Deep Survival Analysis Frameworks for Personalized Prognosis Prediction

This thesis presents a series of novel computational frameworks designed to enhance survival analysis in biomedical research by integrating advanced artificial intelligence methods with privacy-preserving data sharing and mechanistic modeling. First, the SYNDSURV framework is a synthetic data-based approach, enabling decentralized survival analysis without compromising patient sensitive data. By generating synthetic survival data at local nodes—combining Bayesian data synthesis with Accelerated Failure Time modeling—it facilitates centralized training of survival models while maintaining privacy constraints through differential privacy techniques. Next, the VAE-Surv model introduces a deep-learning-based survival analysis approach that leverages Variational Autoencoders to extract latent representations from high-dimensional genetic and cytogenetic data. This representation, enriched with clinical features, is concatenated with a deep survival network that achieves high prognostic predictions and patient stratification, particularly in the context of Myelodysplastic Syndromes (MDS). This approach not only improves predictive accuracy but also enhances interpretability, bridging the gap between deep learning and clinical applicability. Finally, this thesis explores a stochastic diffusion process approach, conceptualizing survival as a First Fitting Time (FHT) problem. Here, patient-specific parameters are inferred via neural networks, offering mechanistic insights into individual survival trajectories and capturing non-linear progression patterns that traditional survival models often overlook. The experimental results across multiple datasets demonstrate that these frameworks consistently outperform traditional survival analysis methods, providing interpretable and clinically actionable insights. By addressing key challenges in federated learning, data privacy, and multimodal integration, this work contributes to the development of more effective, personalized strategies for managing rare and complex diseases, especially in the context of prognosis prediction. During my PhD, I have collaborated with the Center for Health Data Science, University of Copenhagen, under the supervision of Professor Anders Krogh. I also thank Prof. Matteo Della Porta (Humanitas University) and Prof. Gastone Castellani (University of Bologna), my external supervisors within the GenoMed4All consortium, for their guidance and continuous support throughout my PhD. Their expertise and collaboration have significantly enriched many of my research projects. I