This retrospective longitudinal study included hospitalized patients with COVID-19 in an isolation ward who received negative pressure ventilation from February 2020 to October 2020. COVID-19 was diagnosed by confirmatory quantitative reverse transcription–polymerase chain reaction using upper or lower respiratory tract samples. A protocol was used for systematic questionnaires and anthropometric measurements to ascertain demographic information, initial symptoms, and comorbidities of the patients. Within 24 hours of hospitalization, the patients routinely underwent blood tests and anteroposterior-view CXR. Medical decisions regarding treatment and discharge to home were made by each physician based on guidance from the Korea Disease Control and Prevention Agency. We excluded patients with missing clinical information or when there were technical difficulties in reading CXR images due to compromised software compatibility.

For model training, validation, and internal testing, we used CXR images and clinical information from patients who were admitted at Boramae Medical Center (BMC), Seoul, Korea. For external testing, we used the Korean Imaging Cohort of COVID-19 (KICC-19) data set, which collects imaging data and clinical information from patients with COVID-19 at 17 medical centers in Korea. Details on the profiles of the KICC-19 data set were published in an earlier report [14].

Data were collected on baseline characteristics, including age, sex, BMI, smoking status, and comorbidities. A protocolized questionnaire was used at the BMC to identify the symptoms of patients with COVID-19, including abnormal smell or taste, myalgia, sore throat, cough, sputum, chest discomfort, dyspnea, fever or chills, rhinorrhea, and nausea or diarrhea. In the KICC-19 data set, we extracted information on symptoms such as cough, dyspnea, and fever. We obtained laboratory test results, including white blood cell counts and lymphocyte percentage, as well as C-reactive protein (CRP), procalcitonin, troponin-I, and lactate dehydrogenase levels. Information on treatment and disease severity was also acquired. Information was obtained on clinical outcomes, including hospital length of stay (LOS) and oxygen supplementation, as well as the use of a high-flow nasal cannula (HFNC), mechanical ventilator (MV), or extracorporeal membrane oxygenation (ECMO). ARDS was operationally defined as a medical condition needing an HFNC, MV, or ECMO.

We obtained 26,684 CXR images with information on the location and extent of pneumonia provided by the Radiological Society of North America (RSNA) pneumonia-detection challenge. We extracted the initial CXR images from the electronic medical records of patients hospitalized for COVID-19 at the BMC and registered them in the KICC-19.

The primary outcome was the performance of the AI model in predicting clinical outcomes such as hospital LOS ≤2 weeks, need for oxygen supplementation, and development of ARDS based on CXR images. The output value from this model was defined as the CXR score. The secondary outcome was whether the performance of the prediction model using clinical information could be improved by combining it with the CXR score.

The DL model was implemented using the open-source PyTorch library (version 1.7.0+cu101) with the CUDA/cuDNN (versions 10.1 and 7.6.3, respectively) computing frameworks on a single graphics processing unit (Geforce RTX 3090; NVIDIA). The overall data flow and proposed model architecture is summarized in Figure 1. We developed our model in two stages, to ensure robust performance: (1) backbone training and (2) model training. First, we trained the backbone of our model with a data set (n=26,684) including information on the location and extent of pneumonia from the RSNA pneumonia-detection challenge to boost performance by learning robust features from a large quantity of data. The EfficientNet B5 architecture was used as the backbone architecture due to its computational efficiency and performance compared to those of other convolutional neural network architectures, such as ResNet and DenseNet. For backbone training, we attached a region proposal network, a region-of-interest pooling layer, and a classifier network to the backbone, which was pretrained on ImageNet, to configure the faster region-based convolutional neural network (RCNN) architecture (Multimedia Appendix 1, Figure S1). We trained the faster-RCNN model with the Adam optimizer under the following settings: learning rate of 1×10^–6, learning decay rate of 0.8, learning rate decay step size of 4, and batch size of 3. We selected the model with the minimum validation loss, which was a combination of classification and bounding box regression loss. After the backbone training stage, the backbone learns to extract pertinent features for detecting pneumonia from posteroanterior or anteroposterior CXR images.

