Memory is key in capturing COVID-19 epidemiological dynamics

SARS-CoV-2 virus has spread over the world rapidly creating one of the largest pandemics ever. The absence of immunity, presymptomatic transmission, and the relatively high level of virulence of the COVID-19 infection led to a massive flow of patients in intensive care units (ICU). This unprecedented situation calls for rapid and accurate mathematical models to best inform public health policies.We develop an original parsimonious discrete-time model that accounts for the effect of the age of infection on the natural history of the disease. Analysing the ongoing COVID-19 in France as a test case, through the publicly available time series of nationwide hospital mortality and ICU activity, we estimate the value of the key epidemiological parameters and the impact of lock-down implementation delay.This work shows that including memory-effects in the modelling of COVID-19 spreading greatly improves the accuracy of the fit to the epidemiological data. We estimate that the epidemic wave in France started on Jan 20 [Jan 12, Jan 28] (95% likelihood interval) with a reproduction number initially equal to 2.99 [2.59, 3.39], which was reduced by the national lock-down started on Mar 17 to 24 [21, 27] of its value. We also estimate that the implementation of the latter a week earlier or later would have lead to a difference of about respectively −13k and +50k hospital deaths by the end of lock-down.The present parsimonious discrete-time framework constitutes a useful tool for now- and forecasting simultaneously community incidence and ICU capacity strain.