Kong Na, Xia Yu, Wang Zhilong, Zhang Hui, Duan Liyan, Zhu Yingchun, Huang Chenyang, Yan Guijun, Mei Jie, Li Wujun, Sun Haixiang
Nanjing Drum Tower Hospital, Drum Tower Clinical Medical College, Nanjing University of Chinese Medicine, Nanjing, 210008, Jiangsu, China.
Center for Reproductive Medicine and Obstetrics and Gynecology, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
J Transl Med. 2025 May 15;23(1):545. doi: 10.1186/s12967-025-06562-8.
Individualizing follicle-stimulating hormone (FSH) dosing during controlled ovarian stimulation (COS) is critical for optimizing outcomes in assisted reproduction but remains difficult due to patient heterogeneity. Most existing models are limited to static predictions of initial doses and do not support real-time adjustments throughout stimulation.
We developed a deep learning model that integrates cross-temporal and cross-feature encoding (CTFE) to predict personalized daily FSH doses in patients undergoing COS using the GnRH agonist long protocol. A total of 13,788 IVF/ICSI cycles conducted between January 2018 and December 2020 were retrospectively analyzed. Women with baseline antral follicle counts between 7 and 30 were included. Data were randomly divided into training (n = 6761), validation (n = 2898), and test (n = 4135) sets. The model encodes both static (e.g., age, BMI, basic hormone levels) and dynamic (e.g., follicle development, hormone trends during COS) variables across stimulation days. Final dose predictions were generated using a K-nearest neighbor algorithm applied to low-dimensional latent representations derived from the deep encoder layers.
The CTFE model achieved a dose classification accuracy of 0.737 (± 0.004) and a weighted F1-score of 0.732 (± 0.005) on the test set. On key stimulation days 1 and 5, the CTFE model significantly outperformed traditional LASSO regression models (F1-score: 0.832 vs 0.699 on day 1; 0.817 vs 0.523 on day 5; p < 0.001). Prediction performance was maintained beyond day 13 using a sliding window mechanism, despite reduced data availability in longer stimulation cycles.
This is the first study to apply a cross-temporal and cross-feature deep learning framework for daily, individualized FSH dose prediction across the full duration of COS. The model demonstrated superior performance over conventional approaches and offers a promising tool for standardizing COS management. Although currently limited by its retrospective, single-center design, the model may support future clinical decision-making and improve COS outcomes. Prospective, multicenter validation studies are warranted to confirm its utility and generalizability.
在控制性卵巢刺激（COS）过程中个体化促卵泡激素（FSH）剂量对于优化辅助生殖结局至关重要，但由于患者的异质性，这仍然具有挑战性。大多数现有模型仅限于对初始剂量的静态预测，不支持在整个刺激过程中进行实时调整。
我们开发了一种深度学习模型，该模型集成了跨时间和跨特征编码（CTFE），以预测使用GnRH激动剂长方案进行COS的患者的个性化每日FSH剂量。回顾性分析了2018年1月至2020年12月期间进行的总共13788个IVF/ICSI周期。纳入基线窦卵泡计数在7至30之间的女性。数据被随机分为训练集（n = 6761）、验证集（n = 2898）和测试集（n = 4135）。该模型对整个刺激周期中的静态变量（如年龄、BMI、基础激素水平）和动态变量（如卵泡发育、COS期间的激素趋势）进行编码。最终剂量预测是使用应用于从深度编码器层导出的低维潜在表示的K近邻算法生成的。
CTFE模型在测试集上实现了剂量分类准确率为0.737（±0.004）和加权F1分数为0.732（±0.005）。在关键刺激日第1天和第5天，CTFE模型显著优于传统的LASSO回归模型（F1分数：第1天为0.832对0.699；第5天为0.817对0.523；p < 0.001）。尽管在较长刺激周期中数据可用性降低，但使用滑动窗口机制在第13天之后仍保持预测性能。
这是第一项应用跨时间和跨特征深度学习框架在COS的整个持续时间内进行每日个体化FSH剂量预测的研究。该模型表现出优于传统方法的性能，并为标准化COS管理提供了一个有前景的工具。尽管目前受其回顾性、单中心设计的限制，但该模型可能支持未来的临床决策并改善COS结局。有必要进行前瞻性、多中心验证研究以确认其效用和可推广性。
