Predictive determinants of overall survival among re-infected COVID-19 patients using the elastic-net regularized Cox proportional hazards model: a machine-learning algorithm.

ABSTRACT

BACKGROUND: Narrowing a large set of features to a smaller one can improve our understanding of the main risk factors for in-hospital mortality in patients with COVID-19. This study aimed to derive a parsimonious model for predicting overall survival (OS) among re-infected COVID-19 patients using machine-learning algorithms. METHODS: The retrospective data of 283 re-infected COVID-19 patients admitted to twenty-six medical centers (affiliated with Shiraz University of Medical Sciences) from 10 June to 26 December 2020 were reviewed and analyzed. An elastic-net regularized Cox proportional hazards (PH) regression and model approximation via backward elimination were utilized to optimize a predictive model of time to in-hospital death. The model was further reduced to its core features to maximize simplicity and generalizability. RESULTS: The empirical in-hospital mortality rate among the re-infected COVID-19 patients was 9.5%. In addition, the mortality rate among the intubated patients was 83.5%. Using the Kaplan-Meier approach, the OS (95% CI) rates for days 7, 14, and 21 were 87.5% (81.6-91.6%), 78.3% (65.0-87.0%), and 52.2% (20.3-76.7%), respectively. The elastic-net Cox PH regression retained 8 out of 35 candidate features of death. Transfer by Emergency Medical Services (EMS) (HR=3.90, 95% CI 1.63-9.48), SpO2≤85% (HR=8.10, 95% CI 2.97-22.00), increased serum creatinine (HR=1.85, 95% CI 1.48-2.30), and increased white blood cells (WBC) count (HR=1.10, 95% CI 1.03-1.15) were associated with higher in-hospital mortality rates in the re-infected COVID-19 patients. CONCLUSION: The results of the machine-learning analysis demonstrated that transfer by EMS, profound hypoxemia (SpO2≤85%), increased serum creatinine (more than 1.6 mg/dL), and increased WBC count (more than 8.5 (×109 cells/L)) reduced the OS of the re-infected COVID-19 patients. We recommend that future machine-learning studies should further investigate these relationships and the associated factors in these patients for a better prediction of OS.