Functional Independent Component Analysis for Random Survival Forests with Censored Functional Data

G. Loffredo, E. Romano, A. M. Aguilera del Pino, F. Maturo, M. Vidal

This study addresses survival prediction with mixed data, where scalar covariates coexist with sparse and irregular functional trajectories subject to right-censoring. Standard approaches rely on Functional Principal Component Analysis (FPCA), which is variance-driven and may overlook non-Gaussian patterns relevant for survival outcomes. We propose a two-stage framework where functional trajectories are first reconstructed under the Censored Functional Data paradigm using PACE, yielding FPCA scores. Then, a kurtosis-based Independent Component Analysis (ICA) is applied to extract approximately independent functional features. These are combined with scalar covariates and used as inputs to a Random Survival Forest. Empirical results on the SOFA dataset show improved predictive performance and stability compared to FPCA-based representations, particularly in smaller samples.

Palabras clave: Survival analysis, Functional data analysis, Independent component analysis, Random survival forests, Censored data

SI Sesión Hispano-Italiana
3 de septiembre de 2026  11:10
Aula 20

F. Porro, M. Restaino, J. E. Ruiz Castro, M. Zenga

A. Sarra, T. Di Battista, A. Evangelista, E. Nissi, N. Di Deo