五运六气+现代统计模型:叠加还是干扰?

Written by

in

Language

Evidence level: B

五运六气与现代 DLNM 模型的底层逻辑同构,但”运气粗筛能提升预测精度”是一个必须用同数据集头对头检验的假设,而不是已知结论。

正文

五运六气的核心主张,是”气候先兆 → 滞后发病”:某年岁运太过、某气失常,之后 1-3 年可能出现特定瘟疫。这个框架与现代流行病学的分布式滞后非线性模型(DLNM)在结构上惊人地同构——DLNM 的核心就是承认暴露对健康的影响”非线性且延迟”,用交叉基函数(cross-basis)同时刻画暴露-反应关系和滞后结构(PMID 20812303)。换句话说,古人用天干地支编码气候的长期节律,现代人用样条和滞后矩阵编码同样的节律,两边在做同一件事:把”过去的气候”翻译成”未来的疾病风险”。

那么问题来了:既然逻辑同构,把五运六气当作”粗筛器”(先按干支把年份分层,再进 DLNM 精算),能否带来真实的增量价值?这是本主题要检验的核心假设。

先看现代模型自身的气象预测能力。南非林波波省基于 1998-2015 年周度气象与疟疾病例建立的时间序列 DLNM,1-2 周提前预测的相关系数超过 0.8,即使提前 16 周仍保持在 0.7 以上(PMID 31784563)——气象驱动的时间序列模型本身就能干活。

但增量价值的证据是分化的。广东 2019-2023 年 RSV 研究把 DLNM 与 XGBoost+SHAP 结合后发现:对预测性能贡献最大的是滞后的 RSV 自身指标,气象变量只提供”适度”的增量价值(PMID 42602072)。泉州流感研究给出了更精确的数字:在 ARIMA 基础上加入滞后 2 周的降水作为外生变量,RMSE 仅改善 2.9%、MAE 改善 1.3%(PMID 41200038)。这说明气象信息有用,但增量小、且容易被疾病自身的历史信息吸收。

这对运气粗筛是双重警告。其一,如果纯气象变量的增量都只有个位数百分比,那么比气象更间接的”干支分层”增量可能更小、更难从噪声中分离。其二,运气理论的最大卖点是 1-3 年的长滞后(”三年化疫”),但现有 DLNM 研究普遍只用 0-3 周的短滞后就能取得不错的预测(PMID 42602072),长滞后是否真的携带独立信息,尚无人用同协议检验。

正确的研究设计其实不复杂:同一套传染病长序列数据,跑三个模型——纯 DLNM、按运气年份分层的 DLNM、运气分层+DLNM 精算的混合模型——用 AUC、校准度、Brier 分数等统一指标比较,再做时间外验证。如果混合模型不能显著胜出,那”叠加”就只是”干扰”;如果胜出,则说明干支编码确实编码了 DLNM 样条遗漏的慢周期信息。

证据与局限

已验证: DLNM 是气象-疾病建模的成熟框架(PMID 20812303);气象驱动的时间序列模型能在 1-16 周窗口内维持较高预测性能(PMID 31784563);气象变量对传染病预测的增量价值真实存在但有限(PMID 41200038、PMID 42602072)。

部分验证: 滞后疾病自身信息通常比气象因子贡献更大(PMID 42602072),提示任何”气象先兆”类方法(含运气)的增量空间可能天然受限。

待验证: “干支分层作为 DLNM 粗筛器能提升预测精度”——目前没有同数据集头对头比较研究;长滞后(1-3 年)气象-疾病关联是否携带独立于短滞后的信息,同样无直接证据。

相关阅读

证据等级: B
参考文献: PMID 20812303, PMID 31784563, PMID 41200038, PMID 42602072


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。

Five Elements and Six Qi Plus Modern Statistical Models: Overlapping or Interfering?

Evidence level: B

The underlying logic of Wu Yun Liu Qi and modern DLNM models is isomorphic; however, the hypothesis that “Yun Qi coarse screening enhances prediction accuracy” must be validated via head-to-head comparison using the same dataset, rather than being treated as an established conclusion.

