Evidence level: C
表观 AUC 与交叉验证 AUC 的差值随格子样本量增大而单调收缩,两条曲线的交叉拐点就是三质联合分层格子的最小可行人数——这个数应当被实测出来,而不是沿用现成的 120 人惯例。
正文
问题出在哪:用三质分型加 HRV 特征做族裔、性别、年龄的联合分层判别时,每个格子(比如”某族裔×女性×30-45 岁”)的人数同时决定两件事:一是模型在同一批数据上算出的表观 AUC 有多乐观,二是交叉验证 AUC 的置信区间有多宽。表观 AUC 天然带着过拟合税——Riley 团队的预测模型样本量计算框架(PMID 32188600)与 TRIPOD 声明(PMID 25569120)都明确要求用内部验证修正这份乐观性;2026 年一项系统性模拟研究比较了 k 折交叉验证与 bootstrap 两类内部验证方案,证实交叉验证 AUC 是更可信的报告口径(PMID 41391734)。格子越小时,过拟合税越大、CV-AUC 越不稳,两条曲线分得越开;格子越大时,两线逐渐收敛。收敛点落在哪一档人数上,在三质×HRV 这个具体场景里从来没有被实测过——这正是本文要裁决的拐点。
裁决设计:在多中心大样本 HRV 加三质分型的数据池里,按格子样本量四档(30/50/100/200)做重抽样,每档重复不少于 200 次。每次抽样后用同一套三分类模型算两个数:表观 AUC 与 k 折交叉验证 AUC。对每档拟合两条曲线,取差值的中位数与 bootstrap 95% 置信区间。交叉拐点的判定标准预先注册:差值中位数首次低于 0.02、且 CV-AUC 置信区间宽度首次低于 0.05 的那一档,就是”联合分层格子至少需要多少人”的实证下限。再与现行 120 人下限对照:若实测拐点显著偏离 120,就把这条行业惯例改写成数据驱动的标定。
三条实施细节:第一,特征选择与调参必须全部放进交叉验证循环内部完成,防止选择泄露把 CV-AUC 抬高、把拐点人为推低;第二,每档重抽样须按族裔、性别、年龄分层进行,保持格子结构不被摊薄,否则会出现”总体达标、格子不达标”的假象;第三,若 200 人档的差值仍高于 0.02,只能报告”拐点 ≥200″的下界并把该格子规模标记为不可裁决,不许外推。
为什么用重抽样而不是解析公式:样本量计算公式(PMID 32188600、34031906)需要预先假设目标精度、预期 AUC 与事件率等参数,这些参数在三质×HRV 的每个格子里都不一样,公式只能给出笼统的区间;而重抽样直接在真实数据上重演”这个格子到底多少人时表观与交叉验证才开始合流”,把参数假设的误差整个绕开。代价是计算量大、且答案绑定在当前数据池的分布上——所以拐点结论必须同时报告数据池的构成(族裔、性别、年龄分布),否则换一个数据池拐点可能漂移。这正是权衡曲面比单一数字更有用的地方:曲面可以随数据池更新重画,惯例不能。
为什么这件事值一篇裁决:小样本预测模型的”数据饥饿”已被模拟研究证实——小样本下多种现代建模技术的判别性能被系统性高估,所需样本量远超每变量 10 个事件的老经验(PMID 25532820)。三质是多分类问题,多分类 logistic 的最小样本量计算框架(PMID 36660777)与二分类外部验证的最小样本量方法(PMID 34031906、38253388)共同说明:格子人数够不够,取决于目标精度与预期性能,不是一个固定常数。HRV 领域有现成先例——2007 年一项关于 HRV 正常值、信度与样本量估计的研究(PMID 17451377)就是把”多少人才算得稳”当作独立问题来处理。三质与 HRV 的既有证据链(Prakriti 与倾斜试验 HRV 的关联,PMID 36106418;三质与 HRV 参数的分层关联,PMID 29269119)提供了模型可用的特征池,但格子规模的标定始终缺位。把拐点测出来,跨文化分层的每一条后续结论才有统一的样本量地基。
证据与局限
已验证: 预测模型表观 AUC 的乐观性、内部验证的必要性与样本量计算框架,均有方法学共识与系统综述支撑(PMID 32188600、25569120、25532820)。
部分验证: k 折交叉验证与 bootstrap 的相对表现有 2026 年系统性模拟证据(PMID 41391734);HRV 常模的样本量估计有先例(PMID 17451377);三质×HRV 的判别关联有初步人体证据(PMID 36106418、29269119)。
待验证: 三质×HRV 场景下格子样本量-AUC 权衡曲面的具体形态与交叉拐点位置,属本文提出的研究设计提案,尚无直接实测;拐点对模型族(logistic 与树模型)、对 CV 折数的敏感性也均未测。
相关阅读
L1 网络状态:该主题尚未进入 L1 对应网络——EDGE_TABLE 与 CORRESPONDENCE_NETWORK 对格子样本量、AUC 权衡、重抽样、交叉验证均无对应边,仅 E-060 标签信度方法学背景边相邻;证据等级按无 L1 边规则封顶 C。
证据等级: C
参考文献: PMID 32188600, PMID 25569120, PMID 41391734, PMID 25532820, PMID 36660777, PMID 34031906, PMID 38253388, PMID 17451377, PMID 36106418, PMID 29269119
证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索
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。
Empirical Localization of the Crosspoint on the Grid Sample Size-AUC Trade-off Surface: Turning \”How Many Per Cell\” from Guesswork into Data-Driven Decision
Evidence level: C
