三模态交叉校验的误差不是沿链累加:三点之间没有「链」,只有闭合差

Written by

in

Language

Evidence level: C

三点两两校验构成的是闭合三角而非串联链,「误差随环节数放大」这个斜率在三点设计里根本没有自由度可估;真正能算的只有闭合差。

正文

一、三点不是链,估不出「放大率」

链条的前提是有 N 个串联环节,每环把上一环的输出当输入。三模态交叉校验不是这样:FOT 负载、EIT phantom、RIP 肺量计是三个并列测同一批受试者的系统,两两交叉只产生三条边,不产生「第 N 环」。要在数据里估「误差随环节数放大」的斜率,自变量至少要有四个以上取值——三点设计里这个斜率没有自由度。更根本的是,(A−B)+(B−C)+(C−A)=0 是恒等式,不是结论。能观测的只有「闭合差不为零」,这个残差是三台设备方法学误差的可观测投影,不是传播过程的快照。

二、可互换性不能沿链继承

这里要把两个被混在一起的量分开。不确定度可以沿校准层级传播、按误差结构相加;可互换性不能继承——参考物质必须在每一级单独证明它与新鲜临床样本的响应一致。2023 年《Clinical Chemistry》的指引专门回答了「校准层级中哪些校准品必须与临床样本可互换」,答案是逐级都要(PMID 36660772)。而且合格的可互换参考物质本身极稀缺:同型半胱氨酸、碱性磷酸酶、血浆肾素活性、免疫球蛋白 M 的候选评估都是逐个做的专项工作,结论常是不合格(PMID 35977428、PMID 32271089、PMID 37470745、PMID 37571860)。因此「链上第 N 环的一致性界限随 N 放大」这个提法预设了「一致性是可传递的量」,而这个预设已被计量学否定。

三、放大、相消,还是不动,取决于误差是否相关

同一批误差相加,三条路给出数量级完全不同的答案:独立随机误差按方差相加,两环合成的标准差是单环的 1.41 倍;同向相关的系统误差按标准差线性相加,是 2 倍;若两环共用同一上游校准件,该误差在差值里被完全消掉。三模态恰好同时踩中后两种。如果 RIP 与 FOT 的容积定标都用同一支 3 L 注射器,这两条边的容积尺度就被钉在同一把尺子上、误差完全相关,差值恒为零——此时的「一致性极好」是共模的结果,不是两台各自准确。不确定度沿链叠加的前提是各环相互独立(PMID 21795407、PMID 32238602)。

四、三模态的偏差都是乘性的,不是加性的

如果三台设备的偏差都是常数偏移,闭合差检验会非常好用。但它们不是。EIT 与总肺容积的关系本身是参数化非线性的(PMID 15886435),线性度在最大努力呼吸与用力呼气动作下会破(PMID 28004642),非插管受试者还需专门的容积校准(PMID 31195395)。FOT 的参考负载校准是两点线性程序(PMID 9538557),两点之间的线性假设恰是它在偏离区失效的地方。phantom 侧更麻烦:3D 打印二维 phantom(PMID 38267531)与 active element 设计(PMID 10984936)验证的是算法与硬件,验证不了电极-皮肤界面在活体上的行为。三台设备在活体上的偏差因此难以互相抵消——来源不重叠,方向也不相反。

五、三条边的一致性界限不能直接比大小

即使决定做误差预算,还有两个坑。第一,Bland-Altman 一致性界限的置信区间有专门的样本量算法(PMID 27838682);三条边若各用部分样本,每条 LoA 的 CI 都更宽,「哪条更差」很可能只是在比 CI 的重叠程度。第二,重复测量数据不独立——同一受试者被三台设备各测多次,按独立样本处理会把 LoA 的方差估小(PMID 10907753、PMID 9820922)。叠加第四节的乘性偏差,常数 LoA 假设已破,须改比例 LoA 或先做对数变换,否则「放大率」回归的自变量与因变量同时被污染。

