Prediction of COVID-19 Mortality in Patients with Cancer

Background: Coronavirus disease 2019 (COVID-19) is associated with an increased mortality rate in patients with cancer. Machine learning approaches have been used successfully to predict mortality in infected patients from the general population. We aimed to devise an algorithm to estimate the risk of death in the first 45 days after COVID-19 diagnosis for patients with cancer.

Methods: As a first step we used the ensemble learning method random survival forests. Computations were performed with R packages randomForestSRC, ROCR and cvAUC. The cohort consisted of Memorial Sloan Kettering patients diagnosed with COVID-19 before June 12, 2020 (n=820). The primary endpoint was mortality at 45 days after diagnosis of COVID-19. Patients were included in the study regardless of whether or not they were admitted to the hospital. Predictors assessed at the time of COVID-19 diagnosis and retained for the model include: cancer type (thoracic, hematologic or other), metastatic status, age, sex, smoking, presence of a comorbidity (any of hypertension, diabetes, hyperlipidemia, chronic hepatic or renal disease, coronary artery disease, asthma, thromboembolism, chronic obstructive lung disease, congestive heart failure, valvular heart disease, stroke, HIV and peripheral arterial disease), body mass index (BMI), presence of hypoxia (defined as need to use supplemental oxygen 2 L/min or more), blood glucose, total bilirubin, total white blood count (WBC), hemoglobin, red cell distribution width (RDW), platelet count, absolute neutrophil count (ANC), absolute lymphocyte count, serum albumin, blood urea nitrogen (BUN) , creatinine, aspartate transaminase (AST), alanine transaminase (ALT), ferritin, lactate dehydrogenase (LDH), prothrombin time (PT), activated partial thromboplastin time (aPTT), D-dimers, race and smoking history. Model performance was assessed with repeated 10-fold cross-validation.

Results: 11% (n=91) of patients had died by day #45 of observation. The prevalence of hematologic malignancy in the cohort at large was 26% (n=210), compared with 9% (n=74) for primary thoracic neoplasms. Metastatic solid cancer was present in 40% (n=332) of individuals. Information on cell counts and RDW was complete for 51% of patients, compared to 43% for liver function tests, 54% for renal function measurements, 23% for PT, 22% for PTT and 7% for D-dimers. Records were manually reviewed and assessed successfully in all patients for cancer type, metastatic status, presence of a comorbidity, presence of hypoxia and smoking history. Only 23% of patients were admitted at the time of COVID-19 diagnosis. The mean cross-validated area under the receiver operating characteristic (ROC) curve for a model including all available covariates was 0.81 (95% confidence interval = 0.80-0.82). The averaged ROC curve is shown in Figure A. Comorbidities, BUN, serum albumin, age, hypoxia, cancer type, RDW, hemoglobin, platelet count, metastatic status and smoking history all contributed significantly to the model, as measured with the Ishwaran-Kogalur Variable Importance (VIMP). Scaled VIMP values are reported in Figure B.

Conclusions: COVID-19 is associated with substantial mortality in patients with cancer. As observed in prior reports of patients from the general population, the presence of comorbidities, age, BUN and hypoxia at presentation are highly predictive of mortality in this setting. Interestingly, serum albumin appears to be an important predictor in patients with cancer, as are cancer-specific covariates including tumor type and metastatic status. The latter suggests that optimal predictive models of COVID-19 mortality in oncology patients will differ substantially from those trained on more general medical cohorts. Additional work remains to be done in order to validate this model before its application in clinical practice.

View largeDownload slide

View largeDownload slide

Close modal