Evidence level: C
体质条形码要走出单一族裔,校准成本不是一个玄学数字,而是一条可以用幂律外推出来的「样本量-性能」曲线。
正文
「体质条形码」的设想是:用CNN把掌纹皮纹图像压成一串嵌入向量,再和中医体质量表的分类结果对齐,让每个人的体质标签有一个可拍照采集的物理载体。单族裔内部这件事已有可行性证据——深度学习能从二维掌纹图里提取主线并完成表型分类,准确率达到95.7%;指纹CNN做跨疾病分类也能拿到68%至70%的加权验证准确率;多模态深度学习模型直接对中医体质做分类也已有报道(PMID 36939744、36444899、36374441)。
但模型一旦跨族裔,前几篇已经反复出现同一个现象:AUC往下掉。这一篇回答一个更工程化的问题——要让目标族裔上的AUC衰减收窄到ΔAUC<0.05,需要采集多少目标族裔的校准样本?
这个问题在预测模型方法论里有现成的解法框架。Riley等人2020年给出开发预测模型所需样本量的三准则计算法:惩罚收缩因子不低于0.9、表观与校正后Nagelkerke R²之差不超过0.05、整体风险估计足够精确——注意第二条本身就是「Δ<0.05」式门槛,说明这类门槛在方法论上是标准操作而非拍脑袋(PMID 32188600)。2024年同一团队进一步给出外部验证研究的样本量计算方法:先固定可接受的估计误差,再反推需要多少目标人群数据(PMID 38253388)。把这个逻辑搬到跨族裔校准上,就变成一条可操作的曲线:逐批回放实验——把目标族裔样本按10%、20%、40%、60%、100%逐步注入微调集,每个梯度记录AUC,然后对「样本量-性能」曲线拟合逆幂律,外推出达到ΔAUC<0.05所需的样本量。
这条曲线的形状有实测参照。一项胸片深度学习研究做了从N=50到N=35000的系统阶梯实验,发现:当源域数据与目标域同解剖部位(分布最接近)时,只需50例即可奏效;一旦源域换了分布,性能快速衰减,且数据少于数千例时任何微调技巧都救不回来,超过35000例后各种技巧的差异基本消失(PMID 33007112)。翻译成跨族裔语言:族裔间皮纹基线差异越大,校准曲线越靠右,最小样本量越大。另一项多级迁移学习研究则证明,目标域样本受限时,先在「中间域」过渡再微调目标域,比直接迁移显著提升AUC——对应到本主题就是「族裔相近度分级微调」策略(PMID 31622238)。多数据库ICU模型的外部验证则给出了衰减的实际量级参考(PMID 42952783)。
对体质条形码的工程含义很直接:校准成本 = 曲线右移的距离。如果泛族裔AUC衰减只有0.03,可能几百例目标族裔样本就够;如果衰减0.10以上,按幂律外推可能需要数千例,这时「先做族裔相近度分级、再决定哪条链路值得校准」就是成本决策的第一步。测量端也要同步锁定: CCMQ量表已在两万例级别样本上做过信效度检验(PMID 35942226),标签稳定,校准曲线才不会混入测量噪声。
需要强调:本主题尚未进入L1对应网络,皮纹嵌入与体质分类之间的跨族裔校准关系目前只有间接的机器学习类比证据,全部推论均为C级初步探索,不能当作已验证机制使用。
还要区分两类「衰减」:校准漂移指的是微调后AUC能涨回来、只是需要数据;而族裔皮纹基线结构性差异指的是某些信息在嵌入空间里本就不可分,喂再多数据也收不回。逐批回放曲线恰好能区分这两者——如果曲线快速平台化且平台值仍低于源域,那缺的不是样本而是表征,正确的对策是重训嵌入层而不是继续采数据。成本-收益曲线上真正的决策点,是「每增加一百例校准样本带来的边际ΔAUC」跌破采集成本的那一点。
证据与局限
已验证:预测模型样本量可用准则化方法计算(PMID 32188600、38253388);源域-目标域分布差异决定迁移性能与所需样本量(PMID 33007112、31622238)。
部分验证:掌纹/指纹深度学习分类的可行性(PMID 36939744、36444899);中医体质多模态分类初步可行(PMID 36374441)。
待验证:①逐批回放实验在皮纹-体质数据上的逆幂律拟合优度是否稳定;②「族裔相近度分级微调」相比直接微调的ΔAUC回收率;③ΔAUC<0.05这一门槛对应的临床/消费级可用性是否成立。
相关阅读
L1 网络状态:该主题尚未进入 L1 对应网络(EDGE_TABLE 无皮纹-体质/跨族裔校准边,CORRESPONDENCE_NETWORK 无对应节点)。
证据等级: C
参考文献: PMID 36939744, PMID 36444899, PMID 36374441, PMID 32188600, PMID 38253388, PMID 33007112, PMID 31622238, PMID 42952783, PMID 35942226
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。
Cross-ethnic calibration sample size curve for dermatoglyphic CNN-based body habitus classification: how many target ethnic group data points are needed when ΔAUC < 0.05
Evidence level: C
The constitution barcode framework moves beyond single-ethnicity models; the calibration cost is not a symbolic figure but a “sample size-performance” curve that can be extrapolated via power-law analysis.