六、三角验证的收敛口径用错了范畴

心理测量学里三角验证有成熟框架:MTMM 矩阵,要求同一构念被不同方法测到并收敛(PMID 37530967 用它评估临床胜任力,PMID 26794016 是方法学底本)。但它成立的前提是 trait 与 method 可分离——不同方法测的必须是同一个构念。三模态测的是三个不同物理量:FOT 测阻抗谱变化,EIT 测胸腔内电导率分布重构,RIP 测胸腹表面位移。它们不是「一个构念的三种方法」,是三个构念。要求三者收敛,等于要求三个不同物理量给出同一个数——这是范畴错误。

反转:能答的不是放大率,是闭合差

真正可算、且不需要金标准的只有一个量:闭合差。它是三条边两两差值绕一圈的总和,无偏时恒为零;实际不为零的部分,就是「三条边是否属于同一坐标系」的直接度量。它有三个好处:不需要金标准、不需要知道谁对谁错、不依赖样本量假设。而「误差随 N 放大」在三点设计里不可识别,四环以上才可谈。所以正确顺序是:先算闭合差,通过了再谈误差预算。

四格判据

  • 闭合差 ≤ 三边最小 LoA,且不随体型/容积幅度系统变化 ⇒ 三边同坐标系,可进误差预算;但须补第四环才有自由度估放大率
  • 闭合差显著大于任一边 LoA ⇒ 至少一台带固定偏移,先定偏移源(换第四台或换可互换参考物质);此阶段任何「累积率」数字无意义
  • 闭合差小但三边 LoA 都很宽 ⇒ 三台一起漂在同一方向,三角验证给出虚假信心——最危险的一格,也最容易被写成「三模态高度一致」
  • 闭合差随容积幅度或体型系统变化 ⇒ 偏差是乘性的,常数 LoA 假设失效,须改比例 LoA 或 log 变换后重算

证据与局限

已验证:可互换性须在校准层级的每一级单独证明、不可从上级继承(PMID 36660772);候选参考物质的可互换性须逐物质专项评估且常不合格(PMID 35977428、PMID 32271089、PMID 37470745、PMID 37571860);不确定度沿链叠加以各环独立为前提(PMID 21795407、PMID 32238602);EIT 与肺容积关系非线性、在大力呼吸下线性度破坏(PMID 15886435、PMID 28004642);Bland-Altman 一致性界限的 CI 需专门样本量算法、重复测量数据须按非独立处理(PMID 27838682、PMID 10907753、PMID 9820922)。

部分验证:phantom 能验证算法与硬件但无法验证活体电极-皮肤界面(PMID 38267531、PMID 10984936 均为 phantom 设计与验证,未覆盖在体界面);MTMM 框架要求 trait 与 method 可分离,该前提在三模态场景是否成立属推论、未直接实证(PMID 37530967、PMID 26794016)。

待验证:三台设备在同一批受试者上的闭合差量级从未被实测;闭合差是否随体型/呼吸幅度系统变化未知;三模态是否共用同一容积定标(共模风险)在多数研究报告中未注明;四环以上设计的误差放大率目前无任何呼吸力学数据;FOT 参考负载与 EIT phantom 之间是否存在第四个独立参照,尚无方案。

L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 的节点表无「测量学维度(参考物质 / 可互换性 / 闭合差 / 误差预算)」这一类节点,边表 3.1–3.4 只登记五行↔五脏/五色/五音/五方/五季/五志、八卦↔五行、天干地支↔五行,无任何以三模态交叉校验、参考样本链或误差传播为源或目标的边;EDGE_TABLE_science-yixue.md 仅有概念计数,无 L1/L2/L3 层级与 active/open/conflict 状态判定。

相关阅读

证据等级: C
参考文献: PMID 36660772, PMID 35977428, PMID 32271089, PMID 37470745, PMID 37571860, PMID 21795407, PMID 32238602, PMID 15886435, PMID 28004642, PMID 31195395, PMID 9538557, PMID 38267531, PMID 10984936, PMID 27838682, PMID 10907753, PMID 9820922, PMID 37530967, PMID 26794016


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-modal cross-validation error does not accumulate along a chain: there is no chain between three points, only a closure error.

Evidence level: C

Three points mutually validated form a closed triangle rather than a serial chain; the slope of “error amplification per link” has no degrees of freedom to estimate in a three-point design. The only calculable quantity is the closure error.

