剂量-反应曲线:是“量”出来的,还是“选”出来的?

Written by

in

Language

Evidence level: B

你想知道辐射剂量和病毒突变之间是线性、有阈值、还是先抑后扬——可这条曲线的形状,有相当一部分不是被数据“量”出来的,而是被拟合尺子“选”出来的。

正文

易学描述“气”从不只讲强弱,还讲“太过”“不及”与“平气”——说的是形状,不是数值。现代剂量-反应分析里也有一场同构的争论:低剂量辐射的影响究竟是线性的(LNT)、有阈值的,还是先抑后扬(hormesis)?搬到“海拔-病毒突变”这个自然实验里,问题反而更清楚——在这里,“曲线形状”恐怕还没到能估的时候,因为它同时撞上四重错位。

一、三种形状,三套世界观

线性无阈(LNT)假设每个粒子事件独立、机体没有净修复,斜率恒定、剂量归零则风险归零;阈值模型假设修复系统有容量,未达阈值不显影;双相/饱和模型假设低剂量激活抗氧化与 DNA 修复,于是低剂量段出现“低于对照”的响应(PMID 30262515)。三者不是细节之争,而是三套生物学假设之争;而在操作层面,选哪个往往由 ΔAIC 最小决定——这就是入口。

二、第一重错位:横轴自己就不确定

要谈剂量-反应,横轴必须是剂量或通量,而不是海拔。可海拔同时捆着低氧、低气压、强 UV、低温与稀疏人口。即便把海拔翻译成通量,翻译本身也有分段争议:1987 年日本团队实测给出“6000 米以下近似二次曲线、以上转线性”的两段行为(PMID 3667276);航空实测把通量随地磁纬度的差异推到 8 倍,纬度权重不亚于海拔(PMID 12033224);蒙特卡洛模拟显示通量在 Pfotzer 峰(约 15-20 公里)之后回落(PMID 26256622)——通量随海拔并非单调,而人类居住带只落在其上升段的前一小截。再叠上飞机结构自屏蔽带来的 12-16% 系统偏差(PMID 37227153):你是在一根自带几十个百分点不确定带、连“是否二次”都待定的横轴上找形状。

三、第二重错位:形状的证据都在够不着的那一段

三模型的分歧点几乎全在低剂量区:高剂量下它们都指向正斜率,形状之争只在低剂量段才有意义——而这一段恰恰是现有检测体系最难识别的(PMID 18648564)。支撑判别的历史证据却来自原子弹幸存者、CT 队列、果蝇照射——急性高剂量率实验;而海拔梯度给的是慢性低剂量率:年有效剂量从海平面约 0.3-0.4 mSv 到 4000 米约 2-3 mSv。拿另一段路的弯道半径描述这一段路,本就是越界。

四、第三重错位:那把尺子只有一半准

AIC 选模在形状判断上的效力,已被两次独立评估质疑。Roberts 与 Martin 用库克县 1987-2000 年数据模拟,发现 AIC 在多种情景下只有约一半时间选对模型,“模棱两可”是常态(PMID 17005626)。Beyea 与 Hoffmann 在 2023 年把原子弹、INWORKS、Techa River 三大队列数据线性化后重跑 5000 次,测得假非线性检出率 24-61%、假阈值出现频率超 50%、平均假阈值约 0.05 Gy——原因是 AIC 对嵌套模型的隐含 α 等价于 0.157,远高于名义 5%(PMID 36787654)。用 AIC 判形状,本身就是一台高假阳机器,多模型平均并不能改善(同文)。

五、第四重错位:环境剂量下的“平曲线”可能是“测不出”

环境人群研究常被读成“支持阈值或 hormesis”:中国阳江高本底地区 10 万人队列全癌 ERR/Sv = -0.10(95% CI -0.67 至 0.69)、剂量组间无差异(PMID 22724612);美国高海拔县全部癌症的 SMR 与率比显著低于近海平面县(同文);另一项 GIS 中位海拔研究给出自然本底辐射与癌症死亡率负相关、而氧浓度与气压正相关的混杂图景(PMID 23550227)。这些结果确实与 LNT 不符——但负 ERR 且置信区间横跨零,说明的是“现有样本量与剂量范围内检测不出正效应”,而非“曲线是平的或向下”。把“低于检测限”读成“形状已确定”,是四重错位里最贵的一次误读。

