Development of a model to predict hospital admission and severe outcome in cancer patients with COVID-19


Year:

Session type:

Rohan Shotton, Cong Zhou, Rebecca Lee, Ann Tivey, Talvinder Bhogal, Oskar Wysoki, Elena Dickens, Prerana Huddar, Hayley McKenzie, Hayley Boyce, Alec Maynard, Michael Rowe, Sam Khan, Leonie Eastlake, Angelos Angelakas, Mark Baxter, Ellen Copson, Laura Horsley, Anne Thomas, Caroline Wilson, Tim Cooksley, Andre Freitas, Carlo Palmieri, Caroline Dive, Anne Armstrong

Abstract

Background

COVID-19 represents a spectrum of clinical syndromes ranging from subclinical infection to death. Patients with cancer are at increased risk of severe COVID-19 infection and adverse outcomes. This novel clinical problem presents a challenge when making decisions regarding admission or discharge of patients with suspected COVID-19. To aid this process, we developed a model to guide which patients should be admitted versus discharged at presentation to hospital, and which patients were likely to experience severe illness.

Method

Consecutive patients with cancer and confirmed COVID-19 infection (PCR-detected) were identified in 12 UK hospitals. A database of patients’ background clinical and presenting laboratory data was created. Clinicopathological data were mapped to clinical outcomes, stratified by discharge within 24 hours of admission with no further complications, need for supplementary oxygen (O2), and death. Logistic regression and random forest methods were applied for model development.

Results

299 consecutive patients were registered from a mix of general hospitals and tertiary oncology centres. Haematological cancer patients showed significantly higher rate of death compared to patients with solid tumours. Albumin, C-reactive protein (CRP), neutrophils and platelets were all independently associated with COVID-19 severity and outcomes. The random forest model showed improved predictive ability compared to logistic regression. It successfully characterised the severity of COVID-19 for cancer patients using nine clinical features (CRP, National Early Warning Score 2 (NEWS2), albumin, age, neutrophils, platelets, number of comorbidities, cancer stage and haematological cancer). The most important drivers in predicting infection severity were NEWS2, CRP and albumin. Considerable variation in clinician decision-making was observed between centres. The random forest model achieved an average true positive prediction rate of 77.4% (+/- 9.9%) and a false positive rate of 20.7% (+/- 27.4%). None of the cases who were admitted despite the model predicting discharge subsequently required supplementary oxygen or died.

Conclusion

We present a model to aid clinical decision-making by identifying patients with cancer and suspected COVID-19 who are suitable for early discharge, and those at risk of severe infection. An online decision support tool is in development.

Impact statement

We present a model to predict which patients with cancer and COVID-19 are at risk of severe infection.