Prediction of Post-COVID-19 Pulmonary Fibrosis: An Integrated Approach Combining CT Radiomics and Clinical Characteristics

By Mujuan Wang¹, Pei Huang¹, Min Jiang², Bing Fan¹

Affiliations

1. Department of Radiology, Jiangxi Provincial People’s Hospital, The First Affiliated Hospital of Nanchang Medical College, Nanchang, China
2. Department of Respiratory Disease, Jiangxi Provincial People’s Hospital, The First Affiliated Hospital of Nanchang Medical College, Nanchang, China

doi: 10.29271/jcpsp.2026.08.1069

ABSTRACT
Objective: To develop and validate a predictive model for post-COVID-19 pulmonary fibrosis (PCPF) by integrating CT radiomics features with clinical characteristics to facilitate early identification and intervention.
Study Design: An observational study.
Place and Duration of the Study: Department of Radiology, Jiangxi Provincial People’s Hospital, The First Affiliated Hospital of Nanchang Medical College, Nanchang, China, from December 2022 to January 2023.
Methodology: This study enrolled 223 patients with COVID-19. Chest CT images and clinical data of the participants during their hospitalisation were collected. Follow-up chest CT scans were performed 3-12 months post-discharge to assess for PCPF. Participants were randomly divided into a training set (n = 156) and a testing set (n = 67). Using the least absolute shrinkage and selection operator (LASSO) regression, six optimised radiomic features were identified, and radiomic scores were calculated (Rad-scores). Univariate and multivariate logistic regression analyses screened clinical features, identifying age, lesion location, length of hospital stay, and lactate dehydrogenase (LDH) as independent predictors. Subsequently, a radiomic model, a clinical model, and a combined nomogram model incorporating Rad-scores and clinical predictors were constructed.
Results: The combined nomogram demonstrated superior prediction. In the testing set, the areas under the curve (AUC) were 0.833, outperforming the clinical (AUC = 0.687) and radiomic (AUC = 0.811) models. The DeLong test confirmed that the nomogram significantly outperformed the clinical model (p < 0.05). Calibration and decision curve analyses verified the nomogram's excellent fit and provided substantial clinical net benefit.
Conclusion: The nomogram model based on CT radiomics and clinical features is promising for predicting PCPF.

Key Words: Computed tomography, COVID-19, Post COVID-19 pulmonary fibrosis, Radiomics, Nomogram.

INTRODUCTION

By the end of 2019, COVID-19, caused by severe acute respi- ratory syndrome coronavirus 2 (SARS-CoV-2), had become a global pandemic. COVID-19 has been demonstrated to cause post-COVID-19 pulmonary fibrosis (PCPF), significantly impac-ting the quality of life of affected patients. The pathogenesis of PCPF can be attributed to several factors.

   First, the rapid replication of SARS-CoV-2, coupled with the consequent death of epithelial cells, may stimulate alveolar macrophages to secrete excessive amounts of cytokines (such as TNF, IL-1β, IL-6, MIP-1, IFN-γ, and VEGF). This excessive cyto-kine release can lead to the phenomenon known as cytokine storms, which exacerbates pulmonary injury and facilitates the onset of acute respiratory distress syndrome (ARDS).^1,2 One recognised sequela of ARDS is PCPF. Second, SARS-CoV-2 may bind to ACE2 receptors, thereby increasing fibrogenic angio-tensin 1 and 2 signalling and alveolar TGF-β signalling, which may stimulate the PCPF pathway.^3,4

As PCPF may cause irreversible lung damage and pulmonary function loss, early identification and prediction are crucial for improving the prognosis of COVID-19 by intervening in the PCPF progression.^5,6 However, conventional imaging examinations have limitations, including a lag in detection and difficulty reversing the typical imaging manifestations of PCPF once they appear.⁷

There is an urgent need to develop new medical imaging analysis methods to fully utilise the in-depth data contained in medical images for clinical information analysis, enabling early assessment and prediction of PCPF.