The gap between apparent AUC and cross-validation AUC monotonically contracts as grid sample size increases; the intersection point of these two curves defines the minimum feasible number of individuals required for the tri-attribute stratified grid. This value must be empirically measured rather than relying on the conventional 120-person heuristic.
Body
Where the problem lies: When joint stratification by ethnicity, sex, and age is performed using the three-constitution typology plus HRV features, the sample size within each cell (e.g., “Ethnicity X × Female × 30–45 years”) simultaneously determines two quantities: (1) how optimistic the apparent AUC will be when computed on the same dataset, and (2) how wide the confidence interval of the cross-validated AUC will be. The apparent AUC inherently carries an overfitting tax—Riley’s team sample size calculation framework for prediction models (PMID 32188600) and the TRIPOD statement (PMID 25569120) both require internal validation to correct this optimism; a 2026 systematic simulation comparing k-fold cross-validation and bootstrap internal validation schemes confirmed that cross-validated AUC is the more reliable reporting metric (PMID 41391734). As cells become smaller, the overfitting tax increases and the CV-AUC confidence interval widens, causing the two curves to diverge; as cells become larger, the curves gradually converge. The specific sample size at which convergence occurs has never been empirically tested in this three-constitution × HRV scenario—this is precisely the inflection point this paper aims to resolve.
Design of the verdict: In a multi-center large-sample data pool combining HRV with three-constitution typology, we perform resampling across four tiers of cell sample sizes (30/50/100/200), repeating each tier at least 200 times. For each resample, the same three-class model computes two metrics: apparent AUC and k-fold cross-validated AUC. For each tier, we fit two curves and extract the median difference along with its bootstrap 95% confidence interval. The criterion for identifying the crossing point is pre-registered: the tier where the median difference first falls below 0.02 and the CV-AUC confidence interval width first drops below 0.05 constitutes the empirical lower bound for “the minimum number of people required per joint stratification cell.” This is then compared against the current industry convention of 120 people per cell; if the observed inflection point significantly deviates from 120, this longstanding convention is revised into a data-driven calibration.
