纯DLNM vs 运气粗筛+DLNM:三模型头对头比较怎么设计才公平?

Written by

in

Language

Evidence level: B

判断”运气粗筛能不能提升传染病预测”,唯一公平的办法是在同一数据集上让”纯DLNM、运气粗筛+DLNM、运气纯分层”三个模型用统一指标做时间外验证——而这个实验至今还没有人做过。

正文

五运六气的现代研究者常常抛出一个诱人的说法:把天干地支当作”粗筛器”,先按运气把年份分层,再交给现代分布式滞后非线性模型(DLNM)精算,预测精度会更高。这句话听起来合理,但它在逻辑上是一个可证伪的增量价值假说——”粗筛+精算”是否真的优于”纯精算”,必须用同一数据集、同一验证协议、同一评价指标,把模型摆在一起比过才知道。可惜的是,尽管 DLNM 在中国传染病研究里已经被广泛使用,这个”三模型头对头”实验至今仍是空白。

先看清三个候选模型长什么样。模型一(纯DLNM):只用气象变量,用交叉基函数(cross-basis)同时刻画暴露-反应的非线性关系和滞后结构(PMID 20812303),这是当前环境流行病学的金标准框架,江苏手足口病(PMID 38562664)、西南农村恙虫病(PMID 35937262)、桐城其他感染性腹泻(PMID 36808040)都用它建立了可靠的暴露-滞后-反应关系。模型二(运气粗筛+DLNM):先按干支年份或六气时段把数据分层,再在每层内拟合 DLNM——这正是张轩团队在北京做过的思路:把结核病发病与当年及前 1-3 年的气象因子关联,发现”三年前至今的多因子组合优于单年数据”(PMID 26525546)。模型三(运气纯分层):不建模气象,纯粹按运气分组描述发病差异,是最弱的对照。

难点在于怎么”公平地比”。这里至少有四个坎。第一是验证协议:不能在训练集上自说自话,必须用滚动时间外验证(expanding-window time series cross-validation)——新加坡登革热预测团队用 2/4/6/8 周前瞻窗口做时间外验证,同时报告 Brier 评分和 ROC-AUC,就是教科书级的做法(PMID 41274877)。第二是指标选择:不能只看 AUC 一个数,AUC 反映排序能力但不反映概率校准,Brier 评分(预测概率与真实结果的均方误差)能惩罚”自信地错”的模型,校准曲线则直接暴露系统性的过/低估。第三是多重比较陷阱:三个模型、多个滞后窗、多个领先时间,比得越多越容易撞上假阳性,必须做多重检验校正。第四是运气分层的先验泄露:干支是固定不变的历法标签,用它分组不会泄露未来信息,这点比很多 ML 特征干净,但如果分层维度(岁运、客气、主气)选得太多,等于变相过拟合,需要用 FDR 控制。

这个设计的价值不止于”谁的 AUC 高”。它真正回答的问题是:运气提供的长滞后信息(1-3年,”三年化疫”),到底有没有独立于短滞后气象(0-3周)的预测力? 北京流感 35 年回顾研究显示,包含前三年的气象数据模型预测精度最高(约49%),暗示长滞后确实携带信息(PMID 26525546 同团队工作;流感六气高发集中在六之气,与冬季干冷吻合)。但纯气象 DLNM 用 0-3 周短滞后也能达到可观预测力(PMID 38562664 极低温滞后1-2天即起效)。如果三模型实验显示”运气粗筛+DLNM”与”纯DLNM”差异不显著,那”粗筛提升精度”就是装饰性的;如果显著且稳定,运气理论的长滞后价值就第一次拿到了现代预测框架内的直接证据。

对中国用户还有一个现实意义:低成本粗筛如果有效,可以服务资源有限的基层预警。概率型预警模型(PMID 40769217)已经证明,从”0/1报警”升级为”连续概率+分层响应”能显著减少误报、提高预警实用性——如果运气粗筛能在同等数据量下达到接近纯DLNM的预测力,那么它在数据匮乏地区就具有真正的部署价值。

