We conducted a multicenter retrospective cohort study in Hubei Province, China. The study was approved by the ethics committee in each participating hospital. Written informed consent was waived because of the observational nature of the study and the epidemic of COVID-19 as an emergency public health event.

We included patients with confirmed COVID-19 pneumonia who were admitted to one of three hospitals in Hubei, all designated for COVID-19, and who underwent at least two chest CT scans after admission. The COVID-19 diagnosis at the three hospitals was initially based on the criteria published by WHO on January 28, 2020 [1], and all cases were later confirmed by real-time reverse-transcriptase polymerase chain reaction analysis of throat swab specimens according to a previously published protocol [2, 4]. Patients with a confirmed diagnosis were treated with effective isolation and protective conditions in the hospitals, and patients with severe and critical disease were admitted to the emergency department or ICU ward. All patients underwent initial laboratory testing, including complete blood count, coagulation profile, renal and liver function, creatine kinase, lactate dehydrogenase, and electrolytes. All patients' vital signs (body temperature, heart rate, blood pressure, and respiratory rate) were monitored during their hospitalization. Patients at all centers received treatment that followed the guidelines established by Jin et al. [10]. All of the patients received antibacterial agents. The use of antiviral therapy and methylprednisolone varied depending on disease severity. For patients with severe disease, continuous renal replacement therapy was applied if acute renal impairment or failure occurred. Invasive mechanical ventilation was administered to patients with hypoxic respiratory failure and ARDS. Extracorporeal membrane oxygenation support was used for patients with refractory hypoxemia that was difficult to correct by prone positioning and protective lung ventilation. More detailed treatment strategies performed at these centers are described by Zhang et al. [11].

To be discharged, patients needed to satisfy all of the following four criteria: afebrile for more than 3 days, two consecutive negative results on reverse-transcriptase polymerase chain reaction tests conducted at least 24 hours apart, significantly improved respiratory symptoms, and improvement in the radiologic abnormalities on chest radiograph or CT.

Demographic, clinical, laboratory, and outcome data were extracted from electronic medical records using a standardized data collection form, which was a modified version of the WHO and International Severe Acute Respiratory and Emerging Infection Consortium case record form for severe acute respiratory infections [12]. Data were extracted from the time of patients' initial hospital admission until either hospital discharge or mortality. This extraction was performed by two physicians (X.H., a pulmonologist with 5 years of experience, and W.T., a pulmonologist with 15 years of experience). A third physician (Z.C., a pulmonologist with 25 years of experience) adjudicated any difference in interpretation between the two primary reviewers.

Cases of COVID-19 were classified as severe or nonsevere at the time of hospital admission using the American Thoracic Society guidelines for community-acquired pneumonia [5, 13]. Patients were classified as having severe disease if they had any of the following: dyspnea, respiratory frequency of 30 breaths/min or more, blood oxygen saturation of 93% or lower, PaO₂:FiO₂ ratio of less than 300, or an increase in lung infiltrates of over 50% within a period of 24–48 hours. Patients with severe disease generally required supplemental oxygen, although this was not a criterion for the classification.

Chest CT scans were performed with a single inspiratory phase using an MDCT scanner (Somatom Definition AS, Siemens Healthineers; Somatom Definition Flash, Siemens Healthineers; or GE BrightSpeed 16, GE Healthcare). Patients were instructed on breath-holding to minimize motion artifacts. All CT scans were obtained with the patient in the supine position during a single breath-hold without IV contrast material. The tube voltage was 120 kV for all scans. The tube current–exposure time product and CT doses varied across scans. Automatic exposure control technology was used to automatically adjust the tube current for each patient.

Two physicians at Zhongnan Hospital of Wuhan University (Z.C. and X.H.), Central Hospital of Wuhan (W.T. and a nonau-thor radiologist with 15 years), and Jingzhou Chest Hospital (P.H., a radiologist with 20 years, and a nonauthor radiologist with 15 years) scored the CT images and were unaware of patients' clinical status. As training before the evaluation, the investigators reviewed CT images from a group of eight patients with severity scores ranging from none to severe. After initial independent evaluations, the two physicians at each hospital discussed any disagreements to reach consensus.

