Fig. 6

Effect of lockdown in Italy. a, The number of ICU patients grows slower than exponential starting from March 14th, i.e., five days after the lockdown of March 9th (green triangle). Five days correspond to the current estimate of median incubation time of COVID-19 [14]. We divided the data points into three groups, before March 14th, between March 14th and March 24th, and up to April 5th. We fit the first group with an exponential curve (red dashed line, ICU patients(t) ∝ exp[r t], t = days), the second group with a line (magenta line, ICU patients(t) ∝ b t, t = days), and the third group with a logistic curve (green dashed line, ICU patients(t) = k/(1 + q exp[−p t]), t = days). The best fit parameters are r = 0.213 (95% CI: 0.197–0.229), b = 104 (95% CI: 100–108), k = 1980 (95% CI: 1960–2000), q = 203 (95% CI: 175–231), p = 0.225 (95% CI: 0.219–0.231). The root mean square error (RMSE) and normalized-RMSE (n-RMSE) of the exponential fit are 13.80 and 3.1%, respectively. RMSE and n-RMSE for the linear fit and the logistic fit are 18.73 and 1.8%, and 20.02 and 1.0%, respectively. b, As for panel a, representing the number of deaths. The linear trend starts eleven days after the lockdown. The best fit parameters are r = 0.197 (95% CI: 0.189–0.205), and b = 230 (95% CI: 226–236). RMSE and n-RMSE for the exponential fit are 20.57 and 1.9%, respectively; RMSE and n-RMSE for the linear fit are 18.84 and 0.9%, respectively. c-d, As for panel a-b, up to 4 months after the lockdown declaration (July 9th). For all panels, we excluded data from the Italian regions where the ICUs saturated, i.e. Lombardy, Piedmont, Marche, Trentino Alto Adige, Valle d’Aosta, which host 28% of the total Italian population. The non-linear least squares problems have been solved using Levenberg-Marquardt algorithm, with NumPy library (Python)