六、该怎么办:先做“形状可辨识性预检”

四条合起来,建议不是“换一个更好的模型”,而是先别谈形状。三步预检:①换轴并封轴——横轴用重建通量而非海拔,并把物理模型自身的系统不确定带明确画进横轴误差范围;②先算功效——在给定跨站点数与采样窗口下用模拟数据检验三模型能否被分开,把 PMID 36787654 的事后证伪前移为事前设计;③设定降级规则——若功效不足,结局从“曲线形状”降级为“是否存在可检测的边际增量”,只报后者区间,不报形状标签。双相模型研究也给了同样前提:要谈兴奋的“剂量区宽度”与“最大刺激幅度”,先得有足够的低剂量点与重复(PMID 23266374, 25523646, 22942868)。低剂量点不够,形状就不可辨识,再好的模型也补不回来。

对易学的意义:《素问·天元纪大论》讲“太过者先天,不及者后天”——古人描述“气”同样分三档:太过、不及、平气。关键在他们怎么定这三档:靠“至而先后”(节气是否按时到、气候是否与所主相合),是一种分类判据,不是估一条函数。这不是不懂定量,而是在自变量(“气”的量)无法直接测量时,把连续量降维成三分,反而更稳健。现代剂量-反应之争走到第四重错位,得到的建议竟一模一样:当横轴不确定、低段数据稀薄、判据又偏假阳,最可靠的结论形态不是“曲线是 A 型”,而是“超出可检测边际增量/没有超出/还不能判”——三分法。古人三分,不是精度不够,是承认了测量的边界。

证据与局限

已验证: 宇宙线中子剂量当量率随海拔呈“6000 米以下二次、以上线性”的分段行为(PMID 3667276);通量随地磁纬度的变化可达 8 倍,纬度权重不亚于海拔(PMID 12033224);通量在 Pfotzer 峰后回落、非单调(PMID 26256622);飞机自屏蔽造成 12-16% 系统偏差(PMID 37227153);AIC 选模在形状判断上正确率约一半(PMID 17005626);大辐射队列中假非线性检出率 24-61%、假阈值超 50%(PMID 36787654);双相模型判据与定量方法已建立(PMID 23266374, 25523646, 22942868)。

部分验证: 低剂量段突变反应难以被现有检测体系识别(PMID 18648564);环境剂量下高本底地区癌症死亡率未见正效应(PMID 22724612);海拔与癌症死亡率呈负相关但混杂于自然本底辐射、氧浓度与气压(PMID 23550227);LNT 作为唯一默认模型的地位仍在争论中(PMID 30262515)。

待验证: “海拔→通量→病毒突变率”的斜率究竟取线性、阈值还是饱和形式;把横轴从海拔换成重建通量后,形状结论是否稳定;在给定跨站点数与采样窗口下,三模型能否被统计功效分开;若把结局从“曲线形状”降级为“是否存在可检测边际增量”,结论会如何改变。

相关阅读

证据等级: B(物理端通量-海拔曲线与统计端 AIC 假阳率均有 PubMed 直接支撑;“四重错位”框架与“形状可辨识性预检”提案属本文件提出的方法学整合,尚无直接研究)
参考文献: PMID 3667276, PMID 12033224, PMID 26256622, PMID 37227153, PMID 30262515, PMID 18648564, PMID 17005626, PMID 36787654, PMID 22724612, PMID 23550227, PMID 23266374, PMID 25523646, PMID 22942868


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。

Dose-Response Curve: Is It Determined by \”Amount\” or \”Selection\”?

Evidence level: B

You want to know if the relationship between radiation dose and viral mutation is linear, threshold-based, or hormetic—but a significant portion of the curve’s shape is not “measured” by data, but “selected” by the ruler of fitting.