Pulmonary disease severity on CT scans was assessed visually. Each scan was assigned to one of the following categories: no ground-glass opacity (GGO) or consolidation, GGO only, consolidation only, or both GGO and consolidation. GGO was defined as an area of haziness with increased attenuation that did not obscure the underlying vascular markings. Consolidation was defined as opacity obscuring the underlying vascular markings. Each scan was also assessed for the presence of a peripheral distribution of opacification, linear opacities, or crazy paving pattern. Each lobe branching off the main bronchus (three on the right, two on the left) was assessed for involvement. A visual severity score was assigned to each lobe using four categories of lobe involvement on the basis of GGO and consolidation: 0, none (0% involvement); 1, mild (< 50% involvement); 2, moderate (50–75% involvement); and 3, severe (> 75% involvement). An overall CT visual severity score was reached by summing the scores for the five lobes (possible range for summed scores, 0–15).

The primary measured outcome for this study was a composite endpoint of major COVID-19–associated complications. Major complications included ARDS, acute cardiac injury, arrhythmia, septic shock, AKI, liver dysfunction, secondary infection, and in-hospital mortality (defined as death from any cause before discharge). ARDS was diagnosed according to the Berlin definition [14], and AKI was diagnosed according to the Kidney Disease Improving Global Outcomes clinical practice guidelines [15]. Acute cardiac injury was diagnosed if serum levels of cardiac biomarkers (i.e., troponin I) were above the 99th percentile upper reference limit or if new abnormalities appeared on ECG and echocardiography [2]. Sepsis and septic shock were defined according to the “Third International Consensus Definitions for Sepsis and Septic Shock” published in 2016 [16]. Liver dysfunction was diagnosed if alanine amino-transferase and aspartate aminotransferase increased during the course of disease [17, 18]. Secondary infection was diagnosed if the patients showed clinical symptoms or signs of nosocomial pneumonia or bacteremia and had a culture that was positive for a new pathogen from a lower respiratory tract specimen (including sputum, transtracheal aspirates, or bronchoalveolar lavage fluid) or from blood samples taken 48 hours or more after admission [2].

The study variables were evaluated using the sample median with interquartile range (IQR) or the number with proportion. The study sample was divided into three groups on the basis of disease severity at hospital admission and survival status: survivors with nonsevere disease, survivors with severe disease, and nonsurvivors. The three groups were compared in terms of clinical features on admission. Categoric variables were compared between groups using the Pearson chi-square test or Fisher exact test, whereas continuous variables were compared using the nonparametric Kruskal-Wallis test. The analysis of interreader agreement was conducted using weighted kappa coefficients.

The findings on lung CT were compared among four groups defined by the amount of time between the onset of symptoms and CT scan: week 1 (e.g., scans performed within the first week after symptom onset), week 2, week 3, and week 4 or later. Non-survivors with severe disease only had scans available in weeks 1 and 2; nonsurvivors were typically critically ill patients who had to use noninvasive or invasive ventilators and did not undergo additional CT scans after 2 weeks of hospitalization. If a patient had multiple CT scans within one of these four time periods, the first of these scans was used for the analysis.

To investigate the nonlinear temporal evolution of CT findings, we built generalized additive models with penalized splines [19], in which the expected value of CT score is related to the scan time period via a nonparametric smooth function. We estimated the nonlinear disease progression curves by considering the time of CT as a continuous predictor. Nonsurvivors were not considered in this analysis given the lack of data. Nonparametric analysis of covariance was conducted to test the parallelism of the nonlinear temporal functions between the groups of survivors with nonsevere and severe disease [20]. A change point detection analysis [21] was performed to estimate the time of peak severity for the three groups.

Univariable and multivariable logistic regression analyses were performed to assess the association between the risk factors and patient outcome. To avoid overfitting, four variables were chosen for the multivariable prediction model on the basis of previous research findings and clinical constraints. In the multivariable model, the CT visual severity score was classified as none or mild (score 0–5), moderate (score 6–10), or severe (score 11–15). Correlation analysis of the predictors was conducted to avoid the multicollinearity in a regression model. Following the guidelines of the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement [22], bootstrap internal validation was conducted for the model, including an evaluation of the model's discriminative ability and calibration [23]. An online risk calculator was created based on the model [24]. The source code for this online risk calculator is available in the Supplemental Methods, which can be viewed in the AJR electronic supplement to this article, available at www.ajronline.org.

External validation was then performed using an independent external cohort to evaluate the overall accuracy of the model. For this external cohort, 135 hospitalized patients with COVID-19 were retrospectively identified from a fourth hospital designated for COVID-19 in Hubei during the same time period as the cohort from the other three hospitals. Data were collected for these patients using the same method as described for the internal cohort. Two readers (W.H., a pulmonologist with 20 years of experience, and a nonauthor radiologist with 20 years) independently reviewed these patients' CT images, which were obtained at the time of hospital admission, and provided a single CT visual severity level: none or mild, < 50% involvement; moderate, 50–75% involvement; or severe, > 75% involvement.

