Table 3.

Performance of CovRNN models on the OPTUM test set before and after fine-tuning

	CovRNN trained only on OPTUM fine-tuning set	CRWD-trained CovRNN before fine-tuning	CRWD-trained CovRNN after fine-tuning using OPTUM fine-tuning set
In-hospital mortality binary prediction	88·6%	87·0%	91·3%
Mechanical ventilation binary prediction	90·4%	72·5%	91·5%
Prolonged hospital stay (>7 days) binary prediction	78·1%	68·0%	81·0%
In-hospital mortality survival prediction*	86·1%	77·1%	88·9%
Mechanical ventilation survival prediction*	90·2%	69·2%	93·7%

Data are area under the receiver operating characteristic curve, unless otherwise indicated. All data are based on evaluation in the OPTUM test set. CRWD=Cerner Real-World COVID-19 Q3 Dataset. OPTUM=Optum deidentified COVID-19 electronic health record dataset.

^*Unlike the binary classifications used in other models, values for the survival models represent the concordance index.