Many countries ramped up the production of real-time reverse transcription polymerase chain reaction (RT-PCR) testing kits to diagnose COVID-19, and thus far, it remains the gold standard for confirmed diagnosis [169]. However, this laboratory-based test is limited by low sensitivity, as reported by several studies [169,170]. As highlighted by both Vaishya et al [20] and Lalmuanawma et al [12], AI can prove helpful in the diagnosis of various infectious diseases (eg, SARS, HIV, and Ebola) when used in conjunction with medical imaging technologies such as CT, magnetic resonance imaging (MRI), and X-ray. Radiological images (CT and X-ray) have been used by clinicians to confirm COVID-19–positive cases; these imaging findings also serve as an important complement to the RT-PCR test [171]. In this systematic review, we found LUS has been used to diagnose COVID-19, in addition to CT and X-Ray. However, we did not find any study using MRI for COVID-19 diagnosis.

Several studies have reported that the use of chest CT for early-stage detection of COVID-19 has proven to have a low rate of misdiagnosis and can provide accurate results even in some asymptomatic cases [172]. We identified 15 studies that used CT to detect COVID-19. One of the most cited studies by Li et al [120] applied DL (COVNet) to differentiate COVID-19 and non–COVID-19 pneumonia CT scans. The area under the receiver operating characteristic (AUROC) curve reported to identify COVID-19 based on chest CT exam was 0.96 and the AUROC curve reported to identify community-acquired pneumonia based on chest CT exam was 0.95. The accuracy reported is slightly higher than that reported by Ardakani et al [93], which was also found in the review by Lalmuanawma et al [12]. However, there are some disadvantages associated with using chest CT for COVID-19 diagnosis, such as the high radiation dose (7 mSv vs 0.1 mSv for chest X-ray) and the fact that chest CT is more expensive than chest X-ray [173,174].

In this systematic review, we identified 23 studies that used chest X-ray and applied AI techniques to diagnose COVID-19 cases. A study by Apostolopoulos et al [91] applied a transfer learning strategy to train convolutional neural network models and then automated the detection of COVID-19 using chest X-ray images. The model (VGG19) achieved an overall accuracy of 97.82% to detect COVID-19 based on a dataset of 224 COVID-19, 700 pneumonia, and 504 normal X-ray images. A similar study was performed by Khan et al [117] using transfer learning and convolutional neural network (Xception) architecture with 71 layers that were trained on the ImageNet dataset. Their model (CoroNet) achieved an average accuracy of 87% to detect COVID-19 based on a dataset of 284 COVID-19, 657 pneumonia (both viral and bacterial), and 310 normal chest X-ray images. These recently published studies successfully used transfer learning strategy to overcome sample size limitation and adapt generalizability; it is noteworthy that such studies were not available in the earlier literature reviews [12,20]. Although chest X-ray is cost-effective and involves a considerably lower radiation dose than chest CT, it is less sensitive, especially in the early stages of the infection and in cases of mild disease [175]. We recommend that new studies develop AI models that can detect COVID-19 by using a combination of CT and X-ray images along with clinical variables to aid clinical practitioners with accurate diagnosis.

During the 2009 influenza (H1N1) epidemic, LUS proved useful in accurately differentiating viral and bacterial pneumonia and were found to have higher sensitivity in detecting avian influenza (H7N9) than chest X-ray [176]. Although clinicians recommend the use of LUS imaging in the emergency room for the diagnosis and management of COVID-19, its role is still unclear [177]. In our review, we identified a study by Roy et al [128] that used a DL model based on an annotated LUS COVID-19 dataset to predict disease severity. The results of the study were reported to be “satisfactory.” Moreover, a study by Joshi et al [115] proposed an ML approach that utilizes only complete blood count and gender information of the patient to predict COVID-19 positivity, as an alternative to the RT-PCR test. These authors built a logistic regression model based on retrospective data collected from a single institute and validated using multi-institute data. Prediction of COVID-19 infection demonstrated a C-statistic of 78% and sensitivity of 93%. The aim of the study was to develop a decision support tool that integrates readily available laboratory test results from patients’ EHRs.

The COVID-19 pandemic has strained global health care systems, especially ICUs, due to the high ICU transfer rates of hospitalized patients with COVID-19 [135]. As the pandemic progressed, the research focus shifted from detecting the presence of the novel coronavirus in patient samples to the prediction of patient recovery and associated risks [178]. Therefore, early systematic reviews included very few studies that focused on DP [12,20]. In this review, we found 19 studies that predicted DP and the likely outcomes in the confirmed COVID-19 population. Prior identification of hospitalized patients who may be at high-risk may aid health care providers to more efficiently plan and prepare for ICU resources (eg, beds, ventilators, and staff) [179].