这个比较实验的设计成本并不高:一套疾病日报 + 一套气象日报 + 一个公开的干支历法表即可启动。难点全在纪律——统一时间外验证窗口、统一报告 Brier 与校准曲线、预先登记要比较的模型清单和多重比较校正方案,避免”事后挑出最好看的结果”。只要守住这套协议,无论结果支持还是否定运气粗筛,产出的都是可复用的方法学模板,这正是当前运气医学研究中稀缺的东西。

证据与局限

已验证: DLNM 是刻画气象-传染病暴露滞后关系的成熟框架(PMID 20812303),在中国多病种上可复现(PMID 38562664、35937262、36808040);运气-气象-疾病关联有北京结核病 15 年回顾支撑(PMID 26525546);概率型预警+Brier/时间外验证方法学已成熟(PMID 41274877、40769217)。

部分验证: “运气粗筛+DLNM”分层思路在北京结核病(PMID 26525546)中初步可行,但样本单中心、指标仅相关系数/准确率,未做同数据集三模型对照。

待验证: 同一传染病数据集上”纯DLNM vs 运气粗筛+DLNM vs 运气纯分层”的滚动时间外三模型头对头比较(AUC/校准度/Brier),及其对长滞后独立信息量的检验——全球尚无此实验。

相关阅读

证据等级: B
参考文献: PMID 20812303, PMID 38562664, PMID 35937262, PMID 36808040, PMID 26525546, PMID 41274877, PMID 40769217


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。

Pure DLNM vs. Luck Screening plus DLNM: How to Design a Fair Head-to-Head Comparison Across Three Models

Evidence level: B

The only fair way to judge whether “luck screening” enhances infectious disease prediction is to compare three models—”Pure DLNM,” “Luck Screening + DLNM,” and “Luck Pure Stratification”—on the same dataset using unified metrics for time-out-of-sample validation. To date, no one has conducted this experiment.

Main Text

Modern researchers in Five Movements and Six Qi often propose an appealing claim: treat the Heavenly Stems and Earthly Branches as a “coarse filter,” stratify years by luck, and then refine predictions using Distributed Lag Nonlinear Models (DLNM), resulting in higher accuracy. While this sounds reasonable, it is logically a falsifiable hypothesis of incremental value—whether “coarse filtering + refinement” truly outperforms “pure refinement” must be determined by pitting the models against each other on the same dataset, using the same validation protocol and evaluation metrics. Unfortunately, despite DLNM’s widespread use in Chinese infectious disease research, this “three-model head-to-head” experiment remains a gap.

First, clarify the three candidate models. Model 1 (Pure DLNM): Uses only meteorological variables, employing cross-basis functions to simultaneously characterize the nonlinear exposure-response relationship and lag structure (PMID 20812303). This represents the gold standard framework in environmental epidemiology; it has established reliable exposure-lag-response relationships for Hand, Foot, and Mouth Disease in Jiangsu (PMID 38562664), Scrub Typhus in Southwest rural areas (PMID 35937262), and other infectious diarrheas in Tongcheng (PMID 36808040). Model 2 (Luck Screening + DLNM): Stratifies data by Heavenly-Stem/Earthly-Branch years or Six Qi periods, then fits DLNM within each stratum. This aligns with the approach taken by Zhang Xuan’s team in Beijing, who associated tuberculosis incidence with meteorological factors from the current year and the preceding 1–3 years, finding that “multi-factor combinations spanning three years to present outperform single-year data” (PMID 26525546). Model 3 (Luck Pure Stratification): Does not model meteorology; purely describes incidence differences by grouping according to luck. This serves as the weakest control.