Radiomics, proposed in 2012, is a method that identifies quantitative features in medical images using high-throughput technology to improve the diagnostic and predictive performance of medical systems.^8,9 Radiomics is primarily used for prognosis assessment, non-invasive disease tracking, and treatment efficacy evaluation.¹⁰ Several studies have used CT radiomics to identify COVID-19 or assess its prognosis, demonstrating that radiomic models have significant value in disease diagnosis and prediction.^11,12 To the best of the authors’ knowledge, the application of radiomics in predicting PCPF has not yet been reported. The objective of this study was to develop and validate a predictive model for PCPF by integrating CT radiomics features with clinical characteristics to facilitate early identification and intervention.

METHODOLOGY

This was a retrospective observational study conducted in the Department of Radiology, Jiangxi Medical College of Nanchang University, Nanchang, China, from December 2022 to January 2023. Ethical approval was obtained from the hospital’s Ethical Review Board (Approval No. 2021-026). Given that this investigation was retrospective, the requirement for obtaining informed consent from patients was waived. A total of 1,376 patients were diagnosed with COVID-19 in Jiangxi Provincial People's Hospital. The diagnostic criteria for COVID-19 were based on the Diagnostic and Therapeutic Protocol for Novel Coronavirus Pneumonia (10th Trial Edition) issued by the National Health Commission.¹³ From these patients, 223 participants were enrolled in the study. The diagnostic criteria for PCPF were based on the Fleischner Society definitions and included traction bronchi-ectasis, grid shadow, and honeycombing.

The study included patients with COVID-19 confirmed by nasopharyngeal swab RT-PCR. Patients were excluded if they had initial or prior CT scans demonstrating definite pulmonary fibrosis, or if their CT images were of suboptimal quality or contained significant artefacts due to poor breath-holding. Additional exclusion criteria included reinfection or other pulmonary infections after discharge; chest CT findings indicative of pulmonary diseases such as tuberculosis, lung cancer, or pneumoconiosis; incomplete clinical data or the absence of a chest CT scan performed between 3 months and 1 year after disease onset.

Clinical data collected from participants included age, gender, underlying diseases, lesion location, length of hospital stay, ICU admission, clinical classification, white blood cell (WBC) count, neutrophil (NEUT#) count, lymphocyte (LYM#) count, lactate dehydrogenase (LDH) level, D-dimer concentration, and C-reactive protein (CRP) concentration.

Patients underwent CT examinations in the supine position, with the head in the first position, and were scanned during breath-hold. The scanning parameters were 120 kV, 100-250 mAs, a layer thickness of 5 mm, a pitch of 1-1.5, and a matrix of 512 × 512. No contrast agent enhancement was perfor-med. All images were reconstructed using high-resolution and conventional algorithms, with a reconstruction layer thickness of 1-2 mm.

Based on the initial CT images of COVID-19, the region of interest (ROI) was manually outlined in the lesion maximal cross-sections using the DARWIN research platform (Beijing Yizhun Intelligent Technology Co., LTD., China; httpss://arxiv. org/abs/2009.00908). All lesion delineations were performed in the lung window (width, 1000HU; level, -700HU). The ROI was drawn along the lesion edge as precisely as possible. A junior radiologist manually delineated the ROI for all patients, and a senior radiologist reviewed the delineations. Both physicians independently assessed the CT features, blinded to the patients' PCPF status, including traction bronchiectasis, grid shadow, and honeycomb lung, and lesion location (subpleural, diffuse, or centrilobular distribution). In cases of disagreement, a consensus was reached after a discussion between the two physicians.

