Evidence level: B
答”加得上”需要三件事同时成立:CCMQ 严重度分层(轻/中/重)稳定 + 严重度在菌群指纹之上仍有非零 ΔAUC/NRI + 增量方向是真实生理信号不是分型代理。
一、问题:评分是装饰还是真信息
中医体质学一直默认一个直觉——痰湿评分越高、阳虚评分越深,体质”偏”得越厉害,干预就该越用力、起效该越明显。但当我们要做”按应答分层的菌群干预 RCT”时,摆在工程团队面前的问题不是”严重度重要不重要”,而是更具体的工程题:
- CCMQ 严重度分层的可信度有多高? 同一个轻/中/重判读,不同医生、不同时间点,分布有多稳?
- 在已经拿到基线 16S/宏基因组菌群指纹的前提下,把 CCMQ 严重度塞进应答预测模型,ΔAUC/NRI 是显著大于零还是塞进去等于没塞?
- 就算 ΔAUC 看起来是显著的正,方向对不对?——会不会只是”严重度高的本来基线菌群就更糟”,模型只是把一个分型代理当作菌群代理在用?
二、严重度梯度的实证底子有多厚
要回答这三件事,先看严重度这个变量在数据里到底有什么”剂量-反应”信号。
在痰湿质这条线,最直接的一组证据是 Luminex 平台的多因子检测(PMID 24063213):101 例 20-50 岁人群按 BMI × 体质两轴四组(痰湿肥胖 / 痰湿非肥胖 / 单纯肥胖 / 平和质),结果 PDC 评分与 TNF-α、IL-6、MCP-1 三种炎症因子水平全部显著正相关(全部 P<0.05)。这个相关性是建立在同一批受试者内部不同评分之间的,不是横断面”痰湿 vs 平和”二分,所以可以算作”严重度梯度”在炎症指纹上的剂量证据。
在虚性体质这条线,2024 年 Front Nutr 多组学整合(PMID 39360281)则给出了”严重度-菌群-贫血倾向”的方向链:气虚/阳虚/阴虚合并的虚性体质人群 Hb 显著低于平和质;菌群上 Firmicutes 升高、Bacteroidetes 降低、α 多样性升高;菌群标志物 R. gnavus 既是虚性体质指示菌,也是贫血倾向的共享菌基础,F. plautii 反向是保护菌。这条链的关键在于”虚性体质的 Hb 偏低”不是二分结论(是/否虚性),而是随评分连续变化的可测量指标。
在阳虚质这条线,2022 年 YADC 健康人群研究(PMID 35294128)则展示了严重度-健康度的”双向平衡”:即使是健康的 YADC 个体,有害菌(Prevotellaceae, LDA>4.0, P=0.0141)和有益菌(Ruminococcaceae/Faecalibacterium, LDA>4.0, P=0.0025/0.048)同步升高,处于”动态稳态”。这条证据说明:评分高未必等于病——但评分高的人菌群已经在做”补偿性调动”,干预后应答基线就跟平和质不一样。
这三组合起来的意思是:严重度梯度在生化层、菌群层、临床表型层都看得到剂量-反应信号;评分不是 0/1 变量,是 1-5 分(CCMQ 子量表每题分值)的连续变量。
三、问题变成了一个 NRI/IDI 工程题
把”严重度梯度”接入到应答预测里,标准方法是用 NRI(净重分类改善)和 IDI(整体判别改善)来回答”加了之后预测力加了多少”。但这条路上有已经反复被点名的坑——Hilden & Gerds 2013(PMID 23553436)的模拟研究证明:在没有先校准两个模型的前提下,NRI/IDI 可能在”零信息增量”的情况下报告出虚高的”改善”——即所谓”造假的最优模型”现象。他们的核心警告是:NRI/IDI 不是 proper scoring rule,必须配合 Brier 分数和 AUC 一并报告,单独看会失真。
这一条放在中医体质场景里特别要紧,因为 CCMQ 评分本身就是有信度上限的(重测相关系数 0.76-0.90,PMID 35294132 全国 21,948 例短表验证),如果不分层建模、不做交叉验证,单看 ΔAUC 很容易把”评分能分”误读为”评分能预测”。所以裁决设计的第一步是:
- 把应答模型拆成三档——基线菌群指纹单模 / CCMQ 严重度单模 / 菌群 + 严重度联合;
- 三档都先在训练集内做 10-fold 交叉验证校准(用 calibration-in-the-large 和 calibration slope 检验),不校准不准算 NRI/IDI(呼应 PMID 23553436);
- 用 ΔAUC、NRI(>0)、IDI、净效益(decision curve analysis, NB)四个维度并报,bootstrap 1000 次取 CI;
- 同时报告 SHAP 方向——严重度项进入模型后,特征重要性方向是否与已知生物机制一致(例如痰湿评分高→SCFA 产菌耗竭→代谢干预起效更慢)。
四、可能出现的四种裁决结局
把上述工程框架套到痰湿/阳虚/虚性三体质 + 益生菌/FMT/中药三种干预上,可能跑出四种结局:
结局 A:ΔAUC 显著(>0.05)且 NRI(>0) 显著且方向一致——这种情况说明严重度是真互补信息,应当入模。代价是要为评分采集增加预算,但边际预测力值。
结局 B:ΔAUC 显著但 SHAP 方向与已知生理信号相反——这是个红旗:很可能严重度只是菌群指纹的代理(严重度高的人本来菌群就更糟,所以评分加进去”看着有用”,但因果上还是菌群在工作)。这种情况下应当用 directed acyclic graph(DAG)配合中介分析检验——若严重度→菌群→应答的间接路径在去掉菌群后严重度→应答的直接路径不显著,则结论降级为”严重度是冗余特征”。
