Machine learning predicts hospital-onset COVID-19 infections using patient contact networks

Accurate and real-time disease prediction is vital for the prevention and control of healthcare-related infections. Although contacts between individuals are primarily responsible for infection chains, most prediction frameworks do not capture the contact dynamics.

A real-time framework using patient contact networks to predict hospital-onset COVID-19 infections

Researchers from the UK recently developed a real-time machine learning framework that uses dynamic patient contact networks to predict hospital-onset COVID-19 infections (HOCIs) at the patient level. They then tested and validated the framework on international multi-site datasets across various epidemic and endemic periods. This study can be found on the medRxiv* preprint server.

The framework extracts dynamic contact networks from hospital data collected routinely and combines them with clinical attributes and background contextual hospital data of the patient to predict the infection status of patients. The research team trained and tested the HOCI framework using over 50,000 patients admitted to a UK National Health Service (NHS) Trust hospital between 01 April 2020 and 01 April 2021, thus spanning COVID-19 surges 1 and 2 in the UK.

They then validated the HOCI framework by applying it to data from 40,057 inpatients gathered from a hospital site in Geneva during an epidemic surge and data from 43,375 inpatients collected from the same NHS Trust from a subsequent period after surge 2, when COVID-19 had become endemic.

Overview of forecasting framework. Patient pathways are extracted from electronic health records which specify the locations each patient has visited over the duration of their hospital stay. Pathways are overlaid with COVID-19 testing results, capturing the space-time positions of patients that tested positive for COVID-19. Forecasting is based on extracting individual patient clinical variables (fixed) and hospital contextual variables (dynamic) during a defined time window, as well as variables capturing the centrality of a patient within the different contact networks (dynamic). We iterate variable extraction over multiple time windows and use the cumulative information for model training and predictions.
Overview of forecasting framework. Patient pathways are extracted from electronic health records which specify the locations each patient has visited over the duration of their hospital stay. Pathways are overlaid with COVID-19 testing results, capturing the space-time positions of patients that tested positive for COVID-19. Forecasting is based on extracting individual patient clinical variables (fixed) and hospital contextual variables (dynamic) during a defined time window, as well as variables capturing the centrality of a patient within the different contact networks (dynamic). We iterate variable extraction over multiple time windows and use the cumulative information for model training and predictions.

The HOCI framework demonstrated high predictive performance while using only contact network variables

Based on the training data from London spanning COVID-19 surges 1 and 2, the HOCI framework showed high predictive performance using all variables but was almost equally predictive using only contact network variables and more predictable than using only hospital contextual or patient clinical variables.

Background hospital infections and contact structure across the study period. Daily new COVID-19 positive patients within the hospital (COCI and HOCI) varied dramatically across the study period. Peaks of 59 and 64 cases were reached on March 30, 2020, and January 06, 2021 respectively, whilst dipping to zero new daily cases over days during July, August, September, and October. The patient contact network also varied across the study period, panels i-iii, with differences in connectivity and size of patient contact clusters between each of the infection surges and during the summer period.
Background hospital infections and contact structure across the study period. Daily new COVID-19 positive patients within the hospital (COCI and HOCI) varied dramatically across the study period. Peaks of 59 and 64 cases were reached on March 30, 2020, and January 06, 2021 respectively, whilst dipping to zero new daily cases over days during July, August, September, and October. The patient contact network also varied across the study period, panels i-iii, with differences in connectivity and size of patient contact clusters between each of the infection surges and during the summer period.

The top 3 risk factors identified by the team comprised one hospital contextual variable, namely, background hospital COVID-19 prevalence, and two contact network variables, namely, the number of direct contacts of infectious patients and network closeness. Moreover, the addition of contact network variables boosted performance related to hospital contextual variables on both the UK and non-UK validation datasets.

“Our study highlights the predictive power that can be mined from networks of patient contacts to aid with personalized predictions of infectious disease in healthcare settings.”

