A total of 246 patients were diagnosed with COVID-19 infection by SARS-CoV-2 nucleic acid test in Jingzhou Central Hospital between 17 December 2019 and 20 February 2020. The patients who had been treated in another hospital for more than 7 days (n = 124) were diagnosed severely ill at the time of the first chest CT scan (n = 24), and missed follow-up chest CT scans in early 7 days (n = 5) were excluded in this study (Figure 1). Among the remaining 93 patients, 61 exhibited enlarged pneumonia lesions, with an enlargement range exceeding 10 cm³. Then, all 61 patients were followed up for 30 days. At last, 42 showed no progression to severe illness and 19 exhibited progression to severe illness.

According to the guidelines of Chinese national diagnosis and treatment protocols for COVID-19 (15) and the guidelines of the American Thoracic Society (16), patients were defined as having severe illness if they reached any of the following endpoints: (a) multi-lobar infiltrates and respiratory failure (arterial oxygen pressure (PaO2) < 60 mmHg, accompanied or unaccompanied with PaCO2 > 50 mmHg); (b) mechanical ventilation, intensive medical care or extracorporeal membrane oxygenation (ECMO) treatment (c); organ damage; and (d) in-hospital mortality. In this study, organ damage include ARDS (according to the Berlin definition), acute cardiac injury (heart function graduation ≥ IV levels by New York Heart Association heart function rating), acute kidney injury (AKI, according to the Kidney Disease Improving Global Outcomes clinical practice guidelines), and liver dysfunction alanine (aminotransferase >5 times upper limit of normal) (7, 13).

We collected the demographic, clinical, laboratory, and outcome data from electronic medical records in the hospital information system (HIS). The baseline characteristics, including sex, age, and comorbidities, were collected. In this study, the comorbidities include diabetes, hypertension, cardiovascular disease, cerebrovascular disease, chronic obstructive pulmonary disease (COPD), chronic hepatitis B infection, and malignant tumors. The examinations of initial and follow-up in 7 days include neutrophil-to-lymphocyte ratio (NLR) and C-reactive protein (CRP). Meanwhile, the changes between the two examinations were depicted.

Chest CT imaging, including initial and follow-up in 7 days, was acquired from the hospital Picture Archiving and Communication System (PACS). If examinations were collected from outpatient and inpatient departments, they should be matched by name, age, exam time, exam number, and ID card number (if applicable) to verify they belong to the same individual.

All CT examinations were performed using a 16-slice scanner (Siemens SOMATOM Emotion, Germany) without contrast injection when the patients held their breath. The position was supine with the head advanced. Imaging parameters were set as follows: 130 kV; automatic tube current; slice width, 5 mm; beam collimation, 1.2 mm; tube rotation time, 0.6 s; reconstructed slice thickness, 1.5 mm; and matrix, 512 × 512.

Computed tomography images were automatically analyzed using an AI system (YT-CT-Lung, YITU Healthcare Technology Co., Ltd., China). This AI system utilized a fully convolutional network along with adaptive thresholding and morphological operations for the segmentation of lungs and pneumonia lesions (17, 18). Two experienced radiologists, each with over 10 years of experience, independently reviewed the CT images to ensure concordance with the AI-based segmentations corresponding to the lesion range in the CT scan. Illustrative examples of the automatically segmented pulmonary tissue and pneumonia lesions from COVID-19 are shown in Figure 2.

Besides the total volume of pneumonia lesions, other three quantitative features were computed by thresholding on CT values. The CT value range of −750 HU to −500 HU, −500 HU to −200HU, and −200HU to 60HU was defined as ground-glass opacity volume (GV), semi-consolidation volume (SV), and consolidation volume (CV), respectively. In this study, semi-consolidation refers to an intermediate, homogeneous area where there is a noticeable increase in density (19).

Patients were divided into two groups according to the results of follow-up, the severe group and the non-severe group. For only one NLR result that was not collected in this study, we used the median value to replace it. During the initial examination, five (8.2%) CRP results were missed, and three were missed during follow-up (4.9%). To address this issue of missing data, we conducted a chained equation approach method with five replications for multiple imputations using the R MI procedure.

We presented continuous variables as the median and interquartile range (IQR), while categorical variables as numbers and percentages. The Wilcoxon rank-sum test and chi-square tests were applied for relevant comparisons between severe and non-severe groups.

