Contrasting factors associated with COVID-19-related ICU admission and death outcomes in hospitalised patients by means of Shapley values

Author summary The design is a retrospective cohort study of 13954 in-patients of ages ranging from 1 to 105 year (IQR: 56, 70, 81) with a confirmed diagnosis of COVID-19 by 28th June 2020. This study used multivariable logistic regression to generate odd ratios (ORs) multiply adjusted for 37 covariates (comorbidities, demographic, and others) selected on the basis of clinical interest and prior findings. Results were supplemented by gradient-boosted decision tree (GBDT) classification to generate Shapley values in order to evaluate the impact of the covariates on model output for all patients. Factors are differentially associated with death and ICUA and among patients. Deaths due to COVID-19 were associated with immunosuppression due to disease (OR 1.39, 95% CI 1.10-1.76), type-2 diabetes (OR 1.31, 95% CI 1.17-1.46), chronic respiratory disease (OR 1.19, 95% CI 1.05-1.35), age (OR 1.56/10-year increment, 95% CI 1.51-1.61), and male sex (OR 1.54, 95% CI 1.42-1.68). Associations of ICUA with some factors differed in direction (e.g., age, chronic respiratory disease). Self-reported ethnicities were strongly but variably associated with both outcomes. GBDTs had similar performance (ROC-AUC, ICUA 0.83, death 0.68 for GBDT; 0.80 and 0.68 for logistic regression). We derived importance scores based on Shapley values which were consistent with the ORs, despite the underlying machine-learning model being intrinsically different to the logistic regression. Chronic heart disease, hypertension, other comorbidities, and some ethnicities had Shapley impacts on death ranging from positive to negative among different patients, although consistently associated with ICUA for all. Immunosuppressive disease, type-2 diabetes, and chronic liver and respiratory diseases had positive impacts on death with either positive or negative on ICUA. We highlight the complexity of informing clinical practice and public-health interventions. We recommend that clinical support systems should not only predict patients at risk, but also yield interpretable outputs for validation by domain experts. Identification of those at greatest risk of death due to the substantial threat of COVID-19 can benefit from novel approaches to epidemiology that leverage large datasets and complex machine-learning models, provide data-driven intelligence, and guide decisions such as intensive-care unit admission (ICUA). The objective of this study is two-fold, one substantive and one methodological: substantively to evaluate the association of demographic and health records with two related, yet different, outcomes of severe COVID-19 (viz., death and ICUA); methodologically to compare interpretations based on logistic regression and on gradient-boosted decision tree (GBDT) predictions interpreted by means of the Shapley impacts of covariates. Very different association of some factors, e.g., obesity and chronic respiratory diseases, with death and ICUA may guide review of practice. Shapley explanation of GBDTs identified varying effects of some factors among patients, thus emphasising the importance of individual patient assessment. The results of this study are also relevant for the evaluation of complex automated clinical decision systems, which should optimise prediction scores whilst remaining interpretable to clinicians and mitigating potential biases.