Study objectives
Design decisions	Optimize for F1-score for all three models for the 30-day mortality prediction task and report all metrics including areas under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, and accuracy
Reasons	30-day mortality is chosen as the target outcome in accordance with clinical precedence [23–25]. The F1-score finds an equal balance between PPV (precision) and sensitivity (recall), which gives a better indication of model performance for unbalanced dataset (mortality is relatively rare compared to survival). Reporting all metrics allows assessment of how the models might perform in populations with a different COVID-19-related mortality distribution. With COVID mortality rates varying with time, model drift can be a real concern under different care delivery parameters during surges [26]