Next, we developed and trained the model for the classification of clinical outcomes. The backbone was followed by 3 branches: one each for hospital LOS ≤2 weeks, oxygen supplementation, and development of ARDS. Each branch consisted of an average pooling layer, a 2D convolution layer, a clinical data channel, and a fully connected layer. The outputs of the model were 3 probabilities between 0 and 1: one each for hospital LOS ≤2 weeks, oxygen supplementation, and development of ARDS; for each output, a probability >0.5 indicated a positive prediction. We initialized the weights of the models using the weights from the previous backbone training stage. We fixed the weights of the backbone during the model training stage to keep the feature extractor untouched. To train the proposed model, we used a configuration of the Adam optimizer with a learning rate of 1×10^–4 and batch size of 12. We selected the model with minimum validation loss, that is, classification loss.

Demographic information, symptoms, laboratory test results, treatments, and study outcomes were compared among the training, validation, internal testing, and external testing sets using the Student 2-tailed t test or Mann-Whitney U test for continuous variables and the Pearson chi-square test or Fisher exact test for categorical variables. Univariate and multivariate logistic regression analyses were performed using the stepwise selection method. For 3 different prediction models for each clinical outcome (model 1, using the CXR score derived from the DL model; model 2, using clinical information derived from the multivariable regression model; and model 3, using both the CXR scores and clinical information from the multivariable regression model), performance was evaluated using sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and the area under the receiver operating characteristic curve (AUROC). We considered an AUROC <0.7 suboptimal performance, 0.7 to 0.79 acceptable, 0.8 to 0.89 excellent, and ≥0.9 outstanding [15]. A comparison of the predictive performance between the 2 different models used the DeLong test [16] or the bootstrap test [17]. Statistical significance was set at P<.05. Calibration of the CXR score–based models (models 1 and 3) was evaluated by plotting the observed versus predicted probabilities and using the P value for the Spiegelhalter statistic [18,19]. Statistical significance in the Spiegelhalter z test indicates poor calibration. All the statistical analyses were performed using R (version 4.1.0; R Core Team).

The Institutional Review Board Committee of the Boramae Medical Center (BMC) approved the study protocol and waived the need for informed consent for access to the electronic medical records (30-2020-307).

We used CXR images and clinical information from hospitalized patients with COVID-19 for model training (n=589), validation (n=75), and internal testing (n=75); we used patients with COVID-19 (n=467) registered in the KICC-19 for external testing. The median interval between symptom onset and CXR was 3 (IQR 1-6) days. The baseline characteristics of the combined total of 1206 patients are summarized in Table 1. The mean age was 53.4 years; 52.3% (n=631) were female; and 9.4% (n=113) were every-day smokers. Comorbidities included hypertension (n=317, 26.3%), diabetes mellitus (n=188, 15.6%), cancer (n=74, 6.1%), cardiovascular disease (n=73, 6.1%), chronic lung disease (n=50, 4.1%), chronic liver disease (n=26, 2.2%), and chronic kidney disease (n=20, 1.7%). The baseline characteristics of the patients in the training, validation, and internal testing sets are described in Multimedia Appendix 1, Table S1.

The clinical features of the 1206 patients are presented in Table 2. Cough, fever, and dyspnea were present in 53% (n=639), 51.7% (n=624), and 17.7% (n=213) of patients, respectively. At baseline assessment, mean white blood cell (WBC) count was 5024 cells/µL, with 29.4% lymphocytes. The median CRP and procalcitonin levels were 0.42 mg/dL and 0.03 ng/mL, respectively. Treatment for COVID-19 was remdesivir in 5.8% (n=70) and corticosteroids in 9.5% (n=115) of patients. Eligible patients were hospitalized for a median of 15 (IQR 11-24) days. HFNC, MV, and ECMO were used in 5.4% (n=65), 5% (n=60), and 1.5% (n=18) of patients, respectively. The clinical features of patients included in the training, validation, and internal testing sets are described in Multimedia Appendix 1, Table S2.

