Multi-state modeling of hospitalized patient data for prediction, etiology, and burden analyses - from hospital-acquired infections to pandemic settings

Abstract: This thesis addressed several strategies for avoiding bias and misinterpretation
in the analysis of data on hospitalized patients. Specifically, the presented
methods avoid competing risks, time-dependent, and selection biases. The first
part of the work outlined how to conduct competing risks analyses in studies into
the epidemiology of S. aureus surgical site infection and pneumonia with complex
sampling designs. In these settings, observation of infection as the main event
of interest was prevented by the competing risks of death and discharge alive in
the hospital. In addition to avoiding a competing risks bias, weights were used to
mimic a population closer to the more generalizable target population. The estimates properly taking into account these challenges diverge greatly from results based on simpler, yet naive alternatives (non-competing risks, unweighted analyses). The results constitute an important contribution to identifying hospitalized patients who would benefit most from interventions against S. aureus infections.
The second section focused on multi-state methodology applied to data on
hospitalized COVID-19 patients. In these analyses, a model was used with one
or more intermediate states (e.g. ICU admission, mechanical ventilation, severe
disease) along with the terminal states of discharge alive and death. In addition
to averting competing risks biases, multi-state methods properly model the
time-varying nature of these intermediate states to avoid time-dependent biases.
These models also enable the prediction of clinical courses and the duration of
clinically relevant states. The models were applied to publicly available data published early on during the pandemic that served as a template for analyses in a
wide range of COVID-19 research. A true strength of the methods is that they
can be performed in real-time, enabling swift analyses in quickly changing circumstances and avoiding a selection bias that can result from omitting current
cases. This aspect was highlighted in the demonstration of analyzing emerging
pandemic variants.
In the third section, strategies were presented for assessing changes in the
burden of infection (modeled as an intermediate event in a multi-state model)
after an intervention has been introduced. Previous research has focused on
estimating the burden of a time-fixed exposure, or the reduction in burden resulting
from the total elimination of a time-varying exposure. In this work, estimation
focused on the more likely situation that a prevention leads to a decrease in the
rate of a hospital-acquired infection. In addition to estimating a reduction in mortality, these methods can be linked with financial data to estimate cost reductions. A simple tool using R code was developed that requires only routinely collected data to estimate the burden.
In summary, the methods detailed in this work constitute important options
for research on hospital data, whether for nosocomial infections, COVID-19, or
future pandemics

Angaben zum Objekt

• Standort: Deutsche Nationalbibliothek Frankfurt am Main
• Umfang: Online-Ressource
• Sprache: Englisch
• Anmerkungen: Universität Freiburg, Dissertation, 2024

Klassifikation und Themen

• Schlagwort: Statistik ; Biostatistik ; Medizinische Statistik ; Ereignisdatenanalyse

Beteiligte, Orts- und Zeitangaben

• Ereignis: Veröffentlichung
• (wo): Freiburg
• (wer): Universität
• (wann): 2024
• Urheber: Hazard, Derek
• Beteiligte Personen und Organisationen: Wolkewitz, Martin ; Rieg, Siegbert ; Institut für Medizinische Biometrie und Statistik. Freiburg im Breisgau ; Albert-Ludwigs-Universität Freiburg. Medizinische Fakultät

Weitere Informationen

• DOI: 10.6094/UNIFR/246956
• URN: urn:nbn:de:bsz:25-freidok-2469567
• Rechteinformation: Open Access; Der Zugriff auf das Objekt ist unbeschränkt möglich.
• Letzte Aktualisierung: 14.08.2025, 10:56 MESZ