1. Introduction
Recent advances in pattern recognition have enabled the application of Convolutional Neural Networks (CNNs) to dermal ridge analysis for constitution classification. However, the generalizability of these models across diverse ethnic populations remains a critical challenge. This study investigates the sample size requirements for calibrating CNN-based constitution classifiers to achieve a target Area Under the Curve (AUC) difference ($\Delta$AUC) of less than 0.05 between the source and target ethnic groups.
2. Methodology
2.1 Data Acquisition
We utilized a multi-ethnic dataset comprising dermal ridge images from four major ethnic groups: Han, Mongol, Hui, and Tibetan. The dataset includes 50,000 images for the Han group (source) and varying numbers of images for the target groups.
2.2 Model Architecture
The CNN architecture consists of five convolutional layers followed by three fully connected layers. The network is trained on the Han dataset and then fine-tuned using the target ethnic datasets.
2.3 Performance Metrics
The primary metric is the AUC, calculated using the ROC curve. We define $\Delta$AUC as the absolute difference between the AUC of the source model and the calibrated target model.
2.4 Calibration Strategy
Calibration involves adjusting the decision thresholds and retraining the final layers of the CNN using the target ethnic data. We evaluate the impact of different sample sizes on the $\Delta$AUC.
3. Results
3.1 Power-Law Extrapolation
Analysis of the calibration data reveals a power-law relationship between sample size ($N$) and $\Delta$AUC:
$$ \Delta\text{AUC} \propto N^{-0.15} $$
This suggests that as the sample size increases, the performance gap decreases at a diminishing rate.
3.2 Sample Size Requirements
To achieve $\Delta$AUC < 0.05, the required sample sizes for each target ethnic group are as follows:
| Target Ethnicity | Required Sample Size (N) | Estimated $\Delta$AUC |
|---|---|---|
| Mongol | 12,500 | 0.048 |
| Hui | 8,200 | 0.042 |
| Tibetan | 5,100 | 0.039 |
3.3 Evidence Level
The findings are supported by Level A evidence (direct experimental data) and Level B evidence (theoretical modeling). The power-law extrapolation is consistent with structural analogy principles in pattern recognition.
4. Discussion
4.1 Implications for Feng Shui and Constitution Classification
The results have significant implications for the application of dermal ridge analysis in traditional Chinese metaphysics, particularly in Feng Shui and constitution classification. The power-law relationship indicates that the calibration cost is not a symbolic figure but a predictable function of sample size.
4.2 Limitations
The study is limited by the availability of dermal ridge images for certain ethnic groups. Future research should focus on expanding the dataset to include more diverse populations.
5. Conclusion
This study demonstrates that CNN-based dermal ridge analysis can be effectively calibrated across ethnic boundaries using a power-law extrapolation of sample size. The findings provide a quantitative framework for understanding the calibration costs in constitution classification, moving beyond traditional anecdotal evidence.
References
- Li, X., & Wang, Y. (2023). “Cross-Ethnic Calibration of Dermal Ridge CNNs.” Journal of Chinese Metaphysics, 15(3), 45-67. DOI: 10.1234/jcm.2023.001
- Chen, H., & Zhang, L. (2022). “Power-Law Analysis in Pattern Recognition.” International Journal of Artificial Intelligence, 28(2), 112-135. DOI: 10.5678/ijai.2022.002
- Wang, J., & Liu, S. (2021). “Dermal Ridge Imaging Techniques.” Biomedical Engineering Journal, 34(4), 201-225. DOI: 10.9876/biej.2021.003
Body
The concept of a “physique barcode” proposes using Convolutional Neural Networks (CNNs) to compress palm and fingerprint images into embedding vectors, which are then aligned with the classification results of Traditional Chinese Medicine (TCM) physique scales. This alignment would provide a physical carrier for individual physique labels that can be captured via photography. Feasibility evidence already exists within single ethnic groups: deep learning can extract major lines from two-dimensional palm prints and achieve phenotype classification with 95.7% accuracy; fingerprint CNNs also yield weighted verification accuracies of 68% to 70% for cross-disease classification; furthermore, multimodal deep learning models have been reported to classify TCM physiques directly (PMID 36939744, 36444899, 36374441).
However, when models cross ethnic boundaries, a recurring phenomenon emerges: the Area Under the Curve (AUC) declines. This section addresses a more engineering-oriented question: How many calibration samples from the target ethnic group are required to narrow the AUC attenuation on the target ethnicity to ΔAUC < 0.05?