Radiomic features were automatically extracted from each ROI using PyRadiomics (httpss://pyradiomics.readthedocs.io), resulting in 1032 features per ROI. The feature values were normalised to the [0, 1] range using min–max normalisation and subsequently standardised. High-relevance features were selected utilising the maximum relevance minimum redundancy (mRMR) algorithm. The least absolute shrinkage and selection operator (LASSO) was utilised to enhance and optimise the feature set, selecting the most appropriate λ value to retain highly relevant feature subsets while avoiding over-fitting. This process included five-fold cross-validation. The final selected features were entered into a support vector machine (SVM) model to develop the predictive model, and a radiomics score (Rad-score) was calculated.

Three distinct models were developed: a clinical model, a radiomic model, and a nomogram model. These models were constructed utilising independent clinical predictors alongside Rad-scores. The predictive performance of the models was assessed by calculating the area under the receiver operating characteristic curve (AUC-ROC), as well as sensitivity, specificity, and overall accuracy. The predictive performance of each model for assessing the risk of PCPF was compared using the DeLong test. Model calibration was evaluated using calibration curves and the Hosmer–Lemeshow (H–L) test. Finally, decision curve analysis (DCA) was performed to determine the clinical net benefit of each model across a range of threshold probabilities.

Statistical analyses were conducted using SPSS 24.0 and the R language (version 4.3.2). Normally distributed data were expressed as mean ± SD, with comparisons made using the independent t-test.

Table I: Clinical features of PCPF and non-PCPF in the training set and testing set.

Variables	Training set (n = 156)		p-values	Testing set (n = 67)		p-values
	non-PCPF (n = 95)	PCPF (n = 61)		non-PCPF (n = 41)	PCPF (n = 26)
Age	63.00 (45.00, 75.50)	76.00 (63.00, 83.00)	<0.001	60.00 (51.00, 72.00)	70.00 (59.25, 83.75)	0.023
Length of hospital stay	11.00 (8.00, 15.00)	16.00 (10.00, 22.00)	<0.001	13.00 (7.00, 16.00)	15.50 (11.25, 19.75)	0.037
Gender, n (%)			0.56			0.187
Female	33 (34.74)	24 (39.34)		14 (34.15)	5 (19.23)
Male	62 (65.26)	37 (60.66)		27 (65.85)	21 (80.77)
Basic diseases, n (%)			0.1			0.061
No	84 (88.42)	48 (78.69)		36 (87.80)	18 (69.23)
Yes	11 (11.58)	13 (21.31)		5 (12.20)	8 (30.77)
Lesion location, n (%)			<0.001			0.4
Subpleural distribution	29 (30.53)	12 (19.67)		16 (39.02)	6 (23.08)
Diffuse distribution	9 (9.47)	24 (39.34)		5 (12.20)	4 (15.38)
Centrilobular distribution	57 (60.00)	25 (40.98)		20 (48.78)	16 (61.54)
Admission to ICU, n (%)			0.442			>0.99
No	90 (94.74%)	55 (90.16)		39 (95.12)	24 (92.31)
Yes	5 (5.26%)	6 (9.84)		2 (4.88)	2 (7.69)
Clinical classification, n (%)			0.004			0.254
Light/medium	75 (78.95)	35 (57.38)		32 (78.05)	17 (65.38)
Heavy/critical	20 (21.05)	26 (42.62)		9 (21.95)	9 (34.62)
WBC (×10⁹/L)	5.51 (4.38, 8.82)	7.10 (4.97, 10.25)	0.062	6.34 (4.15, 9.00)	7.28 (4.14, 10.58)	0.629
LYM# (×10⁹/L)	15.60 (6.80, 22.95)	11.00 (5.80, 17.70)	0.027	15.10 (7.60, 23.10)	10.80 (4.72, 18.25)	0.183
NEUT# (×10⁹/L)	4.07 (2.90, 6.55)	5.71 (3.50, 8.90)	0.032	4.74 (2.90, 7.70)	5.81 (2.91, 8.69)	0.52
LDH (IU/L)	234.00 (197.00, 261.00)	268.00 (223.00, 346.00)	0.001	241.00 (195.00, 280.00)	251.00 (193.75, 330.25)	0.396
CRP (mg/L)	34.00 (15.34, 63.00)	38.30 (17.00, 92.00)	0.153	17.70 (7.00, 71.12)	44.60 (13.55, 99.62)	0.098
D-dimer (mg/L)	0.38 (0.23, 0.66)	0.69 (0.40, 1.62)	<0.001	0.49 (0.23, 0.87)	0.47 (0.34, 0.59)	0.872
Note: The data are presented as median (interquartile range) or number (%). Underlying diseases included hypertension, diabetes mellitus, coronary artery disease, cerebral infarction, and a history of tumours; p <0.05 denotes no statistically significant differences. The p-values were determined using the Mann–Whitney U test for non-normally distributed continuous variables (median [IQR]) and the chi-square test for categorical variables (n, %).

