Data-driven prediction of COVID-19 cases in Germany for decision making

Abstract

Background The COVID-19 pandemic has led to a high interest in mathematical models describing and predicting the diverse aspects and implications of the virus outbreak. Model results represent an important part of the information base for the decision process on different administrative levels. The Robert-Koch-Institute (RKI) initiated a project whose main goal is to predict COVID-19-specific occupation of beds in intensive care units: Steuerungs-Prognose von Intensivmedizinischen COVID-19 Kapazitäten (SPoCK). The incidence of COVID-19 cases is a crucial predictor for this occupation. Methods We developed a model based on ordinary differential equations for the COVID-19 spread with a time-dependent infection rate described by a spline. Furthermore, the model explicitly accounts for weekday-specific reporting and adjusts for reporting delay. The model is calibrated in a purely data-driven manner by a maximum likelihood approach. Uncertainties are evaluated using the profile likelihood method. The uncertainty about the appropriate modeling assumptions can be accounted for by including and merging results of different modelling approaches. The analysis uses data from Germany describing the COVID-19 spread from early 2020 until March 31st, 2021. Results The model is calibrated based on incident cases on a daily basis and provides daily predictions of incident COVID-19 cases for the upcoming three weeks including uncertainty estimates for Germany and its subregions. Derived quantities such as cumulative counts and 7-day incidences with corresponding uncertainties can be computed. The estimation of the time-dependent infection rate leads to an estimated reproduction factor that is oscillating around one. Data-driven estimation of the dark figure purely from incident cases is not feasible. Conclusions We successfully implemented a procedure to forecast near future COVID-19 incidences for diverse subregions in Germany which are made available to various decision makers via an interactive web application. Results of the incidence modeling are also used as a predictor for forecasting the need of intensive care units.

Improved healthcare disaster decision-making utilizing information extraction from complementary social media data during the COVID-19 pandemic

“…However, further suitable state-of-the-art algorithms are available to forecast the COVID-19 incidence. For instance, forecasts based on the SIR-like models that account for Susceptible (S), Infected (I), and Recovered (R) people [ 50 , 55 , 56 ], Random Forest Regression [ 57 ], or Facebook's Prophet model [ 58 , 59 ]. These works rely either on time-series forecasts based on the incidence solely or include different user-generated content like Google Trends data [ 57 , 60 ].…”

Constructing Exact Confidence Regions on Parameter Manifolds of Non-Linear Models

“…As a real-world example for applications in systems biology, where mathematical models are often formulated in terms of differential equations, we illustrated the use integral manifold method via the topical class of SIR models in section 3.4.2. Although we selected the structurally simplest member of this family for the purpose of demonstration, many of today's state of the art methods for predicting the spread of infectious diseases such as COVID-19 are nevertheless direct logical descendants of this model [28,36].…”

“…ODE-based approaches such as this assume that the sub-populations are large enough to be modelled as real numbers and well-mixed. While this basic SIR model is certainly an oversimplification of the mechanisms underlying any real-world outbreaks of infectious diseases, there are various ways to extend this model such that it provides a more accurate description of real disease transmission, for instance by allowing for time dependence in the parameters [27,28]. Table 1: Dataset from an influenza outbreak at an English boarding school in 1978, reproduced from [27].…”

Deep learning and differential equations for modeling changes in individual‐level latent dynamics between observation periods

“…For example, L1-regularized regression could be considered (Inan & Wang, 2017;Schelldorfer et al, 2011Schelldorfer et al, , 2014Wang et al, 2012) or coupling the likelihood of multiple points in time (Schmidtmann et al, 2014;Zöller et al, 2016), as proposed in our own work. Moreover, there are many spline-based approaches that allow for modeling of changes in the dynamics (Bringmann et al, 2017;Hong & Lian, 2012;Meng et al, 2021;Refisch et al, 2022;Wang et al, 2007). Kamalabad and Grzegorczyk (2021) suggested to learn which temporal segments should be coupled, allowing the parameters to be informed by the previous segment, or uncoupled where the parameters of the segment can be estimated without previous information.…”