跨中心 ICC——降幅本身是否可复现;③ 仪器 × 位置交互效应——若交互显著,说明”哪个点位判哪个脏腑”随仪器改变,位置特异性就是仪器伪影而非生理事实。参照 25693606 的做法,前置设备排除标准先筛掉缺陷仪器,再谈映射。这个设计有一个关键优点:双设备采同一批受试者,把个体间变异从设备对比中彻底剥离——两台设备读数的任何差异只能来自仪器本身或仪器与点位的交互,而不是人与人不同。样本量上,跨设备信度研究通常数十人规模即可给出 ICC 的可用置信区间;真正的瓶颈在判别模型侧,各脏腑分组的阳性样本量决定 ΔAUC 置信区间的宽度,建议先做小规模可行性预实验,确认单点位判别力高于机会水平,再上双中心全量采集。

易学一侧,这个设计对准的是一个很古老的承诺。藏医六点位候脏与中医寸关尺”分部候脏”(《素问》三部九候)同属一个结构主张:腕部不同位置承载不同脏腑的信息,位置与脏腑之间存在稳定映射。这类主张要成立,检验标准应当是”结构保持”——换仪器、换中心,六个点位判别力的相对排序(哪个强、哪个弱)应该大体不变,而不是某一个点的绝对 AUC 数值逐位相等。数值漂移可以校正,排序崩塌才是结构性证伪。

证据与局限

已验证: 脉搏波测量的仪器间信度可用标准化手段量化并工程化控制(PMID 25693606);ML 模型跨中心泛化成败两极分化,且风险因素(过拟合、分布漂移、病例构成差异)已被识别(PMID 40577014, 40157430, 37647884)。

部分验证: 新型脉搏采集设备与金标准的跨设备一致性可以做到 ICC 0.78-0.85,但不同设备族之间差异巨大,甚至出现负值 ICC(PMID 29846328 vs 41612666)——”跨设备一致性可好可坏”本身是有证据的,落到具体设备必须逐台验证。

待验证: 藏医六点脏腑映射的跨设备/跨中心判别力一致性——目前没有任何直接研究,本条给出的是完整可测方案:双中心双设备同步采集 + 逐脏腑单点位模型 + ΔAUC/ICC + 仪器×位置交互效应检验。

相关阅读

L1 网络状态:该主题尚未进入 L1 对应网络(2026-11-10 实读 CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md,grep 藏医/脉位/六点/寸关尺/跨设备/泛化/判别力 均无对应边)→ 证据等级按 C 上限处理。

证据等级: C
参考文献: PMID 41612666, PMID 29846328, PMID 25693606, PMID 40577014, PMID 40157430, PMID 37647884

证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


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。

Does the Six-Point Position-Specific Model Remain Robust When Switching Devices? A Design for Cross-Center/Cross-Instrument Generalization Stability

Evidence level: C

A TCM pulse diagnosis model trained on a single-center, single-device setup: how much discriminative power remains after changing equipment and centers—this is the sole threshold for upgrading it from “a lab discovery” to a “reusable diagnostic tool.”

Traditional Tibetan medicine divides the wrist palpation into fixed points, each traditionally associated with specific organ information. Previously, we proposed a cross-organ heterogeneity test for single-point discriminative power (see “Cross-Organ Heterogeneity Test of Single-Point Discriminative Power in the Six-Point Position-Specific Model of Tibetan Medicine”). However, that design left an untested premise: if the conclusion that “a specific point has the strongest discriminative power for a certain organ” holds only in one laboratory and on one acquisition device, it is not the claimed “position-organ structural mapping,” but merely data chance. In measurement science, this is called generalization stability—re-testing across instruments and centers should yield largely unchanged discriminative power (AUC).

Existing literature presents a polarized picture of instrument consistency. At the low end: a 2026 study of adolescent pulse wave velocity and augmentation index using dual devices found poor cross-instrument agreement between PulsePen and Arteriograph, with PWV intraclass correlation coefficient (ICC) ranging from -0.2 to 0.1; authors concluded “data from the two devices should not be directly compared” (PMID 41612666). At the high end: a 2018 study of 254 participants reported excellent cross-device ICC (0.78–0.85) between a novel wrist tension device and the gold-standard SphygmoCor (PMID 29846328). More fundamentally, a 2015 study using artificial pulse wave generators fed identical periodic signals to multiple devices demonstrated that harmonic component variation coefficients could be suppressed below 3%, ICC exceeding 0.95, and preset “device exclusion criteria” automatically eliminated defective instruments (PMID 25693606). This indicates instrument reliability is not predetermined but can be engineered, quantified, and controlled.