结局 C:ΔAUC 不显著(NRI(>0) < 0.1)——说明严重度评分在菌群指纹之上几乎没有边际贡献,可以不入模。这是工程上最舒服的结局(少采集一项就少一份数据负担),但要警惕样本量不足导致的功效不够——按 PMID 23553436 的方法学,二分类应答终点下若想检出 NRI(>0)=0.1,至少要 400-600 例受试者;少于这个量,结论”不显著”不可信。
结局 D:ΔAUC 反向(评分进入模型后预测力变差)——这是最容易被忽略但也最有信息量的情况,通常意味着:评分与菌群存在多重共线性、模型对评分过拟合、或者评分本身在不同严重度层之间没有线性结构(轻度和中度的应答特征可能差异不大,但中度和重度之间有跳跃)。需要改用分层模型或阈值化评分。
五、配套的菌群端要点
不管裁决落点是 A/B/C/D 哪一种,有两条菌群侧的工程细节必须并行做:
- 黏膜 vs 粪便不要混用:Zmora 2018(PMID 30193112)已经证明益生菌/安慰剂的定植抗性主要看黏膜相关菌群,粪便宏标识常常看不出来。严重度梯度如果是作用在黏膜层,粪便采样会低估增量。
- 应答标志物的方向需要预锁:参考 PMID 36250371 在 IBS-D 队列里 F. saccharivorans(应答菌)/ E. faecalis + L. lactis(非应答菌)的方向结论,痰湿质场景下的应答菌方向不能直接外推——之前 #696 已裁决这是”跨体系迁移”问题,本议题里则作为”严重度层是否改变方向”的子问题存在。
六、对中医体质学的工程含义
如果跑出来是 A 或 D 的某些情况,临床意义会反作用于体质学本身:
- 若严重度确实是互补信息,那 CCMQ 就不只是分型工具,而是”剂量-反应”工具,干预方案应当按严重度分档(轻/中/重各有不同菌群剂量、疗程、随访密度),这与”按体质选方”是两条不同的工程问题。
- 若严重度只是冗余(结局 B/C),则工程上可以节省评分采集成本,但体质学就退回到”分型”层面(你是哪一型 → 选哪类干预),失去”严重度”这一维。Wang 2013 综述(PMID 23548117)已经把痰湿质”高分者代谢综合征风险高”作为体质学核心论点之一——这个论点本身仍然成立,但工程上”高分→应答强”的桥就需要重新论证。
证据与局限
已验证:
– 痰湿评分与 TNF-α/IL-6/MCP-1 显著正相关(PMID 24063213,n=101,Luminex 多因子)
– 虚性体质 Hb 偏低 + F/B 比例变化 + R. gnavus 富集(PMID 39360281,多组学)
– 阳虚质健康人群已存在”有害+有益”双向菌群调动(PMID 35294128)
– NRI/IDI 必须先校准再算,否则可能”造假”(PMID 23553436 模拟+癌症实证)
– 益生菌定植抗性主要看黏膜层,粪便宏标识会漏(PMID 30193112)
– CCMQ 重测相关系数 0.76-0.90,8 周内可重复性可接受(PMID 35294132,n=21,948)
部分验证:
– 痰湿质肥胖菌群诊断 AUC 0.80-0.83(PMID 35719350,n=90)—— 菌群端有诊断力,但严重度梯度是否带来 ΔAUC>0 仍待 RCT 验证
– 痰湿质 F. plautii 耗竭 + 植物鞘氨醇→肝 PPARα 通路(PMID 40097405,n=209,Cell Discovery 2025)—— 通路层证据充分,但”严重度↑→F. plautii↓→起效慢”的剂量链在干预场景未实测
待验证:
– 严重度梯度(轻/中/重)在应答模型中真实的 ΔAUC/NRI(>0)/IDI/NB 分布,缺直接 RCT(n≥400-600 才能稳定检出 NRI 0.1)
– 严重度项进入模型后的 SHAP 方向是否与已知生物机制一致
– 评分-菌群共线性下,分层 vs 阈值化 vs 联合建模哪种工程化方案最优
相关阅读
证据等级: B(菌群+炎症层有直接证据;增量预测力尚缺 RCT 标定)
参考文献: PMID 24063213, PMID 39360281, PMID 35294128, PMID 35719350, PMID 40097405, PMID 23553436, PMID 30193112, PMID 35294132
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 severity gradient of constitutionality add value in response prediction? Is the CCMQ score redundant or complementary?
Evidence level: B
To affirm “adds value,” three conditions must simultaneously hold: stable CCMQ severity stratification (mild/moderate/severe) + non-zero ΔAUC/NRI above baseline microbiome fingerprints + incremental direction reflects true physiological signals rather than proxy stratification.
I. Question: Is Scoring Decorative or Genuine Information
