| dc.description.abstract | Fertility treatment requires navigating a series of complex and connected decisions, each with tradeoffs that impact both patient outcomes and clinical workload. In in vitro fertilization (IVF), choices about stimulation protocols, embryo transfer count, and treatment timing are interdependent, and each cycle carries uncertainty. Despite increasing data availability, most clinical tools still address these decisions in isolation. As a result, care remains variable across providers, and patients often face unclear guidance at critical moments in the process. To address this gap, we present an integrative deep learning framework that simultaneously optimizes three critical IVF decision points using data from 33,000 treatment cycles (2021- 2023): (1) embryo transfer count recommendation (MAE = 0.716), (2) ovarian stimulation protocol selection (AUC = 0.85), and (3) biochemical pregnancy prediction (AUC = 0.75). Our comprehensive benchmarking pipeline evaluates classical statistical models, ensemble methods (XGBoost), and novel architectures, including TabPFN, an attention-based probabilistic model that achieved comparable performance to top-performing baselines. To enhance clinical trust, we apply SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME), which consistently uncover biologically meaningful variables, including biomarkers such as Anti-Müllerian Hormone (AMH), patient age, and the day-3 high-quality embryo rate. The final architecture combines: (i) a shared representation learning backbone, (ii) task-specific prediction heads, and (iii) homoscedastic uncertainty weighting for dynamic task balancing. This multitask approach achieves near-state-of-the-art performance (protocol AUC = 0.79, pregnancy AUC = 0.74) while offering practical deployment advantages such as faster inference and fewer parameters than equivalent single-task ensembles. This system’s efficiency, interpretability, and scalability will reshape current IVF practice if adopted in clinical settings, offering a more consistent and data-driven foundation for treatment decisions. | |
