Modelling ICU occupancy and patient flow dynamics during the COVID-19 pandemic
Year: 2025
Type: dissertation
Abstract: In the wake of COVID-19, the world faced the mammoth task of understanding how this new disease functioned, its impact on humans, and, in particular, its implications for public health. The contributions of this thesis are part of the broader collaborative effort of understanding the COVID-19 pandemic by analysing the intensive care unit (ICU) dynamics in Germany. Since only data on ICU occupancy are publicly available at district level, and not on ICU admission or patient length of stay, ICU patient flow dynamics can not directly be analysed. This thus poses a missing data problem. The contributions of this thesis analyse the ICU occupancy and subsequently employ statistical methods in order to disentangle the unobserved patient inflow, length of stay and outflow from the observed occupancy data. Part I of this thesis introduces metrics commonly employed to gain a comprehensive understanding of the impact of an infectious disease on public health, in the context of COVID-19. It also discusses the data available for public health surveillance in Germany and highlights the contextual motivation for the contributions of this thesis, namely the absence of district-level data on ICU patient flow dynamics. Chapter 3 introduces statistical modelling, with a particular focus on parametric models. It lays out the assumptions required for consistent estimation and discusses approaches to quantifying estimator uncertainty. Chapter 4 then links the statistical methodology to the previously introduced data, motivating the contributions of this thesis. Part II analyses the ICU occupancy. The distribution of ICU beds among patients infected with COVID-19, patients not infected with COVID-19, and unoccupied status is modelled using a multinomial model. This approach allows estimation of associations between ICU bed distribution and infection rates in the previous week, spatial correlation (captured through a two-dimensional thin-plate spline), and district-level heterogeneity (modelled via a random intercept). The multinomial assumption is particularly valuable for prediction, as it reflects the mutually exclusive nature of ICU bed allocation. Part III builds on the analysis of Part II by inferring patient inflow and outflow from ICU occupancy. Chapter 6 employs the stochastic Expectation-Maximisation (sEM) algorithm to iteratively simulate from a truncated Skellam distribution with incoming and outgoing intensity parameters, which are estimated using two independent Poisson models. The inflow model incorporates age specific infection rates from the previous week, spatial correlation via a two-dimensional thin-plate spline, temporal correlation via a penalised B-spline over time, and categorical weekday effects. The outflow model is specified analogously, with an additional offset defined as the sum of previously incoming patients, weighted by the exit rate, which is taken to be fixed. Chapter 7 extends this methodology by treating exit rate, also referred to as length of stay, as a random parameter rather than a fixed input. Thus, the outflow model of the M-Step of the sEM is altered, such that the outflow is now solely a linear combination of inflow. This introduces a sum-to-one and non-negativity constraint. Thus, the parameters are estimated via constrained maximum likelihood, which introduces bias. This bias is corrected employing an additional simulation step in the sEM.
DOI: 10.5282/edoc.36468
Author: Rave, Martje
Institution: Ludwig-Maximilians-Universität München
Language: English
Subfield: Modeling and Simulation
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7128301393