Binary logistic regression was used to explore the association between the predictive features and the development of severe illness. We performed a selection process by excluding all variables with p-values of ≥0.05 in the univariate regression analysis. Subsequently, we analyzed the remaining variables using multivariable regression. To simplify the model, we classified all the continuous variables based on quartiles and curve fitting results. Finally, we constructed a mixed nomogram model based on regression coefficients. The calibration of this model was evaluated through bootstrapping with 200 resamples to produce calibration curves and the Hosmer–Lemeshow goodness-of-fit test. The discrimination of the nomogram model was assessed by the receiver operating characteristic curve (ROC) and the area under the curve (AUC). The AUC in the prediction models was compared using the Delong non-parametric approach. The clinical benefits of the model were determined using decision curve analysis (DCA).

Two-tailed tests were conducted, considering p values of <0.05 as statistically significant. Statistical analyses were carried out using R version 3.3.2 (The R Foundation, https://www.r-project.org/) and the Free Statistics software package version 1.7.

Of the total 93 COVID-19 patients, 32 presented reduced pneumonia lesions or increasing volume of less than 10 cm³, and 61 observed increased lesions in follow-up lung CT. There were no statistical differences in age, gender, or comorbidity between them. However, the progression to severe illness was statistically different, with a p value of 0.006 (Supplementary Table 1).

We classified 61 patients with lung lesions greater than 10 cm³ into two groups based on disease severity of follow-up: non-severe (n = 42) and severe (n = 19). Severe patients (54.0 years, IQR 42.0–60.0) were older compared to non-severe patients (38.5 years, IQR 32.2–49.8, p = 0.014). Although there was a higher proportion of males in the severe group compared to the non-severe group, this difference did not reach statistical significance (55.7 vs. 44.3%, p = 0.18). The presence of comorbidities was significantly associated with a worse outcome (p = 0.04). There was no significant difference in the duration between initial examinations and follow-up between the two groups.

In initial examinations, there were no significant differences for all clinic and CT features between severe and non-severe patients. However, in follow-up, NLR (p = 0.032) and CPR (p = 0.027) were significantly higher in the severe patients than in the non-severe patients. About CT features, except GV, the other three (lesion volume, CV, and SV) were significantly higher in the severe group than in the non-severe group (p < 0.05). Regarding the changes in 7 days, the range of increased lesion volume was from 10.71 to 958.69 cm³. All CT features, except GV, exhibited a more pronounced increase in the severe group, whereas the clinical biomarkers (NLR and CPR) did not exhibit any significant changes (Table 1).

The results of the univariate logistic regression and multivariable logistic regression are presented in Table 2. In follow-up, NLR, SV, and CV were significantly associated with progression to the severe illness of COVID-19 patients with increased pneumonia lesions in the early days (p < 0.05). However, calculating the change in 7 days, only CT characteristics (lesion volume, SV, and GV) exhibited a significant association with severe illness when assessing changes over 7 days. Notably, there were minimal disparities in the results between the univariate and multivariable logistic regression analyses.

In our study, the initial examination NLR and changes over 7 days were not significantly associated with progression to severe illness in COVID-19 patients with escalating pneumonia lesions during the early days. However, NLR in the follow-up examination exhibited a significant relationship. According to the clinical experiments, we added it to the mixed model. Finally, age, NLR in the second blood test, and changes in quantitative CT features (lesion volume, SV, and CV) between the two examinations were selected based on the multivariable logistic regression analysis. Combining quartiles and curve fitting results, a mixed model was developed and presented as a nomogram (Figure 3).

The model showed the highest discrimination between severe COVID-19 patients and no-severe patients in those who had increased lung lesions in the early days, with an AUC of 96.05% (91.66–100%). The sensitivity, specificity, and accuracy of the model were 94.74%, 90.48%, and 0.918, respectively.

The calibration curve of the nomogram showed strong agreement between the severe illness predicted and the mixed model (Figure 3C). The Hosmer–Lemeshow test resulted in a p value of 0.666, indicating a good fit for the nomogram. The bootstrap validation technique yielded a p value of 0.854, confirming the model’s acceptable performance.

Figure 3D illustrates the DCA, demonstrating that the use of a nomogram in predicting severe illness provided a higher net benefit compared to both the “treat all” and “treat none” strategies across threshold probabilities ranging from 1 to 81% in the cohort. This highlights the clinical utility of the nomogram.

There was a significant difference in predictive efficacy between the mixed model with categorical variables and with continuous variables (p = 0.009) (Supplementary Figure 1).