Vallée Alexandre, Feki Anis, Moawad Gaby, Ayoubi Jean-Marc
Department of Epidemiology and Public Health, Foch hospital, Suresnes, France.
Department of Gynecology and Obstetrics, University Hospital of Fribourg, Fribourg, Switzerland.
Comput Struct Biotechnol J. 2025 Aug 14;27:3654-3662. doi: 10.1016/j.csbj.2025.08.013. eCollection 2025.
The dynamic interplay of ovarian hormones is central to reproductive physiology, yet the complexity of their cyclic variations poses challenges for analysis, simulation, and teaching. This study presents a framework for generating physiologically constrained, multi-hormone synthetic time series that capture intra- and inter-individual variability across phenotypes.
We developed a semi-mechanistic mathematical framework to generate synthetic multi-hormone profiles (estradiol, FSH, LH, AMH, testosterone, GnRH) using parametric equations embedding known physiological feedbacks (e.g., estradiol-LH delay, estradiol suppression of FSH). Stochastic components were calibrated to reported physiological ranges. Eumenorrheic and PCOS-like phenotypes were defined through parameter adjustments. Data were analysed using Principal Component Analysis (PCA) for phenotype separation, and evaluated in a supervised setting using logistic regression with stratified train/test splitting, reporting accuracy, sensitivity, specificity, and ROC AUC.
Eumenorrheic profiles displayed classical mid-cycle estradiol and LH peaks, biphasic FSH, and stable AMH and testosterone levels. In contrast, PCOS profiles showed elevated LH and testosterone, high AMH, blunted estradiol, and dysregulated GnRH pulsatility. PCA revealed clear separation between phenotypes (PC1 +PC2 = 82 % variance), and k-means clustering (k = 2) accurately grouped individuals without label information. PCA showed clear separation between phenotypes, consistent with known endocrine patterns. Logistic regression achieved 100 % accuracy, sensitivity, and specificity, with an AUC of 1.00, confirming robust, phenotype-discriminative features in the synthetic dataset.
This simulation framework reproduces physiologically accurate hormone dynamics and discriminates ovulatory from anovulatory cycles, offering applications in AI training, phenotype discovery, and medical education.
卵巢激素的动态相互作用是生殖生理学的核心，但它们周期性变化的复杂性给分析、模拟和教学带来了挑战。本研究提出了一个框架，用于生成受生理约束的多激素合成时间序列，该序列能够捕捉不同表型的个体内和个体间变异性。
我们开发了一个半机械数学框架，使用嵌入已知生理反馈（如雌二醇 - 促黄体生成素延迟、雌二醇对促卵泡生成素的抑制）的参数方程来生成合成多激素谱（雌二醇、促卵泡生成素、促黄体生成素、抗苗勒管激素、睾酮、促性腺激素释放激素）。随机成分根据报告的生理范围进行校准。通过参数调整定义了月经周期正常和多囊卵巢综合征样表型。使用主成分分析（PCA）进行表型分离分析数据，并在监督设置中使用逻辑回归和分层训练/测试分割进行评估，报告准确性、敏感性、特异性和ROC曲线下面积（AUC）。
月经周期正常的谱显示出典型的周期中期雌二醇和促黄体生成素峰值、双相促卵泡生成素以及稳定的抗苗勒管激素和睾酮水平。相比之下，多囊卵巢综合征谱显示促黄体生成素和睾酮升高、抗苗勒管激素水平高、雌二醇水平降低以及促性腺激素释放激素脉冲性失调。主成分分析显示表型之间有明显分离（PC1 + PC2 = 82%的方差），并且k均值聚类（k = 2）在没有标签信息的情况下准确地对个体进行了分组。主成分分析显示表型之间有明显分离，与已知的内分泌模式一致。逻辑回归的准确率、敏感性和特异性均达到100%，AUC为1.00，证实了合成数据集中具有强大的、区分表型的特征。
这个模拟框架再现了生理上准确的激素动态，并区分了排卵周期和无排卵周期，在人工智能训练、表型发现和医学教育中具有应用价值。
