Table 4.

Final predictive models

Model	Equation	MAPE in train set	Ljung–Box Test	MAPE in test set
CT(t)	$H\left(t\right)=-\mathbf{3315.30}+\mathbf{17.03}\times \mathit{CT}\left(\mathit{t}\right)+\mathbf{196.63}\times \mathit{Ld}\left(\mathit{t}\right)+\epsilon \left(t\right)+0.84\times \epsilon \left(t-1\right)$, with $\epsilon \left(t\right)\sim \mathcal{N}\left(0,3420285\right)$	6.82	0.0490*	20.02
CT(t − 1)	$H\left(t\right)=\mathbf{7.05}\times \mathit{CT}\left(\mathit{t},-,\mathbf{1}\right)+\mathbf{889.49}\times \mathit{Ld}\left(\mathit{t}\right)+\epsilon \left(t\right)+2.27\times \eta \left(t-1\right)-1.93\times \eta \left(t-2\right)+0.59\times \eta \left(t-3\right)$, with $\epsilon \left(t\right)\sim \mathcal{N}\left(0,924814\right)$	25.82	0.0387*	20.72
CT(t − 2)	$H\left(t\right)=\mathbf{29227}.\mathbf{22}-\mathbf{2.68}\times \mathit{CT}\left(\mathit{t}-\mathbf{2}\right)-\mathbf{1024.94}\times \mathit{Ld}\left(\mathit{t}\right)+\epsilon \left(t\right)+1.95\times \eta \left(t-1\right)-0.96\times \eta \left(t-2\right)$, with $\epsilon \left(t\right)\sim \mathcal{N}\left(0,968004\right)$	30.85	0.2406	127.13
CT(t − 1), CT(t − 2), H(t − 1), H(t − 2)	$H\left(t\right)=\mathbf{8.60}\times \mathit{CT}\left(\mathit{t}-\mathbf{1}\right)-\mathbf{4.27}\times \mathit{CT}\left(\mathit{t},-,\mathbf{2}\right)+\mathbf{149.89}\times \mathit{Ld}\left(\mathit{t}\right)+\mathbf{0.97}\times \mathit{H}\left(\mathit{t}-\mathbf{1}\right)-\mathbf{0.42}\times \mathit{H}\left(\mathit{t},-,\mathbf{2}\right)+\epsilon \left(t\right)+1.83\times \eta \left(t-1\right)$, with $\epsilon \left(t\right)\sim \mathcal{N}\left(0,479558\right)$	24.40	0.1182	5.09

The ‘model’ column gives the predictors entered in the algorithm to predict the number of hospitalisations for the week ‘t’. Hence, ‘t − 1’ and ‘t − 2’ are one and two weeks before (i.e. lag − 1 and lag − 2)

The terms in bold correspond to the regression part of the model, and the other terms to the error η(t) which can be expressed with an auto-regressive integrated moving average (ARIMA) model with ε(t) an uncorrelated error term (i.e. white noise) following a normal law N with variance in parentheses

CT(x), where x in {t, t − 1, t − 2}, corresponds to the number of CT-scans performed in the COVID-19 teleradiological emergency workflow during the week ‘x’

H(x′), where x′ in {t − 1, t − 2}, corresponds to the number of patients hospitalised in mainland French hospitals during the week ‘x′’

Ld(t) is a binary variable that takes the value 1 if France is under national lockdown and 0 otherwise

ARIMA auto-regressive integrative moving average, MAPE mean absolute percentage error

^*p < 0.05