Cross-center generalization of machine learning models is similarly polarized. A 2025 external validation across 45 hospitals showed median AUROC of 0.929 for a surgical transfusion risk model, demonstrating stable cross-center performance (PMID 40577014); yet another 2025 cross-regional (North America/Europe) external validation yielded only AUC 0.631 and calibration slope 0.28 for a spinal metastasis major hemorrhage prediction model, complete failure attributed to overfitting and distribution shift (PMID 40157430). Methodologically, a 2023 three-site model transfer evaluation provided key design insights: multi-site learning cannot improve internal discriminative power but significantly enhances stability during model transfer (PMID 37647884).

Thus, the test design can be directly implemented: dual-center × dual-device synchronous acquisition of six-point pulse waves from the same cohort. After building single-point discriminative models for each organ, calculate three metrics: ① cross-device ΔAUC—how much discriminative power drops when the same model changes devices; ② cross-center ICC of ΔAUC—whether the drop itself is reproducible; ③ instrument × position interaction effect—if significant, it indicates “which point diagnoses which organ” shifts with the instrument, meaning positional specificity is instrumental artifact rather than physiological fact. Following the approach of PMID 25693606, pre-screen device exclusion criteria to eliminate defective instruments before mapping. This design has one key advantage: acquiring the same cohort on both devices completely removes inter-individual variability from device comparisons—any difference between device readings stems solely from the instruments themselves or their interaction with positions, not from person-to-person differences. For sample size, cross-device reliability studies typically require dozens of participants to yield usable ICC confidence intervals; the true bottleneck lies in the discriminative model side, where positive sample sizes per organ group determine ΔAUC confidence interval width. We recommend starting with a small-scale feasibility pilot to confirm single-point discriminative power exceeds chance level before proceeding to full dual-center data collection.

On the Yijing Studies side, this design targets an ancient promise. Tibetan medicine’s six-point organ waiting and Chinese medicine’s “分部候脏” (regional organ waiting, as described in the Su Wen Three-Part Nine-Pulse system) share the same structural claim: different wrist locations carry distinct organ information, implying a stable mapping between position and organ. For such claims to hold, the validation criterion must be “structural preservation”—changing instruments or centers should preserve the relative ranking of discriminative power across the six points (which is strongest, which weakest), rather than requiring absolute AUC values to be equal across positions. Numerical drift can be corrected; ranking collapse constitutes structural falsification.

Evidence and Limitations

Verified: Inter-instrument reliability of pulse wave measurements can be quantified using standardized methods and controlled through engineering means (PMID 25693606); ML models show polarized success or failure in cross-center generalization, with risk factors (overfitting, distribution drift, case composition differences) identified (PMID 40577014, 40157430, 37647884).

Partially Verified: Cross-device agreement between novel pulse acquisition devices and the gold standard can achieve ICC values of 0.78–0.85, yet differences between device families are substantial, sometimes resulting in negative ICC values (PMID 29846328 vs. 41612666)—the evidence confirms that “cross-device agreement can be good or poor,” necessitating individual validation for each specific device.

Pending Verification: Cross-device and cross-center discriminative consistency of the Tibetan medicine six-organ mapping. No direct study currently exists; this entry outlines a fully testable protocol: simultaneous dual-center, dual-device data acquisition; single-point models for each organ; ΔAUC/ICC analysis; and testing of instrument × position interaction effects.

Related Reading

L1 Network Status: This topic has not yet entered the corresponding L1 network (2026-11-10 real-time reading of CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md; grep for Tibetan medicine/pulse position/six points/cun-guan-chi/cross-device/generalization/discriminative power yielded no corresponding edges) → evidence level treated as C upper bound.

Evidence Level: C
References: PMID 41612666, PMID 29846328, PMID 25693606, PMID 40577014, PMID 40157430, PMID 37647884

Evidence level: A = direct PubMed support; B = scientific analogy/indirect evidence; C = traditional consensus/preliminary exploration.


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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *