Developing a risk stratification mechanism among COVID-19 patients helps to facilitate timely assessments, allocate hospital resources, and appropriate decision making [178]. Jiang et al [136] model used demographics, vital signs, comorbidities, and lab results to predict patients that are likely to develop Acute Respiratory Distress Syndrome (ARDS). Of these variables, lab results alanine aminotransferase (ALT), the presence of myalgias, and elevated hemoglobin were the most predictive features. The overall accuracy of predicting ARDS was 80%. Moreover, using ALT alone, the model achieved an accuracy of 70%. However, the conclusion was drawn based on a limited patient (n = 53) set.

Yadaw et al [144] evaluated different ML models to classify COVID-19 patients into deceased or alive classes. The classification was based on five features: age, minimum oxygen saturation during the encounter, type of patient encounter, hydroxychloroquine use, and maximum body temperature. The study revealed age and minimum oxygen saturation during encounters were the most predictive features among different models. The author reports hydroxychloroquine as one of the top mortality predictors identified by XGBoost and Support Vector Machine (SVM) models; however, only a third of the patients in the dataset had been treated with this drug. Overall AUC was reported as 0.91. Ji et al [135] focused on early identification of COVID-19 positive patients that are likely to be at high-risk. Variables used for prediction were demographics, comorbidities, and lab results. The study identified a strong correlation between comorbidities and disease progression as supported by various other studies. The study further claims, a decrease in lymphocyte count and an increase in lactate dehydrogenase are related to disease progression. Overall AUC reported was 0.759. Both studies, Yadaw et al [144] and Ji et al [135] trained their respective ML models on the retrospective data and validated on prospective data.

Zhang et al [145] build a DL diagnostic and prognostic predictive model to detect COVID-19 and identified variables associated with risk factors for early intervention and monitoring. The study used 3,777 patients (5,468 CT scans) to differentiate coronavirus pneumonia from other pneumonia and normal controls. AUROC of the model was reported as 0.97. To improve image segmentation performance a DL model (DeepLabv3) was employed. The study further evaluated the effect of various drug treatments on lesion size and volume changes by comparing the differences between pre-treatment and post-treatment. The work claimed to provide a potential quantitative evaluation of the efficacy of drug treatment. The study employed five senior radiologists to annotate and review image segmentation, the availability of such resources is very expensive. To avoid the time-consuming annotation, Wang et al [143] used a transfer learning strategy to aid COVID-19 diagnostic and prognostic analysis. The study used a two-step transfer learning strategy, first, the model was trained on a large lung cancer CT dataset (4106 patient) along with Epidermal Growth Factor Receptor (EGFR) gene sequencing to learn associations between chest CT image and micro-level lung functional abnormality. Later, the model was trained and validated to differentiate COVID-19 from other pneumonia (AUC 0.87–0.88) and viral pneumonia (AUC 0.86). Moreover, the model classified patients into high-risk and low-risk categories to help manage medical resources and provide early interventions. For COVID-19 disease evaluation and risk stratification, Li et al [137] built a pulmonary disease severity score using X-Rays and neural network models. The score was computed as the Euclidean distance between the patient’s image and a pool of normal images using the Siamese neural network. The score predicted (AUROC 0.80) subsequent intubation or death within three days of hospital admission for patients that were initially not intubated.

This review has some inherent limitations. First, there is a possibility of studies missed due to the search methodology. Second, we removed five publications where full text was not available, and this may have introduced bias. Third, we included studies that were available as preprints. Finally, a comparison of ML model performance was not possible in the quantitative descriptive analysis as variables, sample size, and source of data was diverse across the studies. The current systematic review includes studies that were available online as of June 27^th, as the pandemic progresses, we intend to write a second review on the studies published after the aforementioned date.