Fig. 3

Relative gain (or loss) in classification performance after Phase II optimizations, relative to Phase I. In Phase II, we implemented 4 types of changes to our classification approach in an attempt to improve performance relative to Phase I. For each type of adjustment, the numbers in this figure represent average differences across all relevant classification algorithms. (The class_weight hyperparameter only applies to some classification algorithms; we calculated averages only for the algorithms that supported it). Green indicates relatively high performance compared to Phase I on the test set; purple indicates lower performance. a Performance metrics for data that had been normalized using either the SCAN or FARMS algorithm after batch adjustment with Combat. b Performance metrics after each variable had been scaled, after feature selection, or after dimensionality reduction. c Performance metrics after altering weights assigned to each class label. Numbers indicate weights assigned to the non-DILI vs. DILI class labels. d Performance metrics for variations on the voting-based ensemble approach. The hard-voting approach combined binarized predictions across the algorithms, whereas soft voting used probabilistic predictions. The scaled methods combined predictions from default and non-default hyperparameter combinations for each algorithm