Main Text

Yijing Studies describes “Qi” not merely in terms of strength, but also in terms of “excess,” “deficiency,” and “balanced Qi”—referring to form, not numerical values. A homologous debate exists in modern dose-response analysis: is the effect of low-dose radiation linear (LNT), threshold-based, or hormetic? When transposed to the “altitude-viral mutation” natural experiment, the question becomes clearer—here, the “curve shape” is arguably not yet estimable because it simultaneously encounters four misalignments.

I. Three Shapes, Three Worldviews

The Linear No-Threshold (LNT) hypothesis assumes independent particle events, no net repair capacity, a constant slope, and zero risk at zero dose; the threshold model assumes a repair system with capacity that remains inactive below a certain level; the biphasic/saturation model assumes low doses activate antioxidant and DNA repair mechanisms, resulting in a response “below control” at low doses (PMID 30262515). These are not merely technical details but competing biological assumptions. Operationally, the choice often depends on minimizing ΔAIC—this is the entry point.

II. First Misalignment: The X-Axis Itself Is Uncertain

To discuss dose-response, the x-axis must represent dose or flux, not altitude. However, altitude simultaneously correlates with hypoxia, low pressure, strong UV, low temperature, and sparse populations. Even translating altitude into flux involves segmentation disputes: a 1987 Japanese team measured a “quadratic curve below 6000 meters transitioning to linear above” (PMID 3667276); aviation measurements show flux varies by geomagnetic latitude up to 8 times, with latitude weighting comparable to altitude (PMID 12033224); Monte Carlo simulations indicate flux declines after the Pfotzer peak (approx. 15-20 km) (PMID 26256622)—flux is not monotonic with altitude, yet human habitation zones fall only within the initial rising segment. Adding a 12-16% systematic bias from aircraft structural self-shielding (PMID 37227153): you are searching for shape on an x-axis with a built-in uncertainty band of several tens of percent, where even “quadratic vs. linear” is undecided.

III. Second Misalignment: Evidence for Shape Lies in the Unmeasurable Range

The divergence among the three models occurs almost entirely in the low-dose region: high doses all point to positive slopes, so the shape debate is only meaningful at low doses—precisely the segment hardest to identify with current detection systems (PMID 18648564). Historical evidence supporting discrimination comes from atomic bomb survivors, CT cohorts, and Drosophila irradiation—acute high-dose-rate experiments; whereas the altitude gradient provides chronic low-dose-rate exposure: effective annual dose ranges from approx. 0.3-0.4 mSv at sea level to 2-3 mSv at 4000 meters. Describing this segment with the radius of curvature from another road is inherently out of bounds.

IV. Third Misalignment: The Ruler Is Only Half-Accurate

The validity of AIC for shape selection has been questioned by two independent assessments. Roberts and Martin simulated using Cook County data from 1987-2000, finding AIC selects the correct model in only about half of scenarios, with “ambiguity” being the norm (PMID 17005626). Beyea and Hoffmann re-ran 5000 iterations after linearizing data from three large cohorts (atomic bomb, INWORKS, Techa River) in 2023, measuring a false non-linearity detection rate of 24-61%, false threshold appearance frequency exceeding 50%, and an average false threshold of approx. 0.05 Gy—because AIC’s implied α for nested models equals 0.157, far exceeding the nominal 5% (PMID 36787654). Using AIC to judge shape is essentially running a high false-positive machine; multi-model averaging does not improve this (same paper).

V. Fourth Misalignment: A “Flat Curve” in Environmental Dose May Simply Be “Undetectable”

Environmental population studies are often read as supporting thresholds or hormesis: the 100,000-person cohort in Yangjiang, China (high background area) found a full-cancer ERR/Sv of -0.10 (95% CI -0.67 to 0.69) with no difference between dose groups (PMID 22724612); in the USA, SMRs and rate ratios for all cancers in high-altitude counties were significantly lower than near-sea-level counties (same paper); another GIS study on median altitude showed a negative correlation between natural background radiation and cancer mortality, while oxygen concentration and pressure showed positive correlations (confounding landscape) (PMID 23550227). These results indeed contradict LNT—but negative ERR with a confidence interval crossing zero indicates “no detectable positive effect within current sample size and dose range,” not “the curve is flat or downward.” Reading “below detection limit” as “shape determined” is the most costly misinterpretation among the four misalignments.

