Machine learning infrastructure and methods to support patient care in novel disease pandemics

Thesis

The COVID-19 pandemic exposed critical vulnerabilities in global healthcare systems, particularly in managing novel disease outbreaks. As future pandemics loom, integrating advanced technologies like Machine Learning (ML) offers a promising path to enhance preparedness and patient care. This thesis utilized large Electronic Health Records (EHR) datasets to develop and evaluate ML models for COVID- 19 patient care and pandemic response. We first assessed heuristic Early Warning Score (EWS) systems. We found that existing EWS systems had limited predictive performance in COVID-19 patients, highlighting the need for more sophisticated models. Next, we explored ML-based EWS systems, developing novel models that outperformed traditional methods in predicting patient deterioration. We also tackled data scarcity by using transfer learning, enabling effective model training with non-COVID-19 data. Lastly, we proposed and discussed frameworks for the scalable validation and implementation of clinical ML models, emphasizing continuous adaptation to account for data shifts and variability as a method to ensure ML tools remain effective in dynamic healthcare settings. This thesis advances our understanding of ML’s role in pandemic preparedness and the infrastructure required for its scalable deployment.

Authors: Youssef, A

• DOI: 10.5287/ora-azr9m0mwr
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: pandemic; machine learning; early warning scores; COVID19; electronic health records
• Subjects: Medical records--Data processing; COVID-19 Pandemic, 2020-2023; COVID-19 (Disease)