All analyses were performed using the R software program (version 3.6.3, R Foundation for Statistical Computing). The level of statistical significance was set at p < .05 (two-tailed).

A total of 161 hospitalized patients (55.3% [89/161] men and 44.7% [72/161] women) who underwent at least two CT scans following admission were included (Table 1). The median age was 42.0 years (IQR, 33.0–57.0 years). The most common symptoms on admission were fever (84.5%) and cough (48.4%); vomiting (1.9%) and diarrhea (6.2%) were uncommon.

View Larger Version

TABLE 1: Clinical Characteristics of the Study Patients According to Coronavirus Disease (COVID-19) Severity and Survival

On admission, COVID-19 was nonsevere in 56.5% of patients (91/161) and severe in 43.5% (70/161). Of patients with severe disease at admission, 38.6% (27/70) were admitted or transferred to an ICU, 21.4% (15/70) died, and 78.6% (55/70) recovered and were discharged. The survivors with severe disease and the nonsurvivors were both significantly older (median age, 53.0 years [IQR, 39.0–62.5 years] and 72.0 years [IQR, 64.5–74.5 years], respectively) than the survivors with nonsevere disease (median age, 37.0 years [IQR, 30.5–45.5 years]; p < .001). The survivors with severe disease and nonsurvivors were also more likely to have underlying comorbidities (40.0% [22/55] and 86.7% [13/15], respectively) compared with survivors with nonsevere disease (15.4% [14/91]; p < .001). The nonsurvivors had significantly higher systolic blood pressure (p = .004) and respiratory rate (p < .001) on admission compared with the other two groups.

Table 2 shows the laboratory findings on admission. Lymphocytopenia (lymphocyte count < 1.5 × 10⁹/L) was present in 88.2% of the cohort, leukopenia (WBC count < 4.0 × 10⁹/L) in 40.4%, and thrombocytopenia (platelet count < 150 × 10⁹/L) in 37.3%. WBC count, neutrophil count, d-dimer, aspartate aminotransferase, total bilirubin, blood urea nitrogen, creatinine, creatine kinase myocardial band, lactate dehydrogenase, and high-sensitivity troponin were significantly different among the three groups, showing progressive increases from survivors with nonsevere disease to survivors with severe disease to nonsurvivors (p < .05).

View Larger Version

TABLE 2: Laboratory Test Results for Patients With Coronavirus Disease (COVID-19) on Admission to Hospital, Stratified by Patient Group

All patients showed abnormal CT findings during the course of their disease (Table 3). The median CT visual lung severity score in the first week of hospitalization was 2.0 in survivors with non-severe disease, 4.0 in survivors with severe disease, and 11.0 in nonsurvivors. In week 1, a peripheral distribution was observed in 83.1% survivors with nonsevere disease, 73.7% of survivors with severe disease, and 11.1% of nonsurvivors. A total of 36.1% of survivors with nonsevere disease had three or more lobes affected compared with 65.8% of survivors with severe disease and 100.0% of nonsurvivors. GGO without consolidation was the most common pattern for all three groups (76.3–88.9%); in comparison, 11.1–18.4% had GGO with consolidation in week 1. In week 2, 73.1–100.0% had bilateral lung involvement and 25.6–46.0% had GGO with consolidation. These percentages of patients showing GGO with consolidation were similar in week 3. In week 4, 2.5% of survivors with nonsevere disease and 41.7% of survivors with severe disease showed GGO with consolidation.

View Larger Version

TABLE 3: Distribution and Frequency of Lung Lesions on CT at Different Time Stages by Disease Severity

Figure 1 displays the temporal evolution of CT visual severity scores according to disease severity. For survivors with both nonsevere and severe disease, the CT visual severity score reached a peak in week 2 after symptom onset and then decreased progressively in weeks 3 and 4. Nonsurvivors showed much higher CT visual severity scores compared with survivors and no visually apparent decrease between weeks 1 and 2. Figure 2 displays the regression curve fits for survivors with nonsevere and severe disease. The two curves were not parallel (p < .001), indicating different disease evolution patterns between the two groups. Change point detection analysis showed that the peak CT visual severity score for survivors with nonsevere disease on admission was reached approximate-lyf 9 days (95% CI, 7–11) from symptom onset, whereas the peak score for survivors with severe disease was reached approximately 12 days (95% CI, 11–14) from symptom onset.