Traditional Yijing Studies often default to an intuitive assumption—that higher Phlegm-Dampness scores or deeper Yang Deficiency ratings imply a more “biased” constitution, necessitating stronger interventions and presumably clearer effects. However, when engineering teams design “stratified microbiome intervention RCTs based on response,” the engineering challenge is not whether severity matters, but more specific technical questions:
- What is the reliability of CCMQ severity stratification? How stable are the mild/moderate/severe readings across different practitioners and time points?
- Given baseline 16S/metagenomic fingerprints are already available, does inserting CCMQ severity into a response prediction model yield a significant ΔAUC/NRI, or is it equivalent to adding nothing?
- Even if ΔAUC appears significantly positive, is the direction correct?—Could it merely be that “higher severity inherently correlates with worse baseline microbiomes,” causing the model to treat a stratification proxy as a microbial proxy?
II. How Thick Is the Empirical Foundation for Severity Gradients
To answer these three questions, we first examine what “dose-response” signals the severity variable holds within the data.
Along the Phlegm-Dampness line, the most direct evidence comes from multi-factor detection on the Luminex platform (PMID 24063213): In a cohort of 101 adults aged 20–50, stratified by BMI × constitution into four groups (Phlegm-Dampness Obese / Phlegm-Dampness Non-obese / Simple Obesity / Harmonious Constitution), PDC scores showed significant positive correlations with TNF-α, IL-6, and MCP-1 levels (all P<0.05). This correlation is established within the same subjects across different scores, not as a cross-sectional “Phlegm-Dampness vs. Harmonious” binary; thus, it constitutes dose-level evidence for severity gradients on inflammatory fingerprints.
Along the Deficiency line, a 2024 Front Nutr multi-omics integration (PMID 39360281) provides a directional chain linking “severity-microbiome-anemia tendency”: Individuals with combined Qi/Yang/Yin deficiency showed significantly lower Hb than Harmonious types; microbiome-wise, Firmicutes increased, Bacteroidetes decreased, and α-diversity increased; R. gnavus serves as both an indicator for deficiency and a shared basis for anemia tendency, while F. plautii acts as a protective strain. The critical point here is that “low Hb in deficiency constitutions” is not a binary conclusion (yes/no) but a measurable metric varying continuously with scores.
