This paper proposes a multi-layer analytical framework for customer retention in voluntary health insurance, integrating topological behavioral segmentation (UMAP + HDBSCAN), explainable machine learning (XGBoost + SHAP), survival analysis (Kaplan–Meier, Cox proportional hazards, Random Survival Forest), and financial simulation. The framework is validated using administrative records from a Colombian health insurer, covering approximately 298,000 subscribers. Results show strong predictive performance and reveal structural behavioral patterns, including a "Health Paradox" and a "Fortress Effect". The main contribution lies in the integration of analytical layers that jointly address identity, causation, timing, and atypical cases in churn dynamics. Code and full reproducibility materials are available on GitHub.
Katherin Molina (Mon,) studied this question.