Table II: Univariate and multivariate analyses of clinical features.

Variables	Univariate analysis	p-values	Multivariate analysis	p-values
	OR (95% CI)		OR (95% CI)
Gender	0.82 (0.42-1.59)	0.56	-	-
Age	1.04 (1.02-1.07)	p <0.001	1.04 (1.02-1.06)	0.02
Basic diseases	2.07 (0.86-4.98)	0.105	-	-
Lesion location	-	-	-	-
Centrilobular distribution	1	-	-	-
Diffuse distribution	6.44 (2.33-17.86)	p <0.001	5.02 (1.6-15.82)	0.003
Subpleural distribution	1.06 (0.47-2.41)	0.889	0.92 (0.36-2.35)	0.639
Length of hospital stay	1.09 (1.03-1.14)	0.001	1.09 (1.04-1.16)	0.001
Admission to the ICU	1.96 (0.57-6.74)	0.284	-	-
Clinical classification	2.79 (1.37-5.65)	0.005	-	-
WBC	1.08 (1-1.17)	0.065	-	-
LYM#	0.96 (0.93-1)	0.028	-	-
NEUT#	1.09 (1-1.18)	0.045	-	-
LDH	1.01 (1-1.01)	0.001	1 (1-1.01)	0.126
CRP	1 (1-1.01)	0.214	-	-
D-dimer	1 (0.95-1.04)	0.886		-
p-values in the univariate analysis were calculated using univariate logistic regression. p-values in the multivariate analysis were derived from multivariate logistic regression using stepwise selection based on the AIC.

Table III: AUC, sensitivity, specificity, and accuracy of different models in the training set and testing set.

Variables	AUC (95% CI)	Sensitivity	Specificity	Accuracy
Training set	-	-	-	-
Radiomic nomogram	0.891 (0.836~0.945)	0.869	0.832	0.846
Clinical model	0.813 (0.744~0.882)	0.738	0.789	0.769
Radiomic model	0.859 (0.793~0.925)	0.836	0.842	0.840
Testing set	-	-	-	-
Radiomic nomogram	0.833 (0.733~0.933)	0.731	0.854	0.810
Clinical model	0.687 (0.557~0.817)	0.462	0.854	0.701
Radiomic model	0.811 (0.697~0.924)	0.654	0.927	0.821

Skewed distribution data were expressed as median (percentile). Categorical variables were presented as absolute numbers (n) and proportions (%), with comparative analyses performed using the chi-square test or Fisher's exact test. Clinical features were analysed using univariate and multivariate regression, as well as multivariable stepwise logistic regression using R software. The variables included in the univariable logistic regression analysis were selected on the basis of the available clinical data. The variables that survived the univariate statistical significance test were entered into a multivariate logistic regression, with a p-value of < 0.10 used as the criterion for inclusion.

Figure 1: ROC curves comparing model performances in (A) the training set and (B) the testing set. The radiomics nomogram (C) was established based on an independent clinical predictor and the Rad-score.

This threshold ensures that potentially important predictors are considered in the final multivariable model while reducing the likelihood of excluding relevant predictors because of adjustment for other variables. The feature set with the minimum Akaike information criterion (AIC) was retained. The radiomic, clinical, and nomogram models were constructed using the filtered variables and Rad-score. A p-value of <0.05 was considered statistically significant.