The challenge lies in conducting a “fair comparison.” At least four hurdles exist. First is the validation protocol: self-validation on the training set is unacceptable; expanding-window time series cross-validation must be used, as demonstrated by the Singapore Dengue prediction team using 2/4/6/8-week forward-looking windows while reporting Brier scores and ROC-AUC, representing a textbook approach (PMID 41274877). Second is metric selection: relying solely on AUC is insufficient, as AUC reflects ranking ability but not probability calibration. The Brier score (mean squared error between predicted probabilities and true outcomes) penalizes models that are confidently wrong, while calibration curves directly expose systematic over- or under-estimation. Third is the multiple comparisons trap: comparing three models across multiple lag windows and lead times increases the risk of false positives; therefore, multiple testing correction is mandatory. Fourth is prior leakage from luck stratification: Heavenly Stems and Earthly Branches are fixed calendar labels; using them for grouping does not leak future information, making them cleaner than many machine learning features. However, if too many stratification dimensions (e.g., Age Cycle, Guest Qi, Host Qi) are selected, it amounts to变相 overfitting, requiring False Discovery Rate (FDR) control.

The value of this design extends beyond “who has the higher AUC.” It addresses a fundamental question: Does the long-lag information provided by luck (1–3 years, “three-year epidemic transformation”) possess predictive power independent of short-lag meteorology (0–3 weeks)? A 35-year retrospective study on Beijing influenza shows that models incorporating meteorological data from the previous three years achieve the highest prediction accuracy (~49%), suggesting long lags carry information (same team’s work in PMID 26525546; Six Qi peaks align with winter dry-cold conditions). Yet, pure meteorological DLNM also achieves considerable predictive power using short lags of 0–3 weeks (PMID 38562664 shows extreme cold effects within 1–2 days). If the three-model experiment shows no significant difference between “Luck Screening + DLNM” and “Pure DLNM,” then “screening improves accuracy” is merely decorative. If the difference is significant and stable, luck theory’s long-lag value will finally receive direct evidence within a modern predictive framework.

For Chinese users, there is also practical significance: If low-cost coarse screening is effective, it can serve resource-limited grassroots warning systems. Probabilistic warning models (PMID 40769217) have demonstrated that upgrading from “0/1 alerts” to “continuous probability + stratified response” significantly reduces false alarms and enhances warning utility. If luck screening achieves predictive power comparable to pure DLNM with equivalent data volume, it holds genuine deployment value in data-scarce regions.

The design cost of this comparison experiment is low: one disease daily report, one meteorological daily report, and a public Heavenly-Stem/Earthly-Branch calendar are sufficient to launch it. The difficulty lies entirely in discipline—maintaining unified time-out-of-sample windows, standardizing Brier and calibration curve reporting, and pre-registering the model comparison list and multiple testing correction scheme to avoid “post-hoc cherry-picking of the most favorable results.” Adhering to this protocol ensures that regardless of whether luck screening is supported or refuted, the output is a reusable methodological template—a scarce resource in current luck medicine research.

Evidence and Limitations

Verified: DLNM is a mature framework for characterizing meteorological-infectious disease exposure lag relationships (PMID 20812303) and is reproducible across multiple diseases in China (PMID 38562664, 35937262, 36808040); associations between luck, meteorology, and disease are supported by a 15-year Beijing tuberculosis retrospective study (PMID 26525546); probabilistic warning methods combined with Brier scores and time-out-of-sample validation are established (PMID 41274877, 40769217).

Partially Verified: The stratification approach of “Luck Screening + DLNM” was preliminarily feasible in the Beijing tuberculosis study (PMID 26525546), but the sample was single-center, metrics relied only on correlation coefficients/accuracy, and no head-to-head comparison on the same dataset was conducted.

To Be Verified: Rolling time-out-of-sample head-to-head comparison of “Pure DLNM vs. Luck Screening + DLNM vs. Luck Pure Stratification” on the same infectious disease dataset (evaluating AUC/calibration/Brier) and its testing of independent information content in long lags—no such experiment exists globally.

Related Reading

Evidence Level: B
References: PMID 20812303, PMID 38562664, PMID 35937262, PMID 36808040, PMID 26525546, PMID 41274877, PMID 40769217


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 *