Fig. 8

SNF cluster labels prediction using elastic net regression and causal model. A Features selected by the elastic net multinomial logistic regression to predict cluster labels across ten repetitions of 10-fold cross-validation. The bar graph indicates the number of times each analyte was selected out of the 100 trained models. Only analytes selected in at least half of the models are shown. B, C A causal graphical model depicting the predictive clinical variables and analytes directly linked to the cluster labels (B), and specifically in cluster III (C), given all other strongly predictive clinical variables and analytes. Edge thickness indicates the stability of each adjacency in the causal model across 100 bootstrap samples. Different edge types indicate different causal information, inferred by the FCI-Max algorithm: A → B indicates that A causes B, A ↔ B indicates that there is a latent confounder of A and B, A o→ B indicates that B is not a cause of A, but it is unclear if A causes B or if a latent confounder causes A and B, and A o–o B indicates that there is an interaction between A and B but the causal direction of the interaction cannot be determined. Adjacencies in the causal graphical models are controlled at an FDR < 0.05, calculated using the Benjamini-Hochberg procedure