Utilizing machine learning for survival analysis to identify risk factors for COVID-19 intensive care unit admission: A retrospective cohort study from the United Arab Emirates

Abstract

Background: The current situation of the unprecedented COVID-19 pandemic leverages Artificial Intelligence (AI) as an innovative tool for addressing the evolving clinical challenges. An example is utilizing Machine Learning (ML) models-a subfield of AI that take advantage of observational data/Electronic Health Records (EHRs) to support clinical decision-making for COVID-19 cases. This study aimed to evaluate the clinical characteristics and risk factors for COVID-19 patients in the United Arab Emirates utilizing EHRs and ML for survival analysis models.

Methods: We tested various ML models for survival analysis in this work we trained those models using a different subset of features extracted by several feature selection methods. Finally, the best model was evaluated and interpreted using goodness-of-fit based on calibration curves,Partial Dependence Plots and concordance index.

Results: The risk of severe disease increases with elevated levels of C-reactive protein, ferritin, lactate dehydrogenase, Modified Early Warning Score, respiratory rate and troponin. The risk also increases with hypokalemia, oxygen desaturation and lower estimated glomerular filtration rate and hypocalcemia and lymphopenia.

Conclusion: Analyzing clinical data using AI models can provide vital information for clinician to measure the risk of morbidity and mortality of COVID-19 patients. Further validation is crucial to implement the model in real clinical settings.

Fig 1. Flow diagram of the cohort.

Inclusion and exclusion criteria for determining patients consider in this study. SEHA extracted 1,800 COVID-19 patients from March 1, 2020 to April 20, 2020. After applying the inclusion and exclusion criteria, our study included 1787 patients.

Fig 2. Models hyperparameters tuning: Heatmap of mean C-index (5-fold cross validation) of the best combination of the hyperparameters and feature selections methods; also we tune the threshold for selecting the highest importance features.

Fig 3

Calibration curves and time-dependent ROC curve of the gradient boosting machine (GBM) model: Calibration curves of predicted compared with observed ICU access after 2 Days (a), 3 Days (b) and 5 Days (c) of hospital administration. Time-dependent ROC curve for the ICU admission predicting after after 2 Days (d),3 Days (e) and 5 Days (f) of hospital administration.

Fig 4. Partial dependence plots inferred from gradient boosting machine (GBM) model using random forest minimal depth features subset: The lines present the change in the risk to access the ICU across selected variable of interest whilst holding other variables constant.