View larger version (21K)

Fig. 1 —Box-and-whisker plots show CT visual severity scores by time interval for survivors with nonsevere coronavirus disease (COVID-19) (left, red), survivors with severe disease (middle, green), and patients who did not survive (right, blue). Upper and lower limits of whiskers indicate represent 1.5 times interquartile range above and below upper and lower quartiles, respectively, and dots indicate outliers.

View larger version (88K)

Fig. 2 —Nonlinear regression curve fits with 95% CIs. Graphs show lung CT visual severity scores for survivors with nonsevere coronavirus disease (COVID-19) (left) and survivors with severe disease (right). Peak CT score occurred approximately 9 days from symptom onset (95% CI, 7–11 days) for patients with nonsevere disease compared with 12 days (95% CI, 11–14 days) for survivors with severe disease. Gray dots and lines represent CT visual severity scores for individual patients; curved blue lines and shaded areas represent estimated mean functions and 95% CIs, respectively.

Representative serial CT images from patients with nonsevere and severe disease are shown in Figures S1 and S2, respectively, which can be viewed in the AJR electronic supplement to this article, available at www.ajronline.org.

A total of 77.0% of patients (124/161) had no complications of COVID-19, and 23.0% (37/161) had at least one complication. Complications included ARDS (20.5%), AKI (7.5%), liver dysfunction (6.8%), acute cardiac injury (6.8%), septic shock (6.2%), arrhythmia (5.6%), and secondary infection (3.1%) (Table S1, which can be viewed in the AJR electronic supplement to this article, available at www.ajronline.org). All complications were significantly more common (p < .001) in patients with severe disease. In uni-variable logistic regression analysis (Table 4), age, sex, respiratory rate, neutrophil count, lymphocyte count, procalcitonin, d-dimer levels over 1 mg/L, alanine aminotransferase, creatinine, lactate dehydrogenase, and the first CT severity score following admission were all associated with the composite complication endpoint (p < .05).

View Larger Version

TABLE 4: Risk Factors Associated With the In-Hospital Composite Endpoint

The multivariable model for predicting in-hospital complications was constructed using the following four variables: age, sex, lymphocyte count, and the first CT visual severity score after admission (Table 4). In the model, the factor with the highest odds for predicting in-hospital complications was the CT visual severity score (odds ratio, 5.86 [95% CI, 1.70–20.23] for moderate score [p = .005] and 31.28 [95% CI, 2.97–329.80] for severe score [p = .004]). Complications were also significantly associated with older age (odds ratio, 1.09 [95% CI, 1.09–1.14] per 1-year increase; p < .001) and lower lymphocyte count (odds ratio, 0.03 [95% CI, 0.00–0.29] per 1 × 10⁹/L increase; p = .002). Sex was not a significant independent predictor in this model (p = .45).

The internal bootstrap validation of the model for predicting in-hospital complications showed a bias-corrected C-index of 0.94 (Fig. S3, which can be viewed in the AJR electronic supplement to this article, available at www.ajronline.org). In the calibration plot (Fig. S4, also available in the AJR electronic supplement to this article at www.ajronline.org), predicted probabilities closely matched the actual probabilities, indicating that the model performed reasonably well. Although slight underprediction or overprediction occurred at lower predicted probabilities (< .5), the prediction was highly accurate at higher predicted probabilities (≥ .5).

Tables S2–S4 (which can be viewed in the AJR electronic supplement to this article, available at www.ajronline.org) present the demographic, laboratory, and outcome variables, respectively, for the 135 patients in the external validation cohort from a different hospital. The median age was 53.0 years (IQR, 38.0–66.0 years). This cohort included 70 (51.9%) men and 65 (48.1%) women. The weighted kappa interreader agreement between the two readers for the CT-based visual severity level was 0.83 (95% CI, 0.73–0.92).

In the external validation cohort, the prediction model showed a C-index of 0.95 and 0.92 using the CT visual severity scores from the two readers (Fig. S5, which can be viewed in the AJR electronic supplement to this article, available at www.ajronline.org). When using p ≥ .5 as the cut point for predicting complications on the basis of the CT scores, reader 1 achieved an accuracy of 86.7%, sensitivity of 75.6%, and specificity of 91.5%, whereas reader 2 achieved an accuracy of 85.2%, sensitivity of 70.7%, and specificity of 91.4%.