Table 3.

Model performance of extreme gradient boosting (XGB) models with 200 features, baseline logistic regression (LR) model, and LR model with 200 features predicting 24-hour discharge in the test dataset (01 February 2019 to 31 January 2020) and an additional test dataset post-COVID (01 February 2021 to 31 January 2022)

Admission type	Model	Accuracy	Balanced accuracy	Sensitivity/Recall	Specificity	PPV/Precision	NPV	F1-score	AUROC	AUPRC	MAE (%)
Test data: 01/02/2019 to 31/01/2020
Elective	XGB (200 features)	0.823	0.767	0.673	0.861	0.555	0.911	0.609	0.871	0.658	8.9
Elective	LR (baseline)	0.464	0.596	0.820	0.372	0.252	0.889	0.385	0.629	0.269	10.7
Elective	LR (200 features)	0.696	0.740	0.815	0.666	0.386	0.933	0.524	0.821	0.538	10.6
Emergency	XGB (200 features)	0.844	0.756	0.616	0.896	0.571	0.912	0.593	0.860	0.644	4.9
Emergency	LR (baseline)	0.637	0.654	0.682	0.626	0.292	0.897	0.409	0.708	0.349	5.8
Emergency	LR (200 features)	0.718	0.738	0.769	0.707	0.372	0.931	0.501	0.813	0.507	5.6
Overall	XGB (200 features)	0.837	0.752	0.615	0.888	0.561	0.909	0.587	0.859	0.634	4.6
Overall	LR (baseline)	0.589	0.642	0.726	0.558	0.276	0.898	0.400	0.694	0.327	5.4
Overall	LR (200 features)	0.700	0.733	0.787	0.680	0.363	0.932	0.497	0.809	0.497	5.0
Test data: 01/02/2021 to 31/01/2022
Elective	XGB (200 features)	0.825	0.753	0.638	0.869	0.532	0.911	0.580	0.864	0.614	11.6
Emergency	XGB (200 features)	0.835	0.703	0.501	0.906	0.528	0.896	0.514	0.820	0.543	10.0

The baseline LR models only included age, sex, day of the week, and hours since admission. PPV: Positive predictive value, NPV: Negative predictive value, AUROC: Area under the receiver operating curve, AUPRC: Area under the precision-recall curve, MAE: Normalised mean absolute error (mean difference in predicted and actual discharges per day divided by the mean number of discharges per day).