Fig. 3. Model performance by subgroups in the test dataset (01 February 2019 to 31 January 2020).

Balanced accuracy (a), positive predictive value (PPV) (b), and negative predictive value (NPV) (c) were compared. IMD=index of multiple deprivation score (higher scores indicate greater deprivation). ‘Weekday’ refers to the day of the week of the index date. Comorbidity was calculated using Charlson comorbidity score. ‘Source’ refers to the source of admission. Overall performance is shown by the dashed line in each plot. 95% confidence intervals were calculated using bootstrap (n = 500). F1 score, area under the receiver operating curve (AUROC), and area under the precision-recall curve (AUPRC) are shown in Supplementary Fig. 9.