Figure 3.

The structured grouping strategy improves model interpretability when the grouping patterns are clinically interpretable by domain experts

(A–C) We used four algorithms to measure feature importance of composite indices: logistic regression, decision tree, random forest, and xgboost. The feature importance was evaluated on five composite embeddings from (A) the composite model_g1, where the importance measures from logistic regression showed slightly different results but commonly showed that the liver function group is relatively more important, (B) the composite model_g2, where all four algorithms indicated that comp1 is relatively more important than others, and (C) the composite model_g3, where importance of all seven composite indices is comparable. The error bars indicate standard deviations.