Along the Yang Deficiency line, a 2022 YADC healthy population study (PMID 35294128) demonstrates a “two-way balance” between severity and health: Even in healthy YADC individuals, harmful bacteria (Prevotellaceae, LDA>4.0, P=0.0141) and beneficial bacteria (Ruminococcaceae/Faecalibacterium, LDA>4.0, P=0.0025/0.048) simultaneously increased, indicating a “dynamic steady state.” This evidence suggests: High scores do not necessarily equal disease—but high scorers’ microbiomes are already engaging in “compensatory mobilization,” meaning response baselines differ from Harmonious types.
Taken together, these combinations indicate that severity gradients exhibit dose-response signals across biochemical, microbiome, and clinical phenotype layers. Scoring is not a 0/1 variable but a continuous one (1–5 points per CCMQ subscale item).
III. The Question Becomes an NRI/IDI Engineering Problem
Integrating “severity gradients” into response prediction standards utilize Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) to quantify “how much predictive power increases.” However, pitfalls have been repeatedly flagged—Hilden & Gerds 2013 (PMID 23553436) demonstrated via simulation that without prior calibration of both models, NRI/IDI may report inflated “improvement” under “zero information increment”—a so-called “optimal fake model” phenomenon. Their core warning is: NRI/IDI are not proper scoring rules; they must be reported alongside Brier scores and AUC. Viewing them alone leads to distortion.
This is particularly critical in the TCM constitution context because CCMQ scores inherently have reliability ceilings (retest correlation 0.76–0.90, PMID 35294132 national validation of 21,948 short-form cases). Without stratified modeling or cross-validation, observing ΔAUC alone risks misreading “scoring can differentiate” as “scoring can predict.” Therefore, the first step in adjudication design is:
- Split response models into three tiers—Baseline Microbiome Fingerprint Single Model / CCMQ Severity Single Model / Microbiome + Severity Joint Model;
- Calibrate all three tiers within the training set using 10-fold cross-validation (testing calibration-in-the-large and calibration slope); do not calculate NRI/IDI without calibration (aligning with PMID 23553436);
- Report across four dimensions: ΔAUC, NRI (>0), IDI, and Net Benefit (decision curve analysis, NB), using bootstrap 1000 iterations for confidence intervals;
- Simultaneously report SHAP directions—after entering severity terms, do feature importance directions align with known biological mechanisms (e.g., high Phlegm-Dampness score → depletion of SCFA-producing bacteria → slower metabolic intervention efficacy)?
IV. Four Possible Adjudication Outcomes
Applying the above engineering framework to Phlegm-Dampness/Yang Deficiency/Deficiency constitutions across three interventions (probiotics/FMT/herbal medicine), four outcomes may emerge:
Outcome A: Significant ΔAUC (>0.05) with significant NRI (>0) and consistent direction — This indicates severity provides genuine complementary information and should be included in the model. The cost is increased budget for score collection, but the marginal predictive gain justifies it.
Outcome B: Significant ΔAUC but SHAP direction opposes known physiological signals — This is a red flag: Severity likely acts merely as a proxy for microbiome fingerprints (higher severity inherently correlates with worse microbiomes, so adding scores “looks useful,” but causality remains microbial). In such cases, use a Directed Acyclic Graph (DAG) with mediation analysis. If the indirect path Severity → Microbiome → Response is significant, but the direct path Severity → Response becomes non-significant after removing microbiome, downgrade the conclusion to “severity is a redundant feature.”
Outcome C: Non-significant ΔAUC (NRI (>0) < 0.1) — This suggests severity scores contribute little marginal value above microbiome fingerprints, allowing exclusion from the model. This is the most comfortable engineering outcome (less data burden), but caution against insufficient sample size leading to low power—per PMID 23553436 methodology, detecting NRI (>0)=0.1 in binary response endpoints requires 400–600 subjects; below this, “non-significant” conclusions are unreliable.
Outcome D: Reverse ΔAUC (predictive power worsens after adding severity) — This is often overlooked yet highly informative, typically implying: multicollinearity between score and microbiome, overfitting to scores by the model, or lack of linear structure in severity itself (response characteristics between mild/moderate may differ little, but moderate/severe show jumps). Consider stratified models or thresholded scoring instead.