Three implementation details: First, feature selection and hyperparameter tuning must be fully embedded within the cross-validation loop to prevent selection leakage from inflating CV-AUC and artificially lowering the inflection point. Second, each tier’s resampling must be stratified by ethnicity, sex, and age to preserve the cell structure; otherwise, averaging across heterogeneous groups creates a false impression of “overall adequacy despite inadequate cells.” Third, if the difference at the 200-person tier remains above 0.02, we can only report the lower bound “inflection point ≥ 200” and mark that cell size as unresolved, prohibiting extrapolation.
Why resampling rather than analytical formulas: Sample size calculation formulas (PMID 32188600, 34031906) require pre-specified assumptions about target precision, expected AUC, and event rate—parameters that vary across every cell in the three-constitution × HRV space. Consequently, formulas yield only broad intervals. Resampling directly re-enacts on real data the question “at what sample size do apparent and cross-validated AUCs begin to converge,” thereby bypassing all parameter assumption errors. The trade-off is higher computational cost and answers tied to the current data pool’s distribution; thus, the inflection point conclusion must simultaneously report the data pool composition (ethnicity, sex, age distribution), otherwise shifting to a different data pool may shift the inflection point. This illustrates why the convergence surface is more informative than a single number: the surface can be redrawn as the data pool updates, whereas conventions cannot.
Why this warrants a formal verdict: The “data hunger” of small-sample prediction models has been confirmed by simulation studies—under small samples, the discriminative performance of various modern modeling techniques is systematically overestimated, requiring far more events per variable than the traditional rule of thumb (PMID 25532820). Since three-constitution is a multiclass problem, the minimum sample size framework for multiclass logistic regression (PMID 36660777) and the minimum sample size method for external validation of binary classifiers (PMID 34031906, 38253388) together demonstrate that whether a cell has sufficient sample size depends on target precision and expected performance, not on a fixed constant. The HRV domain already has precedents: a 2007 study on HRV normal ranges, reliability, and sample size estimation (PMID 17451377) treats “how many people are needed for stable estimates” as an independent question. Existing evidence linking Prakriti with tilt-test HRV (PMID 36106418) and stratified associations between three-constitution and HRV parameters (PMID 29269119) provides the feature pool for modeling, yet calibration of cell size remains missing. Determining the inflection point establishes a unified sample size foundation for all subsequent cross-cultural stratification conclusions.
Evidence and Limitations
Verified: Optimism in apparent AUC of prediction models, the necessity of internal validation, and frameworks for sample size calculation all have methodological consensus and systematic review support (PMID 32188600, 25569120, 25532820).
Partially Verified: Relative performance of k-fold cross-validation and bootstrap has 2026 systematic simulation evidence (PMID 41391734); sample size estimation for HRV norms has precedents (PMID 17451377); discriminative associations of Triple Quality × HRV have preliminary human evidence (PMID 36106418, 29269119).
To Be Verified: The specific shape of the lattice sample size-AUC trade-off surface and the location of inflection points under Triple Quality × HRV scenarios belong to the research design proposal presented here and lack direct measurement; sensitivity of inflection points to model families (logistic and tree models) and to CV fold numbers also remains untested.
Related Reading
- Three Substances ΔAUC Head-to-Head Minimum Detectable Effect and Sample Size Feasibility Boundaries
- Directional Encoding “Binning Granularity” Optimal Scanning: Robustness vs. Granularity Trade-off Curve
- Cross-Cultural Frequency Cognition Research Pool
L1 Network Status: This topic has not yet entered the corresponding L1 network—there are no edges linking EDGE_TABLE and CORRESPONDENCE_NETWORK to grid sample sizes, AUC trade-offs, resampling, or cross-validation. Only the E-060 label reliability methodology background edge is adjacent; evidence level is capped at C per the “no L1 edge” rule.
Evidence Level: C
References: PMID 32188600, PMID 25532820, PMID 36660777, PMID 34031906, PMID 38253388, PMID 17451377, PMID 36106418, PMID 29269119, PMID 36660777, PMID 34031906
Evidence level: A = direct PubMed support; B = scientific analogy or indirect evidence; C = traditional consensus / preliminary exploration.
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