Risk prediction of clinical adverse outcomes with machine learning in a cohort of critically ill patients with atrial fibrillation