AutoSEIR: Accurate Forecasting from Real-time Epidemic Data Using Machine Learning

ABSTRACT

Since the SARS-CoV-2 virus outbreak has been recognized as a pandemic on March 11, 2020, several models have been proposed to forecast its evolution following the governments interventions. In particular, the need for fine-grained predictions, based on real-time and fluctuating data, has highlighted the limitations of traditional SEIR models and parameter fitting, encouraging the study of new models for greater accuracy. In this paper we propose a novel approach to epidemiological parameter fitting and epidemic forecasting, based on an extended version of the SEIR compartmental model and on an auto-differentiation technique for partially observable ODEs (Ordinary Differential Equations). The results on publicly available data show that the proposed model is able to fit the daily cases curve with greater accuracy, obtaining also a lower forecast error. Furthermore, the forecast accuracy allows to predict the peak with an error margin of less than one week, up to 50 days before the peak happens.