Findings show that integration of patient contact networks in clinical care could improve infection prevention and help early diagnosis

According to the authors, this is the first study to predict individual patient HOCIs using routine patient and hospital data and dynamic contact networks. The research team combined machine learning with network analysis and developed a HOCI prediction framework with the help of routine electronic health data.

Epidemiology curves of validation datasets. Newly identified COVID-19 cases are reported across time and are broken down by HOCI and COCI case types. Panel A shows the non-UK (Geneva) hospital case load during an epidemic surge of cases. Panel B shows the UK hospital group post pandemic surges 1 and 2, when COVID-19 became endemic and non-surging.
Epidemiology curves of validation datasets. Newly identified COVID-19 cases are reported across time and are broken down by HOCI and COCI case types. Panel A shows the non-UK (Geneva) hospital case load during an epidemic surge of cases. Panel B shows the UK hospital group post pandemic surges 1 and 2, when COVID-19 became endemic and non-surging.

The results showed that dynamic patient contact networks could be a strong predictor of transmission of respiratory viral infections in hospitals. Thus, the integration of these networks in clinical care could improve individualized prevention of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection as well as early diagnosis.

“The framework provides daily patient-individualized predictions and can identify dynamic disease acquisition risk factors.”

They found that centrality measures calculated from patient contact networks together with some hospital contextual variables represent a key HOCI risk factor, which can boost the performance of predictive models in a pandemic as well as endemic conditions.

To demonstrate the transferability of the framework, they applied it to data from a Geneva-based hospital during COVID-19 surge 1 and showed increased predictive power by including contact network risk factors. However, access was only available for room-level contacts that were less predictive in the London data.

With no end in sight to the COVID-19 pandemic, this new predictive framework and contact risk factors offer a valuable tool to assist identification of potential HOCI cases across various epidemic and endemic periods.

“Further work will be needed to extend this work to other healthcare-acquired infections by assessing how the inclusion of variables capturing a patient’s environment within an underlying contact network could aid prediction with a view to informing infection prevention and control measures.”

*Important Notice

medRxiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.

Journal reference:
Susha Cheriyedath

Written by

Susha Cheriyedath

Susha has a Bachelor of Science (B.Sc.) degree in Chemistry and Master of Science (M.Sc) degree in Biochemistry from the University of Calicut, India. She always had a keen interest in medical and health science. As part of her masters degree, she specialized in Biochemistry, with an emphasis on Microbiology, Physiology, Biotechnology, and Nutrition. In her spare time, she loves to cook up a storm in the kitchen with her super-messy baking experiments.

Citations

Please use one of the following formats to cite this article in your essay, paper or report:

  • APA

    Cheriyedath, Susha. (2021, October 03). Machine learning predicts hospital-onset COVID-19 infections using patient contact networks. News-Medical. Retrieved on October 15, 2021 from https://www.news-medical.net/news/20211003/Machine-learning-predicts-hospital-onset-COVID-19-infections-using-patient-contact-networks.aspx.

  • MLA

    Cheriyedath, Susha. "Machine learning predicts hospital-onset COVID-19 infections using patient contact networks". News-Medical. 15 October 2021. <https://www.news-medical.net/news/20211003/Machine-learning-predicts-hospital-onset-COVID-19-infections-using-patient-contact-networks.aspx>.

  • Chicago

    Cheriyedath, Susha. "Machine learning predicts hospital-onset COVID-19 infections using patient contact networks". News-Medical. https://www.news-medical.net/news/20211003/Machine-learning-predicts-hospital-onset-COVID-19-infections-using-patient-contact-networks.aspx. (accessed October 15, 2021).

  • Harvard

    Cheriyedath, Susha. 2021. Machine learning predicts hospital-onset COVID-19 infections using patient contact networks. News-Medical, viewed 15 October 2021, https://www.news-medical.net/news/20211003/Machine-learning-predicts-hospital-onset-COVID-19-infections-using-patient-contact-networks.aspx.

Comments

The opinions expressed here are the views of the writer and do not necessarily reflect the views and opinions of News Medical.
You might also like... ×
COVID-19 booster for immunosuppressed patients with inflammatory bowel disease