AI-DRIVEN MORTALITY ANALYSIS: DESIGN OF AN INTERACTIVE ECOSYSTEM FOR THE ADVANCED DEVELOPMENT OF MORTALITY MODELS AND THEIR DYNAMIC VISUALIZATION

This thesis presents the design and implementation of the Reproducible Framework for Survival Analysis (SURF), a modular survival analysis framework conceived as a reproducible and deployment-oriented analytical system. The framework integrates statistical, Machine Learning (ML), and Deep Learning (DL) survival models within a unified analytical environment. The system architecture enforces consistent preprocessing, configuration stability, and a clear separation between training and inference stages. The analytical core implements a standardized pipeline for Time-to-Event (TTE) data processing, including survival-aware validation and deterministic dataset partitioning. Models are serialized together with preprocessing artifacts and metadata, enabling artifact-based reconstruction during deployment. This mechanism ensures reproducibility across experimental and operational environments. A dashboard architecture provides structured interaction with the analytical pipeline while maintaining separation between computation and presentation logic. Role-Based Access Control (RBAC) mechanisms distinguish Standard User (USER) and Technical User (TECH) roles, enforcing governance over model lifecycle management and deployment procedures. The framework integrates interpretability modules based on SHapley Additive exPlanations (SHAP) for local attribution and global feature importance. Structured explanations are translated into natural language through a controlled Large Language Model (LLM) interface grounded in analytical artifacts. Experimental evaluation across heterogeneous survival models, including Support Vector Machine (SVM), a Deep Neural Network implementing a non-linear Cox proportional hazards model (DeepSurv), Random Survival Forest (RSF), and Gradient Boosting (GB), showed comparable discrimination performance across models. RSF achieved the highest observed Harrell's Concordance Index (C-index) in this experimental setting. Artifact-based inference experiments illustrated consistent propagation from training to deployment, and interpretability results showed coherence between survival trajectories and local attribution signals. Overall, the experimental and engineering results support the architectural principles underlying the framework, including configuration stability, artifact-based deployment, and structured model lifecycle management. SURF therefore provides a reproducible environment for survival modeling, comparative experimentation, and operational inference.