TM économie comportementale
Étude de cas : TM économie comportementale. Recherche parmi 300 000+ dissertationsPar Ludovic Graf • 15 Mars 2024 • Étude de cas • 7 290 Mots (30 Pages) • 94 Vues
Predicting hospital-onset COVID-19 using dynamic networks of patient contact: an observational study
Ashleigh Myall1,2, James R Price1,3, Robert Peach2,4,5, Mohamed Abbas6,7, Sid Mookerjee3, Nina Zhu1, Isa Ahmad3, Damien Ming1, Farzan Ramzan1, Daniel Teixeira7, Christophe Graf 8, Andrea Weisse9,10, Stephan Harbarth7, Alison Holmes1,3, and Mauricio Barahona2*
1Department of Infectious Disease, Imperial College London, London, UK. 2Department of Mathematics, Imperial College London, London, UK. 3Imperial College Healthcare NHS Trust, London, UK. 4Department of Brain Sciences, Imperial College London, London, UK. 5Department of Neurology, University Hospital of Wuerzburg, Germany. 6 MRC Centre for Global Infectious Disease Analysis, Imperial College London, London, UK. 7Infection Control Programme, Geneva University Hospitals and Faculty of Medicine, Geneva, Switzerland. 8Department of Rehabilitation and Geriatrics, Geneva University Hospitals and Faculty of Medicine, Geneva, Switzerland. 9Biological Sciences, University of Edinburgh, Edinburgh, UK. 10School of Informatics, University of Edinburgh, Edinburgh, UK.
*m.barahona@imperial.ac.uk
Abstract[a]
Background. Real-time disease prediction is key to prevent and control healthcare-associated infections. Yet current prediction frameworks fail to fully capture the dynamic, contact-driven nature of infectious diseases. We develop a machine learning framework that incorporates dynamic patient contact networks and perform testing and validation on international multi-site data to predict and identify dynamic risk factors of hospital-onset COVID-19 infection (HOCI).
Methods. Our framework uses both dynamic contact network variables and traditional variables (patient attributes and background contextual measurements) extracted from routinely collected hospital data. We train and test the framework to predict HOCIs using 51,157 hospital patients admitted to a large London NHS Trust covering UK COVID-19 surges 1 and 2. We then apply the model to additional post surge 2 data from the same hospital group (total 43,375 inpatients) and to data from a non-UK site (Geneva COVID-19 surge 1; 40,057 total inpatients).
Findings. The framework achieved high predictive performance using all variables (AUC-ROC 0.89, (0.88-0.90; 95% CI)) was almost as predictive using only contact network variables (AUC-ROC 0.88, (0.86-0.90; 95% CI)), and more so than using hospital contextual (AUC-ROC 0.82, (0.80-0.84; 95% CI)) or clinical (AUC-ROC 0.64 (0.62-0.66); 95% CI)) variables. Through feature elimination, we established a highly predictive model (AUC-ROC 0.88 (0.86-0.90; 95% CI)) with only six variables: five contact network variables (including ‘network closeness' to all recently infectious patients and number of direct contacts) and one contextual variable (background hospital disease prevalence). This reduced model was applied to post surge 2 data from the London site (AUC-ROC 0.68 (0.64-0.70; 95% CI)[b]) and non-UK site (AUC-ROC 0.88 (0.86–0.90); 95% CI), validating the generalisability of both the modelling framework and HOCI risk factors.
Interpretation. Our analysis supports the inclusion of dynamic contact network centrality variables, accessible from routinely collected electronic health records, for forecasting hospital infections. Given the generalisability and flexibility of our framework, its integration into clinical care can drastically improve personalised infection prevention[c].
Funding. Medical Research Foundation, World Health Organisation, Engineering and Physical Sciences Research Council, National Institute for Health Research.[d]
Introduction
Healthcare transmission of coronavirus disease 2019 (COVID-19) has been well documented throughout the pandemic1. Reports have cited hospital-onset COVID-19 infections (HOCIs) accounting for 12–15% of all COVID-19 cases identified in healthcare settings2, and as high as 16.2% at the peaks of the pandemic3. While their impact is yet to be fully quantified, HOCIs amplify the impact of the pandemic by seeding further outbreaks.
The ability to predict HOCIs can dramatically impact prevention and control reducing both illness and the workload burden experienced during outbreaks4. Traditionally, prediction has involved the identification of patient risk factors of disease acquisition by fitting statistical models to a combination of patient clinical variables (e.g., age, gender identity, co-morbidities) and hospital contextual variables (e.g., colonisation pressure, patient length of stay)5. Although such risk factors perform reasonably well, they however ignore the fact that the spread of infectious diseases depends largely on patient contacts6, which are heterogenous7 and variable over time8.
Isolation and cohorting of infected patients prevent onward spreading by interrupting transmission chains9.I Investigations of contacts to known infected patients have been an effective epidemiological tool for identifying at-risk secondary cases and disease ‘super-spreaders’10, and have played a pivotal role in national COVID-19 responses11–13. However, considering the full contact network (not just the direct links to known infections)provides greater depth to characterise disease spreading14. For example, early in the COVID-19 pandemic, population mobility and interactions guided national policy to reduce transmission15. In healthcare settings, the total number of direct contacts created by patient transfers has been found predictive of disease acquisition16–19. Yet these studies fail to take advantage of the rich predictive power present in the full dynamic contact networks to describe transmission routes20.
...