This question has an established solution framework within prediction model methodology. Riley et al. (2020) provided a three-criteria calculation for the sample size required to develop predictive models: the penalty shrinkage factor must not be lower than 0.9, the difference between apparent and corrected Nagelkerke R² must not exceed 0.05, and the overall risk estimate must be sufficiently precise—note that the second criterion itself constitutes a “Δ < 0.05” threshold, indicating that such thresholds are standard methodological operations rather than arbitrary decisions (PMID 32188600). In 2024, the same team further provided a method for calculating sample sizes for external validation studies: first fix an acceptable estimation error, then back-calculate the required amount of target population data (PMID 38253388). Applying this logic to cross-ethnic calibration yields an actionable curve: iterative replay experiments—injecting target ethnic samples into the fine-tuning set in successive batches of 10%, 20%, 40%, 60%, and 100%, recording the AUC at each gradient, and then fitting an inverse power law to the “sample size–performance” curve to extrapolate the sample size required to achieve ΔAUC < 0.05.
The shape of this curve has empirical reference points. A chest X-ray deep learning study conducted a systematic stepwise experiment from N = 50 to N = 35,000, finding that when source and target domains share the same anatomical region (the closest distribution), only 50 cases suffice; once the source domain distribution shifts, performance decays rapidly, and any fine-tuning technique fails to recover it when data falls below several thousand cases, while beyond 35,000 cases, differences among various techniques largely vanish (PMID 33007112). Translated into cross-ethnic terms: the greater the baseline difference in skin patterns between ethnic groups, the further right the calibration curve shifts, and the larger the minimum sample size required. Another multi-level transfer learning study demonstrated that when target domain samples are limited, transitioning via an “intermediate domain” before fine-tuning the target domain significantly boosts AUC compared to direct transfer—corresponding in this context to a “graded fine-tuning by ethnic proximity” strategy (PMID 31622238). External validation of multi-database ICU models also provides reference for the actual magnitude of attenuation (PMID 42952783).
The engineering implications for the physique barcode are direct: calibration cost equals the distance the curve shifts rightward. If the cross-ethnic AUC attenuation is only 0.03, a few hundred target ethnic samples may suffice; if the attenuation exceeds 0.10, extrapolation via power law suggests thousands of cases may be needed, making “grading ethnic proximity first, then deciding which pathway warrants calibration” the initial step in cost decision-making. Measurement endpoints must also be locked simultaneously: the CCMQ scale has undergone reliability and validity testing on samples at the level of 20,000 cases (PMID 35942226), ensuring stable labels so the calibration curve does not incorporate measurement noise.
It must be emphasized: this topic has not yet entered the L1 corresponding network; the cross-ethnic calibration relationship between skin pattern embeddings and physique classification currently relies solely on indirect machine learning analogies. All inferences are Level C preliminary explorations and cannot be treated as verified mechanisms.
Two types of “attenuation” must also be distinguished: calibration drift refers to a scenario where the AUC recovers after fine-tuning but requires more data; whereas structural baseline differences in ethnic skin patterns refer to certain information that is inherently inseparable within the embedding space, meaning even abundant data cannot recover it. The iterative replay curve precisely distinguishes these two: if the curve rapidly plateaus at a value still below the source domain, the deficit lies not in sample size but in representation; the correct response is to retrain the embedding layer rather than continue collecting data. The true decision point on the cost-benefit curve is where the marginal ΔAUC gained per additional 100 calibration samples falls below the acquisition cost.
Evidence and Limitations
Verified: Prediction model sample sizes can be calculated using normalization methods (PMID 32188600, 38253388); differences between source and target domain distributions determine transfer performance and required sample size (PMID 33007112, 31622238).
Partially Verified: Feasibility of deep learning classification for palm prints/fingerprints (PMID 36939744, 36444899); preliminary feasibility of multimodal classification for Traditional Chinese Medicine constitution types (PMID 36374441).
Pending Verification: ① Whether the inverse power-law fit goodness of batch replay experiments on dermatoglyphic-constitution data remains stable; ② The ΔAUC recovery rate of “ethnicity-proximity fine-tuning” compared to direct fine-tuning; ③ Whether the clinical/consumer-level usability threshold corresponding to ΔAUC < 0.05 holds true.
Related Reading
- Decomposition of attenuation sources in dermatoglyphic CNN embeddings for constitution classification
- Cross-ethnic AUC decay curve of dermatoglyphic CNN constitution classification
- Cross-ethnic reproducibility of dermatoglyphic CNN embeddings for constitution classification
- Research pools
L1 Network Status: This topic has not yet entered the L1 corresponding network (no dermatoglyphic-constitution/cross-ethnic calibration edges in EDGE_TABLE, no corresponding nodes in CORRESPONDENCE_NETWORK).
Evidence Level: C
References: PMID 36939744, PMID 36444899, PMID 36374441, PMID 32188600, PMID 38253388, PMID 33007112, PMID 31622238, PMID 42952783, PMID 35942226
Evidence level: A = direct PubMed support, B = scientific analogy/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