Federated Learning for Embryo Evaluation in Assisted Reproduction: development of distributed predictive models for assessing embryonic development

Choosing which embryo to transfer is one of the most delicate steps of an in vitro fertilization (IVF) cycle. It still relies largely on the embryologist’s visual assessment of morphology, which is subjective and hard to reproduce. Time-lapse imaging and machine learning offer a way to describe embryo development quantitatively. However, in order to achieve good predictive performance, a sufficient amount of data need to be collected for training; those data could be spread across clinics and protected by strict regulations such as the GDPR and the EU AI Act. Federated Learning (FL) addresses exactly this situation: it trains a shared model without the data ever leaving the institution that owns them. In this thesis, FL has been applied to the prediction of blastocyst formation from optical-flow time series extracted from time-lapse videos of embryos development. The dataset contains 5011 embryos acquired between 2013 and 2020. To mimic a realistic multi-center setting, the dataset has been split in 8 different subsets, 7 used for the training of the models and 1 kept for the global evaluation. Five model configurations have been compared: ROCKET, ConvTran, a hybrid LSTM-FCN network, and the MOMENT foundation model, used both with full fine-tuning and as a frozen feature extractor with a small classification head on top. Every model is trained with FedAvg, using the Flower framework, for 2 to 7 clients, and compared with the same model trained on the pooled data. LSTM-FCN gives the best and most stable results. In the federated setting it reaches an accuracy of 0.84–0.87 on the global test set, matching or beating its centralized counterpart (0.83–0.84). The MOMENT-embedding model reaches about 0.81 with seven clients while exchanging only the classification head, which keeps the communication cost very low. A repeated-measures ANOVA shows a strong effect of the architecture (p < 0.001) and a significant interaction between architecture and training mode (p = 0.030), whereas the training mode alone has no significant effect (p = 0.093). Finally, a federated explainability procedure based on SHAP is proposed, which is called SHAP-FL. Each client computes Shapley values locally on its own data and sends the server only an aggregated importance vector, which the server combines into a global explanation. Applied to LSTM-FCN, SHAP-FL, shows that the model bases its decision on the last time period of culture, when cavitation and blastocyst expansion take place. the explanations of the seven centers agree closely (mean Spearman correlation 0.95).