The AI model was trained, validated, and internally tested using the CXR images of the hospitalized patients with COVID-19. The performance of each model in the internal testing set is described in Multimedia Appendix 1, Figure S2. The probability of each prespecified outcome (CXR score) was calculated in the external testing set (ie, KICC-19); the AUROCs were 0.602 (95% CI 0.540-0.664) for hospital LOS ≤2 weeks, 0.647 (95% CI 0.586-0.708) for oxygen supplementation, and 0.782 (95% CI 0.720-0.845) for development of ARDS. Representative heat maps visually explain how DL preferentially recognized pneumonic lesions in the CXR images (Figure 2). The heat maps used the feature maps from the clinical data channel of the model weighted by the output probabilities.

We identified clinical variables that were significantly associated with clinical outcomes in hospitalized patients with COVID-19 (Multimedia Appendix 1, Tables S3-4; Table 3). Patients with hypertension, chronic liver disease, low lymphocyte count, or corticosteroid treatment were less likely to be discharged from the hospital within 2 weeks. Patients who needed oxygen supplementation were older; had hypertension, diabetes, or dyspnea; and had a higher level of inflammatory markers, including CRP and procalcitonin. ARDS was more common in those who were older, had dyspnea, or had higher procalcitonin levels. The performance of the prediction models using significant clinical variables was evaluated in the external testing set, with AUROCs of 0.618 (95% CI 0.558-0.678) for hospital LOS ≤2 weeks, 0.567 (95% CI 0.501-0.632) for oxygen supplementation, and 0.878 (95% CI 0.835-0.920) for development of ARDS.

The sensitivity, specificity, PPV, NPV, and accuracy of each model are summarized in Multimedia Appendix 1, Table S5. A comparison of the performance of the different prediction models for each clinical outcome is shown in Figure 3. We found no significant difference in the performance of the 3 prediction models in predicting hospital LOS ≤2 weeks (Multimedia Appendix 1, Figure S3). Model 3 showed an AUROC of 0.704 (95% CI 0.646-0.762) in predicting oxygen supplementation, which was significantly superior to models 1 and 2 (P<.001 and P=.02, respectively; Multimedia Appendix 1, Figure S4). Model 2 showed better performance in predicting ARDS than model 1 (P=.01) (Multimedia Appendix 1, Figure S5). Model 3 showed an AUROC of 0.890 (95% CI 0.853-0.928) for ARDS, which was significantly superior to model 1 (P=.004).

The calibration of models 1 and 3 in the internal and external test data sets is described in Table 4 and Multimedia Appendix 1, Figures S6 and S7. The Spiegelhalter z test showed good calibration of model 3 for hospital LOS ≤2 weeks, oxygen supplementation, and ARDS in the internal test set. Model 3 showed appropriate calibration only for ARDS in the external test data set (P=.86).

We developed and externally validated an AI model to predict prespecified clinical outcomes based on DL using CXR. The performance of the AI model using CXR and the logistic regression model using clinical information were suboptimal for predicting hospital LOS ≤2 weeks or oxygen supplementation; there were no differences between the 2 models. The combined model, with both CXR score and clinical information, performed better in predicting oxygen supplementation. The AI prediction model for ARDS using the CXR score performed acceptably but was inferior to the model using clinical information. The combined model did not perform better in predicting ARDS than clinical information alone. CXR score calibration was appropriate for ARDS in the external test data set but not for hospital LOS ≤2 weeks or oxygen supplementation, suggesting that the CXR score may be important in identifying COVID-19 patients at high risk of progression to severe illness or ARDS. However, it is desirable to refrain from making prognoses for patients with COVID-19 based on CXR alone, considering that the predictive performance of the AI model using CXR was inferior to that of the model combining CXR score and clinical information.