A study by Cheng et al [135] developed an ML-based model to predict ICU transfers within 24 hours of hospital admission. The random forest model was used for COVID-19 prediction and was based on multiple variables such as vital signs, nursing assessment, laboratory test results, and electrocardiograms collected during the patient’s hospital stay. The overall AUROC curve of the model was reported to be 79.9%. Similar work was done by Shashikumar et al [141] to predict the need for ventilation in hospitalized patients 24 hours in advance. The prediction was not only limited to patients with COVID-19 but also included other hospitalized patients. These authors studied 40 clinical variables, including 6 demographic and 34 dynamic variables (eg, laboratory results, vital signs, sequential organ failure assessment, comorbidity, and length of hospital stay). In contrast to the traditional ML model used by Cheng et al [135], Shashikumar et al [141] resorted to a DL model (VentNet) for prediction with an area under the curve (AUC) of 0.882 for the general ICU population and 0.918 for patients with COVID-19. Both the aforementioned studies relied on clinical variables for prediction, whereas a study by Burian et al [132] combined clinical and imaging parameters for estimating the need for ICU treatment. The major finding of the study was that the patients needing ICU transfers had significantly elevated interleukin-6, C-reactive protein, and leukocyte counts and significantly decreased lymphocyte counts. All studies in this category applied AI techniques to facilitate the efficient use of clinical resources and help hospitals plan their flow of operations to fight the ongoing pandemic.

Prediction and risk stratification of COVID-19 cases that are likely to have adverse outcomes will help to streamline health care resources for patients that need urgent care. In our review, Yadaw et al [146] evaluated different ML models to classify COVID-19 cases as deceased or alive classes. This classification was based on 5 features: age, minimum oxygen saturation during the encounter, type of patient encounter, hydroxychloroquine use, and maximum body temperature. Their study revealed that age and minimum oxygen saturation during encounters were the most predictive features among the different models examined. The overall AUC was reported as 0.91. On the other hand, Ji et al [137] focused on the early identification of COVID-19 cases that are likely to be at high-risk. Variables used for this prediction model included demographics, comorbidities, and laboratory test results. They found a strong correlation between comorbidities and DP as supported by various other studies. The study further suggests that a decrease in lymphocyte count and an increase in lactate dehydrogenase levels are related to DP. The overall AUC reported was 0.759. In both studies, Yadaw et al [146] and Ji et al [137], the ML models were trained on the retrospective data and validated on prospective data. Li et al [139] 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 3 days of hospital admission for patients that were initially not intubated.

COVID-19 pneumonia is associated with high morbidity and mortality, and it is critical to differentiate COVID-19 from general pneumonia [180]. In the study by Jiang et al [138], their model used demographics, vital signs, comorbidities, and laboratory test results to predict patients that are likely to develop ARDS. Of these variables, laboratory levels of alanine aminotransferase (ALT), the presence of myalgias, and elevated hemoglobin were found to be the most predictive features. The overall accuracy of predicting ARDS was 80%. Moreover, using ALT alone, the model achieved an accuracy of 70%. Zhang et al [147] built a DL diagnostic and prognostic predictive model to detect COVID-19 and identified variables associated with risk factors for early intervention and monitoring of the disease. The study comprised 3777 patients (5468 CT scans) to differentiate COVID-19 pneumonia from other types of pneumonia and normal controls. The AUROC of the model was reported as 0.97.

In general, DL techniques are employed to improve prediction accuracy by training models on large volumes of data [182]. In our review, several studies applied AI techniques, either using smaller imaging datasets specific to the organization, or mid- to large-sized datasets from publicly available repositories. However, there are substantial costs associated with the development and maintenance of such repositories [183]. To overcome data size and cost limitations, Xu et al [110] proposed a decentralized AI architecture to build a generalizable model that is distributed and trained on in-house client datasets, eliminating the need for sharing sensitive clinical data. The proposed framework is in the early phase of adoption and needs technical improvements before it is widely employed by participating health care organizations.

There are classification challenges associated with the diagnosis of COVID-19 using patients’ radiological imaging data, which consists of multiple steps. In general, the initial steps involved in image classification are preprocessing, annotation, and feature extraction. Annotation of radiological images is time-consuming and depends on the sheer expense of the expert radiologist. Several strategies have been proposed to address this challenge, such as self-supervised and transfer learning techniques. Our review identified a study by Wang et al [145] that used a transfer learning strategy to aid COVID-19 diagnostic and prognostic analyses. The study used a 2-step transfer learning strategy: first, the model was trained on a large lung cancer CT dataset (n=4106) along with epidermal growth factor receptor gene sequencing to learn associations between chest CT image and micro-level lung functional abnormalities. Thereafter, the model was trained and validated to differentiate COVID-19 from other pneumonia (AUC 0.87-0.88) and viral pneumonia (AUC 0.86) types. We believe such techniques will significantly improve the computational costs associated with training the models.