VI. What to Do: First Conduct a “Shape Identifiability Pre-Check”

Taken together, the recommendation is not “switch to a better model,” but “stop discussing shape first.” Three-step pre-checks: ① Change axis and lock it—use reconstructed flux rather than altitude for the x-axis, and explicitly include the physical model’s systematic uncertainty band within the x-axis error range; ② Calculate power first—use simulated data to test if the three models can be distinguished given a fixed number of sites and sampling window, moving the post-hoc falsification from PMID 36787654 to a priori design; ③ Set downgrade rules—if power is insufficient, downgrade the outcome from “curve shape” to “existence of a detectable marginal increment,” reporting only the latter interval without shape labels. Biphasic model studies also establish this prerequisite: to discuss the “width of the excited dose range” and “maximum stimulation amplitude,” one must first have sufficient low-dose points and replication (PMID 23266374, 25523646, 22942868). Without sufficient low-dose points, shape is unidentifiable; no model can compensate for this.

Significance for Yijing Studies: The Su Wen · Tian Yuan Ji Da Lun states “excess precedes the heavens, deficiency follows the earth”—ancients described “Qi” in three tiers: excess, deficiency, and balanced. The key lies in how these tiers are defined: via “arrival timing” (whether solar terms arrive on schedule and whether climate matches the governing element), serving as a classification criterion, not estimating a function. This is not ignorance of quantification, but rather dimensionality reduction from continuous to ternary when the independent variable (“Qi” quantity) cannot be directly measured, which is actually more robust. The modern dose-response debate reaching the fourth misalignment yields an identical recommendation: when the x-axis is uncertain, low-dose data are sparse, and criteria yield high false positives, the most reliable conclusion form is not “the curve is Type A,” but “exceeds detectable marginal increment / does not exceed / cannot yet judge”—a ternary approach. Ancient ternary classification was not due to insufficient precision, but an acknowledgment of measurement boundaries.

Evidence and Limitations

Verified: Cosmic ray neutron dose equivalent rate shows segmented behavior (“quadratic below 6000 meters, linear above”) with altitude (PMID 3667276); flux varies by geomagnetic latitude up to 8 times, with latitude weighting comparable to altitude (PMID 12033224); flux declines after the Pfotzer peak and is non-monotonic (PMID 26256622); aircraft self-shielding causes a 12-16% systematic bias (PMID 37227153); AIC selects the correct model in only about half of cases for shape judgment (PMID 17005626); false non-linearity detection rate is 24-61% and false threshold frequency exceeds 50% in large radiation cohorts (PMID 36787654); criteria and quantitative methods for biphasic models are established (PMID 23266374, 25523646, 22942868).

Partially Verified: Low-dose mutation responses are difficult to identify with current detection systems (PMID 18648564); no positive effect of cancer mortality observed in high-background areas under environmental dose (PMID 22724612); negative correlation between altitude and cancer mortality confounded by natural background radiation, oxygen concentration, and pressure (PMID 23550227); LNT’s status as the sole default model remains debated (PMID 30262515).

To Be Verified: Whether the slope from “altitude → flux → viral mutation rate” takes linear, threshold, or saturated forms; whether shape conclusions remain stable after switching the x-axis from altitude to reconstructed flux; whether the three models can be statistically separated given a fixed number of sites and sampling window; how conclusions would change if the outcome is downgraded from “curve shape” to “existence of a detectable marginal increment.”

Related Reading

Evidence Level: B (Physical flux-altitude curves and statistical AIC false-positive rates have direct PubMed support; the “Four Misalignments” framework and “Shape Identifiability Pre-Check” proposal represent methodological integration proposed in this document without direct research support)
References: PMID 3667276, PMID 12033224, PMID 26256622, PMID 37227153, PMID 30262515, PMID 18648564, PMID 17005626, PMID 36787654, PMID 22724612, PMID 23550227, PMID 23266374, PMID 25523646, PMID 22942868


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 *