CXR images in patients with COVID-19 pneumonia show various features, including diffuse ground-glass opacities, patchy reticular or nodular opacities, and consolidation [20]. The radiological features of CXR are related to the COVID-19 prognosis [21,22]. The CXR severity scoring system, which is based on radiological interpretation, is significantly associated with the prognosis of patients with COVID-19 [23]. To automate the quantification of the extent and opacity of lung lesions and the consequent assessment of radiological severity, DL models using CXR have been evaluated in COVID-19 pneumonia [24]. Recently, a DL model showed acceptable performance for predicting COVID-19 pneumonia based on CXR [12]. In our study, the CXR score derived from an AI predictive model showed superior performance for oxygen demand and comparable performance for ARDS compared to clinical information. With the application of DL techniques, CXR may need to be reconsidered as a beneficial tool for making COVID-19 pneumonia prognoses.

Hospital LOS is a clinical indicator of disease severity and time to recovery in patients with COVID-19. Prolonged hospital LOS has been associated with specific demographic characteristics and underlying comorbidities [25]. In COVID-19 patients with pneumonic infiltration in CXR, the time to negative conversion is prolonged [26]. However, none of our prediction models using CXR or clinical information were suitable for predicting whether a patient could be discharged within 2 weeks. This is consistent with a previous study reporting that radiological progression in chest CT and hospital LOS were not correlated [27]. Therefore, evidence to support an AI model using only baseline CXR to predict early recovery from COVID-19 or hospital LOS ≤2 weeks is insufficient.

Severe illness in COVID-19 is defined as a condition with SpO₂ ≤94% in room air, including supplemental oxygen demand [28]. Progression to severe COVID-19 increases mortality risk [29]. Early prognosis and intervention may improve mortality risk in patients with COVID-19 [30,31]. Early administration of dexamethasone is associated with less progression to severe COVID-19 [32]. Therefore, many studies have attempted to predict severe COVID-19 using all available medical information, including CXR, but previous prediction models are not sufficiently validated [3]. Our prognostic model was externally validated to verify its performance in predicting oxygen supplementation need in patients with COVID-19. The AI prediction models using CXR and logistic regression with clinical information were suboptimal for predicting oxygen supplementation in patients with COVID-19. However, the predictive performance for oxygen supplementation improved to an acceptable level after clinical information was combined with the CXR score. Our results suggest that clinical and radiological information are complementary in predicting the need for oxygen supplementation.

Because of high mortality and morbidity, predicting COVID-19–associated ARDS is important [33]. CXR and clinical information, when applied to a time-dependent DL model to predict MV use, showed good predictive performance [12]. Recently, the Berlin definition of ARDS was expanded to include patients treated with HFNC [34]. Therefore, our study operationally defined ARDS as an event in which HFNC, MV, or ECMO were administered. Our AI model using CXR score showed acceptable performance; the combined model using CXR score and clinical information showed excellent performance in predicting the use of HFNC, MV, and ECMO. Further clinical trials are needed to ascertain whether early detection of patients at high ARDS risk can improve outcomes through early treatment for severe COVID-19.

Our study had some limitations. First, the interval between symptom onset and CXR imaging varied. The natural course of COVID-19 suggests that radiological abnormalities would have been ground-glass opacities at the beginning, progressing to consolidative lesions. Therefore, if the interval between symptom onset and CXR imaging was too short, our AI model might have underestimated the severity of prognoses among patients with COVID-19. Second, the performance of our prediction model may have changed according to vaccination history. As the vaccination rate increases, the progression to severe illness or ARDS decreases. Therefore, the PPV of our AI model may decrease in vaccinated patients. Third, most of the included patients were diagnosed with COVID-19 before results on the efficacy of dexamethasone or antiviral agents were reported [35,36]. Therefore, it is necessary to validate the performance of the AI prediction model under recently introduced standard treatments.

A prediction model combining CXR score and clinical information was externally validated as having acceptable performance in predicting progression to severe illness and excellent performance in predicting the use of HFNC or MV in patients with COVID-19. We hypothesize that making prognoses with AI models using CXR could be applied for patients with COVID-19 in different settings.