Fig. 1

Initial feature stability metrics. Pearson correlation coefficients, Kendall-Tau scores, Jaccard similarity indices, and rank plots were calculated for each pairwise grouping of validation folds. This provides a quantitative analysis of feature stability by evaluating the effect of training set perturbations on the resulting top-ranked variant list. The rank plots compare the magnitude of coefficient scores assigned to features across a pair of folds, with the axes representing the relative rank of a particular variant. In the presence of perfect model stability, the ranked variant lists would be equivalent between a pair of folds, resulting in a trend line with a slope of 1. However, in this cross-fold analysis, the scatter plots show a high degree of randomness, with trend line slopes approaching zero