RESULTS

A total of 223 COVID-19 patients were enrolled in the study, including 136 non-PCPF cases (89 males, 47 females) and 87 PCPF cases (58 males, 29 females). Participants were randomly assigned in a 7:3 ratio using a random number generator into a training set (n = 156) and a testing set (n = 67).

Figure 2: The calibration curve for the radiomic nomogram in (A) the training set and (B) the testing set. The analysis of decision curves for various models (C).

Table I summarises the demographic and clinical features of the participants. Univariate analysis revealed that the PCPF group was older and had a longer hospital stay, wider lesion distribution, and a more severe clinical classification in the training set. Additionally, compared to the non-PCPF group, the PCPF group exhibited elevated neutrophil count, LDH, and D-dimer levels, while lymphocyte count did not show a statistically significant increase.

Six optimised radiomic features were developed using LASSO regression, and ROC was performed to evaluate the predictive performance of the radiomic label to assess the correlation among the six radiomic features. The Rad-score was calculated based on regression coefficients.

Clinical features and statistically significant differences identified through univariate analysis were included in a multivariate analysis, performed using a stepwise regression method based on the AIC. The feature set with the minimum AIC was retained. Ultimately, age, length of hospital stay, lesion location, and LDH were identified to be independent risk factors for PCPF. Independent risk factors for PCPF were confirmed using a two-stage statistical procedure. A univariable analysis was first performed to identify variables significantly associated with PCPF based on p-values. Statistically significant variables were subsequently entered into a multivariable logistic regression model, with stepwise variable selection performed according to the AIC. The procedure involved iteratively adding or removing variables to identify the model with the lowest AIC, which reflected the best model with the fewest parameters. Age, length of hospital stay, lesion location, and LDH remained in the final model and were identified as independent risk factors, each contributing significantly to the prediction of PCPF after adjustment for the other variables (Table II). Clinical prediction models were then established based on these factors.

A nomogram model was constructed using logistic regression based on age, length of hospital stay, lesion location, LDH, and Rad-score. Table III summarises the sensitivity, specificity, accuracy, and AUC of different models. Figure 2 A-B illustrates the ROC curves for different models as evaluated on both the training and testing datasets. The nomogram model exhibited the highest AUC values, surpassing both the clinical and Radiomic models. The results of the DeLong test indicated that the nomogram model provided significantly improved predictive accuracy compared with the clinical model in both datasets (p <0.05). However, no statistically significant difference was found between the nomogram and the radiomic model (p >0.05).

To enhance the evaluation of these models' efficacy, continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) metrics were employed. In the training set, NRI demonstrated a 61% improvement in predictive performance for the nomogram model compared with the radiomic model (p <0.001), and IDI showed a 5% improvement (p = 0.01). In the testing set, NRI indicated a 44% improvement in predictive performance for the nomogram model compared to the radiomic model (p = 0.069), and IDI showed an 8% improvement (p = 0.029). Overall, the nomogram had the best predictive performance. The variables included in the radiomic nomogram were visualised using the nomogram (Figure 1C). According to the calibration curves, the probabilities predicted by the radiomics nomogram showed good agreement with the actual probabilities of PCPF in both the training and testing cohorts (Figure 2A–B). The results of the H–L test indicated satisfactory goodness-of-fit, with p-values of 0.514 and 0.959 for the training and testing cohorts, respectively. Furthermore, the DCA results revealed that both the nomogram and radiomic models provided greater net benefits within risk thresholds ranging from 0.13 to 0.75 than the clinical model (Figure 2C).

DISCUSSION

A radiomic nomogram was established by integrating radiomic features with clinical characteristics to forecast the likelihood of PCPF. The resulting nomogram exhibited enhanced predictive accuracy, with AUC values of 0.891 in the training cohort and 0.833 in the validation cohort, surpassing the performance of radiomic and clinical models when utilised independently. This model effectively highlighted individuals predisposed to PCPF, facilitating timely interventions and improving clinical decision-making.

