Fig. A8.

Summary of analysis of model randomness against feature noise. Effect sizes are shown for the features that are significantly different between false negative and true positive populations (green). Effect sizes are also shown for the features that are significantly different between false positive and true negative populations (purple). Features with the top ten effect sizes are shown. Analysis was performed on holdout test sets that include patients who were on CRRT for more than seven days. a) Model trained on data from UCLA: CRRT (n=1,270), and evaluated on a holdout test set (n=1,956). b) Model trained on data from UCLA: CRRT (n=1,270), and evaluated on an external test set from Cedars: CRRT (n=2,870). c) Model trained on data from Cedars: CRRT (n=1,069), and evaluated on a holdout test set (n=1,443). d) Model trained on data from Cedars: CRRT (n=1,069), and evaluated on an external test set from UCLA: CRRT (n=3,650). e) Model trained on a combination of data from UCLA: CRRT and Cedars: CRRT (n=2,343), and evaluated on a holdout test set (n=3,427).