Figure 1
Partition of the study data into the training and hold-out test set. Attendances which resulted in planned reattendances (328 attendances) were removed. All attendances (9393 attendances) occurring after 01/02/2020 were also discarded to remove attendances which coincided with the COVID-19 pandemic. Reattendance rates (bottom row of rectangles representing training and test sets) display the observed 72-h reattendance rate for each cohort. Models were trained using (grouped) five-fold cross validation of the training set with patients in the train set not present in the validation set during the training phase. Model hyperparameters were selected such that the validation AUROC (mean out-of-fold AUROC across the five folds) was maximized. The final model was a mean ensemble of the five models (the average prediction of the five models) trained in the cross-validation process, which was then evaluated on the hold-out test set.
