ABSTRACT
Background and objectives: We aim to verify the use of ML algorithms to predict patient outcome using a relatively small dataset and to create a nomogram to assess in-hospital mortality of patients with COVID-19. Methods: A database of 200 COVID-19 patients admitted to the Clinical Hospital of State University of Campinas (UNICAMP) was used in this analysis. Patient features were divided into three categories: clinical, chest abnormalities, and body composition characteristics acquired by computerized tomography. These features were evaluated independently and combined to predict patient outcomes. To minimize performance fluctuations due to low sample number, reduce possible bias related to outliers, and evaluate the uncertainties generated by the small dataset, we developed a shuffling technique, a modified version of the Monte Carlo Cross Validation, creating several subgroups for training the algorithm and complementary testing subgroups. The following ML algorithms were tested: random forest, boosted decision trees, logistic regression, support vector machines, and neural networks. Performance was evaluated by analyzing Receiver operating characteristic (ROC) curves. The importance of each feature in the determination of the outcome predictability was also studied and a nomogram was created based on the most important features selected by the exclusion test. Results: Among the different sets of features, clinical variables age, lymphocyte number and weight were the most valuable features for prognosis prediction. However, we observed that skeletal muscle radiodensity and presence of pleural effusion were also important for outcome determination. Integrating these independent predictors was successfully developed to accurately predict mortality in COVID-19 in hospital patients. A nomogram based on these five features was created to predict COVID-19 mortality in hospitalized patients. The area under the ROC curve was 0.86 ± 0.04. Conclusion: ML algorithms can be reliable for the prediction of COVID-19-related in-hospital mortality, even when using a relatively small dataset. The success of ML techniques in smaller datasets broadens the applicability of these methods in several problems in the medical area. In addition, feature importance analysis allowed us to determine the most important variables for the prediction tasks resulting in a nomogram with good accuracy and clinical utility in predicting COVID-19 in-hospital mortality.
ABSTRACT
Inflammatory states and body composition changes are associated with a poor prognosis in many diseases, but their role in coronavirus disease 2019 (COVID-19) is not fully understood. To assess the impact of low skeletal muscle radiodensity (SMD), high neutrophil-to-lymphocyte ratio (NLR) and a composite score based on both variables, on complications, use of ventilatory support, and survival in patients with COVID-19. Medical records of patients hospitalized between May 1, 2020, and July 31, 2020, with a laboratory diagnosis of COVID-19 who underwent computed tomography (CT) were retrospectively reviewed. CT-derived body composition measurements assessed at the first lumbar vertebra level, and laboratory tests performed at diagnosis, were used to calculate SMD and NLR. Prognostic values were estimated via univariate and multivariate logistic regression analyses and the Kaplan-Meier curve. The study was approved by the local Institutional Review Board (CAAE 36276620.2.0000.5404). A total of 200 patients were included. Among the patients assessed, median age was 59 years, 58% were men and 45% required ICU care. A total of 45 (22.5%) patients died. Multivariate logistic analysis demonstrated that a low SMD (OR 2.94; 95% CI 1.13-7.66, P = 0.027), high NLR (OR 3.96; 95% CI 1.24-12.69, P = 0.021) and both low SMD and high NLR (OR 25.58; 95% CI 2.37-276.71, P = 0.008) combined, were associated with an increased risk of death. Patients who had both low SMD and high NLR required more mechanical ventilation (P < 0.001) and were hospitalized for a longer period (P < 0.001). Low SMD, high NLR and the composite score can predict poor prognosis in patients with COVID-19, and can be used as a tool for early identification of patients at risk. Systemic inflammation and low muscle radiodensity are useful predictors of poor prognosis, and the assessment of these factors in clinical practice should be considered.