Main Text

The core assertion of Wu Yun Liu Qi is “climatic precursors → delayed onset”: when a specific year’s Sui Yun is excessive or a specific Qi is abnormal, specific epidemics may emerge 1–3 years later. This framework is strikingly isomorphic to modern distributed lag nonlinear models (DLNM) used in epidemiology—DLNM’s core premise is acknowledging that exposure impacts health are “nonlinear and delayed,” utilizing cross-basis functions to simultaneously characterize the exposure-response relationship and lag structure (PMID 20812303). In essence, ancients encoded long-term climatic rhythms using Heavenly Stems and Earthly Branches, while moderns encode the same rhythms using splines and lag matrices; both sides perform the same task: translating “past climate” into “future disease risk.”

The question then arises: given this logical isomorphism, can treating Wu Yun Liu Qi as a “coarse filter” (stratifying years by Gan-Zhi first, then refining with DLNM) yield genuine incremental value? This is the core hypothesis this topic aims to test.

First, consider the meteorological forecasting capability of modern models itself. A time-series DLNM established in Limpopo Province, South Africa, using weekly meteorological data and malaria cases from 1998–2015, achieved correlation coefficients exceeding 0.8 for predictions 1–2 weeks in advance, remaining above 0.7 even with a 16-week lead time (PMID 31784563)—meteorological-driven time-series models are inherently functional.

However, evidence regarding incremental value is differentiated. A 2019–2023 RSV study in Guangdong combining DLNM with XGBoost+SHAP found that the lagged RSV indicators themselves contributed most to predictive performance, with meteorological variables providing only “moderate” incremental value (PMID 42602072). A flu study in Quanzhou provided more precise figures: adding precipitation with a 2-week lag as an exogenous variable to an ARIMA model resulted in only a 2.9% improvement in RMSE and a 1.3% improvement in MAE (PMID 41200038). This indicates that meteorological information is useful, but its incremental contribution is small and easily absorbed by the historical information inherent to the disease itself.

This presents a dual warning regarding Yun Qi coarse screening. First, if the incremental value of pure meteorological variables is only single-digit percentages, the incremental value of the more indirect “Gan-Zhi stratification” may be even smaller and harder to separate from noise. Second, the maximum selling point of Yun Qi theory is its long lag of 1–3 years (“three-year epidemic transformation”), yet existing DLNM studies typically achieve satisfactory predictions using only short lags of 0–3 weeks (PMID 42602072); whether long lags truly carry independent information remains unverified under a standardized protocol.

The correct research design is not complex: use the same long infectious disease time-series dataset to run three models—pure DLNM, DLNM stratified by Yun Qi years, and a hybrid model combining Yun Qi stratification with DLNM refinement—compare them using unified metrics such as AUC, calibration, and Brier scores, and perform time-out-of-sample validation. If the hybrid model does not significantly outperform the others, then “stacking” is merely “interference”; if it does outperform, it indicates that Gan-Zhi encoding indeed captures slow-cycle information missed by DLNM splines.

Evidence and Limitations

Verified: DLNM is a mature framework for meteorological-disease modeling (PMID 20812303); meteorological-driven time-series models maintain high predictive performance within a 1–16 week window (PMID 31784563); meteorological variables offer real but limited incremental value for infectious disease prediction (PMID 41200038, PMID 42602072).

Partially Verified: Lagged disease-specific information typically contributes more than meteorological factors (PMID 42602072), suggesting the incremental space for any “climatic precursor” methodology (including Yun Qi) may be naturally constrained.

To Be Verified: “Can Gan-Zhi stratification as a DLNM coarse filter enhance prediction accuracy?”—no head-to-head comparison studies exist on the same dataset; whether long lags (1–3 years) carry information independent of short lags in meteorological-disease associations also lacks direct evidence.

Related Reading

Evidence Level: B
References: PMID 20812303, PMID 31784563, PMID 41200038, PMID 42602072


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 *