Main Text

I. Three Points Are Not a Chain; No “Amplification Rate” Can Be Estimated

The premise of a chain involves $N$ serial links where each link treats the output of the previous one as its input. Trimodal cross-validation does not operate this way: FOT load cells, EIT phantoms, and RIP spirometers are three parallel systems measuring the same cohort of subjects. Mutual cross-validation between pairs generates only three sides, not a “link $N$.” To estimate the slope of “error amplification per link,” the independent variable requires at least four distinct values; in a three-point design, this slope lacks degrees of freedom. More fundamentally, $(A-B)+(B-C)+(C-A)=0$ is an identity, not a conclusion. What can be observed is only “non-zero closure,” which represents the observable projection of methodological errors across three devices, not a snapshot of propagation.

II. Interchangeability Cannot Be Inherited Along the Chain

Here, two conflated quantities must be separated. Uncertainty can propagate along calibration hierarchies and sum according to error structures; interchangeability cannot be inherited—reference materials must be individually proven at each level to demonstrate consistent response with fresh clinical samples. The 2023 guidance in Clinical Chemistry specifically addresses “which calibrators must be interchangeable with clinical samples across calibration levels,” answering that they must all be at each step (PMID 36660772). Furthermore, qualified interchangeable reference materials are extremely scarce: candidate assessments for homocysteine, alkaline phosphatase, plasma renin activity, and immunoglobulin M are all conducted as individual specialized tasks, often concluding with non-compliance (PMID 35977428, PMID 32271089, PMID 37470745, PMID 37571860). Therefore, the claim that “consistency limits scale with $N$ along the chain” presupposes that consistency is a transmissible quantity, a premise negated by metrology.

III. Amplification, Cancellation, or Stasis Depends on Error Correlation

Summing the same errors yields vastly different magnitudes depending on the path: independent random errors sum by variance, making the combined standard deviation of two rings $1.41$ times that of a single ring; co-directional systematic errors sum linearly by standard deviation, doubling the total; if two rings share an upstream calibration artifact, that error is completely canceled in the difference. Trimodal systems simultaneously engage the latter two scenarios. If RIP and FOT volumetric calibration both use a single 3 L syringe, the volumetric scales of these two sides are fixed to the same ruler, errors are fully correlated, and the difference is identically zero—here, “excellent consistency” is a result of common-mode coupling, not individual accuracy. Uncertainty summation along a chain assumes independence among rings (PMID 21795407, PMID 32238602).

IV. Trimodal Biases Are Multiplicative, Not Additive

If biases across three devices were constant offsets, closure testing would be highly effective. They are not. The relationship between EIT and total lung volume is inherently parameterized nonlinear (PMID 15886435); linearity breaks under maximal inspiratory effort and forced expiratory maneuvers (PMID 28004642), and non-intubated subjects require specific volumetric calibration (PMID 31195395). FOT reference load calibration relies on a two-point linear program (PMID 9538557), and the linear assumption between these two points is precisely where it fails in deviation zones. The phantom side presents further complications: 3D-printed two-dimensional phantoms (PMID 38267531) and active element designs (PMID 10984936) validate algorithms and hardware but cannot verify electrode-skin interface behavior in vivo. Consequently, biases across three devices in vivo are difficult to cancel out—sources do not overlap, and directions are not opposite.

V. Consistency Limits Across Three Sides Cannot Be Directly Compared

Even if an error budget is decided, two pitfalls remain. First, confidence intervals for Bland-Altman consistency limits have specific sample size algorithms (PMID 27838682); if three sides use partial samples, each LoA CI is wider, and determining “which is worse” may merely compare CI overlap. Second, repeated measurement data are not independent—measuring the same subject multiple times with three devices; treating them as independent samples underestimates LoA variance (PMID 10907753, PMID 9820922).叠加 the multiplicative biases from Section IV, the constant LoA assumption is violated; proportional LoA or log transformation must be applied before regression, otherwise both the independent and dependent variables in “amplification rate” are contaminated.

VI. The Convergence Criteria for Triangulation Misapply the Category

Psychometrics offers mature frameworks for triangulation: the MTMM matrix, requiring convergence of different methods measuring the same construct (PMID 37530967 uses it to assess clinical competence, PMID 26794016 is the methodological foundation). However, it relies on the premise that trait and method are separable—different methods must measure the same construct. Trimodal systems measure three distinct physical quantities: FOT measures impedance spectral changes, EIT reconstructs intrathoracic conductance distribution, and RIP measures thoraco-abdominal surface displacement. They are not “three methods of one construct,” but three constructs. Requiring convergence among them demands three different physical quantities yield the same number—a category error.

