Development of Machine Learning Prediction Models to Predict ICU Admission and the Length of Stay in ICU for COVID‑19 Patients Using a Clinical Dataset Including Chest Computed Tomography Severity Score Data

Development of ML Models to Predict ICU Admission and LOS for COVID‑19 Patients

Abstract

Objective: The chest computed tomography severity score (CT-SS) is significantly associated with the severity of the disease and subsequently intensive care unit (ICU) admission in coronavirus disease-19 (COVID-19) patients. However, there was a lack of information about the prognostic role of radiological manifestations in combination with demographics, clinical manifestations, and laboratory predictors to predict ICU admission and the length of stay (LOS) in the ICU (ICU LOS) of COVID-19 patients. The machine learning (ML) approach is a new and, non-invasive digital technology that can present an efficient risk prediction model for clinical problems. The purpose of the present study was to develop an effective ML model for predicting ICU admission and ICU LOS for COVID-19 patients using a more comprehensive dataset including imaging findings.
Methods: A COVID-19 hospital-based registry database that contained medical records of 6,854 patients was retrospectively reviewed. The incomplete records with missing values of more than 70% were excluded, and the remaining missing values were imputed using the mean and mode values for the continuous and discrete variables, respectively. The imbalance in the data numbers of groups was resolved using the synthetic minority over-sampling technique algorithm. Two sets of prediction models were separately developed to predict ICU admission and ICU LOSs of COVID‑19 patients. The most important and related predictors selected by the Boruta feature selection method were used to develop ML prediction models. The parameters obtained from the confusion matrix were used to evaluate the performance of the prediction models. The performance evaluation of the developed ML models for predicting ICU LOS of the patients utilized correlation coefficient, mean absolute error, and root mean squared error metrics.
Results: The records of 815 positive reverse transcription polymerase chain reaction (RT-PCR) patients were included in the study after applying the inclusion/exclusion criteria. Of the 815 positive RT-PCR patients, only 185 patients were admitted to the ICU to receive intensive care. The number of records in the ICU admission group was raised to 630 to deal with the data imbalance problem. For predicting the ICU admission of COVID-19 patients, k-nearest neighbors (k-NN) yielded better performance than J48, support vector machine, multi-layer perceptron, Naïve Bayes, logistic regression, random forest (RF), and XGBoostbased ML models. The sensitivity, specificity, accuracy, precision, F-measure, and area under the curve of the k-NN algorithm were 97.0%, 89.7%, 93.3%, 90.4%, 93.6%, and 99.1%, respectively. Results showed that with a correlation coefficient of 0.42, a mean absolute error of 2.01, and a root mean squared error of 4.11, the RF algorithm with a correlation coefficient of 0.42, mean absolute error of 2.01, and root mean squared error of 4.11demonstrated the best performance in predicting the ICU LOS of COVID-19 patients.
Conclusion: The ML approach, utilizing a more comprehensive dataset that includes CT-SS, could efficiently predict ICU admission and ICU LOS of COVID-19 patients. Timely prediction of ICU admission and ICU LOS of COVID-19 patients would improve patient outcomes and lead to the optimal use of limited hospital resources.