Classification and analysis of outcome predictors in non-critically ill COVID-19 patients

Abstract

Background: Early detection of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)-infected patients who could develop a severe form of COVID-19 must be considered of great importance to carry out adequate care and optimise the use of limited resources.

Aims: To use several machine learning classification models to analyse a series of non-critically ill COVID-19 patients admitted to a general medicine ward to verify if any clinical variables recorded could predict the clinical outcome.

Methods: We retrospectively analysed non-critically ill patients with COVID-19 admitted to the general ward of the hospital in Pordenone from 1 March 2020 to 30 April 2020. Patients' characteristics were compared based on clinical outcomes. Through several machine learning classification models, some predictors for clinical outcome were detected.

Results: In the considered period, we analysed 176 consecutive patients admitted: 119 (67.6%) were discharged, 35 (19.9%) dead and 22 (12.5%) were transferred to intensive care unit. The most accurate models were a random forest model (M2) and a conditional inference tree model (M5) (accuracy = 0.79; 95% confidence interval 0.64-0.90, for both). For M2, glomerular filtration rate and creatinine were the most accurate predictors for the outcome, followed by age and fraction-inspired oxygen. For M5, serum sodium, body temperature and arterial pressure of oxygen and inspiratory fraction of oxygen ratio were the most reliable predictors.

Conclusions: In non-critically ill COVID-19 patients admitted to a medical ward, glomerular filtration rate, creatinine and serum sodium were promising predictors for the clinical outcome. Some factors not determined by COVID-19, such as age or dementia, influence clinical outcomes.

Keywords: COVID-19; machine learning; non-critically ill; prediction.

Figure 1

Study design: 176 consecutive COVID‐19 positive patients (pts) from emergency department were retrospectively admitted to the COVID‐19 or non‐COVID‐19 ward and then classified by the outcome in both branches. ED, emergency department; ICU, intensive care unit.

Figure 2

Classic classification decision tree (‘rpart’ package) (M4). A creatinine greater or less than 1.2 mg/dL (106.1 μmol/L), a white blood cell count greater or less than 11 000/mL and the age greater than 79 years are the main determinants of the patient's clinical outcome. This model's accuracy is 65% (95% confidence interval 49–59%) (see the main text for details). ICU, intensive care unit; WBC, white blood cells.

Figure 3

Recursive partitioning regression tree model using permutation tests to hierarchise predictor variables (‘party’ package) (M5). The plasma sodium level (i.e. if greater or less than 144 mEq/L) represents the first predictive node. The following are body temperature (i.e. if higher or lower than 37.7°C), hospitalisation in the ward dedicated to COVID‐19 patients only and the arterial pressure of oxygen and inspiratory fraction of oxygen ratio (PaO₂/FiO₂) higher or lower than 112. The probability percentages for (from left to right) death, discharge or transfer to intensive care (ICU), are reported at the bottom of the decision tree. The accuracy of this model is 79% (95% confidence interval 64–90%). For further details, see the main text.

Figure 4

Classification of predictive factors based on the random forest model (M2). On the left, the predictive factors are reported in descending order from top to bottom according to the accuracy, while on the right, they are shown in descending order from top to bottom according to the Gini coefficient (which measures the inequality among values of a frequency distribution). The accuracy of this model is 79% (95% confidence interval 64–90%). For further details, see the main text.