Fig. 2: Feature selection consistency across five scoring functions using repeated cross-validation.
From: Machine-learning driven strategies for adapting immunotherapy in metastatic NSCLC
Repeated 5-fold cross-validation (n = 30) was performed for five scoring functions: (a) linear model with Poisson loss, (b) generalized additive model (GAM) with squared loss, (c) GAM with Poisson loss, (d) GAM with logistic loss, and (e) gradient additive model with logistic loss. Features were ranked based on their frequency of selection across 150 runs and were considered informative if selected in at least 50% of the runs. These informative features were then refitted on the discovery cohort and evaluated on the external validation cohort. Source data are provided as a Source Data file.