Inversion: What Can Be Answered Is Not Amplification Rate, But Closure Error

The only quantity that is calculable and does not require a gold standard is closure error. It is the sum of differences around the loop formed by pairwise differences of three sides; unbiased, it is identically zero. The non-zero portion directly measures “whether the three sides belong to the same coordinate system.” It offers three advantages: no gold standard required, no need to know who is correct, no reliance on sample size assumptions. “Error amplification with $N$” is unidentifiable in three-point designs; discussion requires four or more rings. Thus, the correct sequence is: calculate closure error first; only if passed, proceed to error budgeting.

Four-Grid Criterion

  • Closure Error ≤ Minimum LoA of any side, and does not vary systematically with body size/volume amplitude ⇒ Three sides share a coordinate system; eligible for error budgeting; but a fourth ring is needed to estimate amplification rate
  • Closure Error significantly greater than any single side’s LoA ⇒ At least one device has a fixed offset; identify the source first (add a fourth device or use interchangeable reference materials); at this stage, any “cumulative rate” figure is meaningless
  • Closure Error small but all three sides have wide LoAs ⇒ All three drift together in the same direction; triangulation provides false confidence—the most dangerous category, yet most easily written as “high trimodal consistency”
  • Closure Error varies systematically with volume amplitude or body size ⇒ Bias is multiplicative; constant LoA assumption fails; must switch to proportional LoA or re-calculate after log transformation

Evidence and Limitations

Verified: Interchangeability must be proven individually at each calibration level, not inherited from higher levels (PMID 36660772); candidate reference material interchangeability requires individual specialized assessment and is often non-compliant (PMID 35977428, PMID 32271089, PMID 37470745, PMID 37571860); uncertainty summation along a chain assumes independence among rings (PMID 21795407, PMID 32238602); EIT relationship with lung volume is nonlinear and linearity breaks under heavy breathing (PMID 15886435, PMID 28004642); Bland-Altman LoA CIs require specific sample size algorithms and repeated measurements must be treated as non-independent (PMID 27838682, PMID 10907753, PMID 9820922).

Partially Verified: Phantoms can validate algorithms and hardware but cannot validate in vivo electrode-skin interfaces (PMID 38267531, PMID 10984936 both address phantom design/validation without covering in vivo interfaces); MTMM framework requires trait-method separability, which holds in trimodal scenarios is inferential and not directly empirically verified (PMID 37530967, PMID 26794016).

To Be Verified: The magnitude of closure error across three devices on the same cohort has never been empirically measured; whether closure error varies systematically with body size/breath amplitude is unknown; whether trimodal systems share a common volumetric calibration (common-mode risk) is not noted in most reports; error amplification rates for designs with four or more rings currently have no respiratory mechanics data; whether a fourth independent reference exists between FOT reference loads and EIT phantoms, no proposal exists.

L1 Network Status: This topic has not yet entered the L1 corresponding network—the node table in CORRESPONDENCE_NETWORK.md lacks a node category for “Metrological Dimension (Reference Materials / Interchangeability / Closure Error / Error Budget),” and edge tables 3.1–3.4 only register Five Phases ↔ Five Organs/Five Colors/Five Tones/Five Directions/Five Seasons/Five Emotions, Eight Trigrams ↔ Five Phases, Heavenly Stems/Earthly Branches ↔ Five Phases; no edges originate or terminate from/to trimodal cross-validation, reference sample chains, or error propagation. EDGE_TABLE_science-yixue.md contains only concept counts, lacking L1/L2/L3 hierarchies and active/open/conflict state determinations.

Related Reading

Evidence Level: C
References: PMID 36660772, PMID 35977428, PMID 32271089, PMID 37470745, PMID 37571860, PMID 21795407, PMID 32238602, PMID 15886435, PMID 28004642, PMID 31195395, PMID 9538557, PMID 38267531, PMID 10984936, PMID 27838682, PMID 10907753, PMID 9820922, PMID 37530967, PMID 26794016


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 *