In this study, age, length of hospital stay, lesion location, and LDH were identified as independent risk factors for PCPF. The identification of age as a significant risk factor for COVID-19 and PCPF is in line with extant literature.¹⁴ The extended length of hospital stay observed (16 days vs. 11 days) likely correlates with disease severity; more severe cases require more prolonged hospitalisations and are more prone to developing PCPF lesions.¹⁵ Although the percentage of diffuse distribution and centrilobular distribution of lesion location in the PCPF group showed no significant difference (39.34% vs. 40.98%), univariate analysis revealed that diffuse distribution was a risk factor for PCPF compared to centrilobular distribution (OR = 6.44, p <0.001). This suggests that widespread lesions may stimulate more alveolar macrophages to produce cytokines, contributing to PCPF. Additionally, the higher LDH levels in the PCPF group (268 mmol/L vs. 234 mmol/L) may be attributed to inflammation and oxidative stress, which can lead to cellular damage and enzyme release.^16,17

The majority of COVID-19 radiomics studies primarily focus on predicting short-term outcomes such as mortality or ICU admission.^18,19 In contrast, this study addresses the less-explored challenge of predicting PCPF. The six optimised radiomic features extracted in this study, including four texture features, one shape feature, and one first-order statistical feature, capture comprehensive, visually imperceptible pulmonary changes. This combined radiomic-clinical model enhances understanding of PCPF progression and provides a tool for early, individualised prediction, thereby filling a critical gap in long-term patient management.

This study has several limitations. As a single-centre retrospective study, it may contain unavoidable selection bias, and the results require validation through multi-centre, large-sample studies. Additionally, the criteria for diagnosing PCPF were based on chest imaging and not confirmed by pathology.²⁰

CONCLUSION

Three predictive models were created utilising CT radiomics alongside clinical characteristics, and their performance in predicting PCPF was assessed. The optimised model led to the construction of a nomogram, which was validated and demonstrated good calibration and discrimination. This nomogram provides a potentially reliable tool for the individualised assessment of PCPF risk, facilitating clinical decision- making and optimising prevention and treatment strategies.

FUNDING:
This study was supported by the National Natural Science Foundation of China (Grant No. 82160335).

ETHICAL APPROVAL: 
This study was conducted after obtaining approval from the Ethics Committee of the Department of Radiology, Jiangxi Provincial People’s Hospital, the First Affiliated Hospital of Nanchang Medical College, Nanchang, China (Approval No. 2021-026).

PATIENTS’ CONSENT: 
Informed consent was obtained from patients to publish the data concerning this case.

COMPETING INTEREST: 
The authors declared no conflict of interest.

AUTHORS’ CONTRIBUTIONS: 
MW, PH: Data collection and analysis.
MW, MJ, BF: Conception of the study, manuscript drafting, and review.
PH, BF: Data analysis and drafting.
All the authors approved the final version of the manuscript to be published.
 

