Machine learning overview.

(A) Workflow for the prediction of AET outcome based on genomic unitig diversity. The one-hot-encoded unitig data was divided into a training and a validation set stratified by the outcome proportions. Feature correlation filters were applied on training data to remove redundant features, thereby reducing the feature dimension from >500k to 4800. We used recursive feature elimination (RFE) in the nested cross validation (NCV) loops to further reduce the number of features in a model dependent manner. Separate modeling was done with no population structure control (nPSC) and population structure control (PSC), which was implemented by blocking the data based on BAPS groups. The feature combinations obtained during NCV were used to fit new predictors on whole training data. Finally, the model was evaluated on the hold-out validation set. The same validation set was used to evaluate the performance of all models. (B) Three different examples to illustrate how samples are split during the NCV’s outer split into training and test, depending on the method used.