Donal Landers Dr, Prerana Huddar Dr, Tim Cooksley Dr, Hayley Boyce Dr, Cong Zhou Dr, Hayley McKenzie Dr, Caroline Wilson Dr, Umair T Khan Dr, Jamie Weaver Dr, Anne C Armstrong Dr, Michael Rowe Dr, Kathryn Banfill Dr, Angelos Angelakas Dr, Alec Maynard Dr, Paul Fitzpartick Dr, Joshua Woodcock Dr, Theingi Aung Dr, Anne Thomas Prof, Christina Hague Dr, Rohan Shotton Dr, Donna Graham Dr, Sophie Williams Dr, Sam Khan Dr, Rebecca J Lee Dr, Louise Lever Dr, Roseleen Sheehan Dr, Talvinder Bhogal Dr, Lance Turtle, Caroline Dive Prof, Tim Robinson Dr, Ellen Copson Dr, Richard Hoskins Dr, Hannah Frost Ms, Julie Stevenson Dr, Andre Freitas Dr, Elena Dickens Dr, Leonie Eastlake Dr, Mark Baxter Dr, Laura Horsley Dr, Oskar Wysocki Dr, Fabio Gomes Dr, Michelle Harrison Dr, Zoe Hudson Dr, Alexander J. Stockdale, Ann Tivey Dr, and Carlo Palmieri Prof
BackgroundCancer patients are at increased risk of severe COVID-19. As COVID-19 presentation and outcomes are heterogeneous in cancer patients, decision-making tools for hospital admission, severity prediction and increased monitoring for early intervention are critical.ObjectiveTo identify features of COVID-19 in cancer patients predicting severe disease and build a decision-support online tool; COVID-19 Risk in Oncology Evaluation Tool (CORONET)MethodData was obtained for consecutive patients with active cancer with laboratory confirmed COVID-19 presenting in 12 hospitals throughout the United Kingdom (UK). Univariable logistic regression was performed on pre-specified features to assess their association with admission (≥24 hours inpatient), oxygen requirement and death. Multivariable logistic regression and random forest models (RFM) were compared with patients randomly split into training and validation sets. Cost function determined cut-offs were defined for admission/death using RFM. Performance was assessed by sensitivity, specificity and Brier scores (BS). The CORONET model was then assessed in the entire cohort to build the online CORONET tool.ResultsTraining and validation sets comprised 234 and 66 patients respectively with median age 69 (range 19-93), 54% males, 46% females, 71% vs 29% had solid and haematological cancers. The RFM, selected for further development, demonstrated superior performance over logistic regression with AUROC predicting admission (0.85 vs. 0.78) and death (0.76 vs. 0.72). C-reactive protein was the most important feature predicting COVID-19 severity. CORONET cut-offs for admission and mortality of 1.05 and 1.8 were established. In the training set, admission prediction sensitivity and specificity were 94.5% and 44.3% with BS 0.118; mortality sensitivity and specificity were 78.5% and 57.2% with BS 0.364. In the validation set, admission sensitivity and specificity were 90.7% and 42.9% with BS 0.148; mortality sensitivity and specificity were 92.3% and 45.8% with BS 0.442. In the entire cohort, the CORONET decision support tool recommended admission of 99% of patients requiring oxygen and of 99% of patients who died.Conclusions and RelevanceCORONET, a decision support tool validated in hospitals throughout the UK showed promise in aiding decisions regarding admission and predicting COVID-19 severity in patients with cancer presenting to hospital. Future work will validate and refine the tool in further datasets.