Prognostic Groups via Interpretable Function Approximation

Le résumé fourni par la source

The Cox ( 1972 ) model remains the most widely used method for assessing associations of patient characteristics and biomarkers with time-to-event outcomes. However, machine learning methods such as Random Survival Forests ( Ishwaran et al. 2008 ), Boosting implementations (e.g., Buhlmann and Hohorn 2007 ) or XGBoost (e.g., Liu et al. 2021 ), and neural networks both in early work by Faraggi and Simon (1995 ) and, much more recently, deep learning methods extensions ( Wiegrebe et al. 2024 ) have made significant inroads in survival analysis, particularly in predicting cancer prognosis. These methods are best suited for large data and some complex data sets when the primary goal is about the best possible predictions or the underlying estimation of conditional regression function. In many cases, these methods can provide more accurate predictions in some settings compared to linear Cox regression. However, they lack interpretability with respect to understanding baseline variable associations with outcomes. It is interesting that Friedman (2024) , an innovator and co-inventor of so many machine learning methods, in recent work has proposed a method with a goal of more transparent machine learning based on a new smooth extension to regression trees called ”function trees.”

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.