V. Key Engineering Details on the Microbiome Side
Regardless of whether the adjudication lands on A/B/C/D, two microbiome-side engineering details must run in parallel:
- Do not conflate Mucosa vs. Stool: Zmora 2018 (PMID 30193112) proved probiotic/placebo colonization resistance primarily depends on mucosa-associated microbiomes; stool metagenomics often fails to detect this. If severity gradients act at the mucosal layer, stool sampling may underestimate incremental gains.
- Pre-lock Response Marker Direction: Refer to PMID 36250371 in IBS-D cohorts regarding F. saccharivorans (response strain) vs. E. faecalis + L. lactis (non-response strains). Directional conclusions in Phlegm-Dampness contexts cannot be directly extrapolated—previous #696 adjudicated this as a “cross-system transfer” issue; here, it exists as a sub-question of whether severity layers alter direction.
VI. Engineering Implications for TCM Constitution Studies
If outcomes A or D (or partials thereof) emerge, clinical implications will feedback into constitution studies themselves:
- If severity is indeed complementary, CCMQ transcends mere typing tools to become a “dose-response” instrument. Intervention protocols should stratify by severity (light/moderate/severe with distinct microbial dosages, durations, and follow-up densities), representing a different engineering problem than “selecting formulas by constitution.”
- If severity is merely redundant (Outcomes B/C), engineering costs for score collection can be saved, but constitution studies retreat to “typing” only (Which type are you → Which intervention class). The dimension of “severity” is lost. Wang 2013 review (PMID 23548117) already posits “high scorers have higher metabolic syndrome risk” as a core TCM constitution argument—this argument itself remains valid, but the engineering bridge “high score → stronger response” requires re-argumentation.
Evidence and Limitations
Verified:
– Phlegm-Dampness scores significantly positively correlated with TNF-α/IL-6/MCP-1 (PMID 24063213, n=101, Luminex multi-factor)
– Deficiency constitution shows lower Hb + F/B ratio shifts + R. gnavus enrichment (PMID 39360281, multi-omics)
– Healthy Yang Deficiency individuals already exhibit “harmful+benevolent” two-way microbiome mobilization (PMID 35294128)
– NRI/IDI require prior calibration before calculation; otherwise may “fake” results (PMID 23553436 simulation + cancer empirical)
– Probiotic colonization resistance primarily observed at mucosal layer; stool metagenomics misses it (PMID 30193112)
– CCMQ retest correlation 0.76–0.90; repeatability within 8 weeks acceptable (PMID 35294132, n=21,948)
Partially Verified:
– Phlegm-Dampness obesity microbiome diagnostic AUC 0.80–0.83 (PMID 35719350, n=90) — Microbiome side shows diagnostic power, but whether severity gradients yield ΔAUC>0 awaits RCT validation
– Phlegm-Dampness F. plautii depletion + phytosphingosine → hepatic PPARα pathway (PMID 40097405, n=209, Cell Discovery 2025) — Pathway-level evidence robust, but “severity↑→F. plautii↓→slower efficacy” dose chain unmeasured in intervention scenarios
To Be Verified:
– True ΔAUC/NRI (>0)/IDI/NB distribution of severity gradients (mild/moderate/severe) in response models; lacks direct RCT (n≥400–600 needed for stable NRI 0.1 detection)
– Whether SHAP directions post-severity entry align with known biological mechanisms
– Under score-microbiome collinearity, which engineering solution is optimal: stratified vs. thresholded vs. joint modeling
Related Reading
- Adjudication on “Cross-system Transfer” of C. aerofaciens Response Markers
- Adjudication on “Conflict Zone” Between Constitution Questionnaires and Microbiome Sequencing
- Adjudication on “Cross-constitution Transfer” of Constitution Response Prediction Models
- Research Pool ④: Environment-Life Interface
Evidence Level: B (Direct evidence exists at microbiome+inflammatory layers; incremental predictive power lacks RCT calibration)
References: PMID 24063213, PMID 39360281, PMID 35294128, PMID 35719350, PMID 40097405, PMID 23553436, PMID 30193112, PMID 35294132
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.
Leave a Reply