REFERENCES

1. Fajgenbaum DC, June CH. Cytokine Storm. New Engl J Med 2020; 383(23):2255-73. doi: 10.1056/NEJMra2026131.
2. Matthay MA, Zemans RL, Zimmerman GA, Arabi YM, Beitler JR, Mercat A, et al. Acute respiratory distress syndrome. Nat Rev Dis Primers 2019; 5(1):18. doi: 10.1038/s41572-019- 0069-0.
3. Allen RJ, Stockwell A, Oldham JM, Guillen-Guio B, Schwartz DA, Maher TM, et al. Genome-wide association study across five cohorts identifies five novel loci associated with idiopathic pulmonary fibrosis. Thorax 2022; 77(8):829-33. doi: 10. 1136/thoraxjnl-2021-218577.
4. Shao H, Qin Z, Geng B, Wu J, Zhang L, Zhang Q, et al. Impaired lung regeneration after SARS-CoV-2 infection. Cell Prolif 2020; 53(12):e12927. doi: 10.1111/cpr.12927.
5. Bai T, Zhou D, Yushanjiang F, Wang D, Zhang D, Liu X, et al. Alternation of the autonomic nervous system is associated with pulmonary sequelae in patients with COVID-19 after six months of discharge. Front Physiol 2021; 12:805925. doi: 10.3389/fphys.2021.805925.
6. Huang L, Li X, Gu X, Zhang H, Ren L, Guo L, et al. Health outcomes in people 2 years after surviving hospitalisation with COVID-19: A longitudinal cohort study. Lancet Resp Med 2022; 10(9):863-76. doi: 10.1016/S2213-2600(22)00 126-6.
7. Franquet T. Imaging of pulmonary viral pneumonia. Radiology 2011; 260(1):18-39. doi: 10.1148/radiol.11092149.
8. Dong D, Tang L, Li ZY, Fang MJ, Gao JB, Shan XH, et al. Development and validation of an individualised nomogram to identify occult peritoneal metastasis in patients with advanced gastric cancer. Ann Oncol 2019; 30(3):431-8. doi: 10.1093/annonc/mdz001.
9. Lambin P, Rios-Velazquez E, Leijenaar R, Carvalho S, van Stiphout RGPM, Granton P, et al. Radiomics: Extracting more information from medical images using advanced feature analysis. Eur J Cancer 2012; 48(4):441-6. doi: 10.1016/j.ejca.2011.11.036.
10. Rizzo S, Botta F, Raimondi S, Origgi D, Fanciullo C, Morganti AG, et al. Radiomics: The facts and the challenges of image analysis. Eur Radiol Exp 2018; 2(1):36. doi: 10.1186/s41 747-018-0068-z.
11. Li C, Dong D, Li L, Gong W, Li X, Bai Y, et al. Classification of severe and critical Covid-19 using deep learning and radiomics. IEEE J Biomed Health 2020; 24(12):3585-94. doi: 10.1109/JBHI.2020.3036722.
12. Wang L, Kelly B, Lee EH, Wang H, Zheng J, Zhang W, et al. Multi-classifier-based identification of COVID-19 from chest computed tomography using generalizable and interpretable radiomics features. Eur J Radiol 2021; 136:109552. doi: 10. 1016/j.ejrad.2021.109552.
    
     
13. Released by National Health Commission of People's Republic of China & National Administration of Traditional Chinese Medicine on January 5, 2023. Diagnosis and treatment protocol for COVID-19 patients (Tentative 10^th Version). Health Care Sci 2023; 2(1):10-24. doi: 10.1002/ hcs2.36.
14. Verity R, Okell LC, Dorigatti I, Winskill P, Whittaker C, Imai N, et al. Estimates of the severity of coronavirus disease 2019: A model-based analysis. Lancet Infect Dis 2020; 20(6): 669-77. doi: 10.1016/S1473-3099(20)30243-7.
15. Rai DK, Sharma P, Kumar R. Post COVID-19 pulmonary fibrosis. Is it real threat? Indian J Tuberc 2021; 68(3):330-3. doi: 10.1016/j.ijtb.2020.11.003.
16. Cameli P, Carleo A, Bergantini L, Landi C, Prasse A, Bargagli E. Oxidant/antioxidant disequilibrium in idiopathic pulmonary fibrosis pathogenesis. Inflammation 2020; 43(1):1-7. doi: 10.1007/s10753-019-01059-1.
17. Sakhaee MH, Sayyadi SAH, Sakhaee N, Sadeghnia HR, Hosseinzadeh H, Nourbakhsh F, et al. Cedrol protects against chronic constriction injury-induced neuropathic pain through inhibiting oxidative stress and inflammation. Metab Brain Dis 2020; 35(7):1119-26. doi: 10.1007/s11011- 020-00581-8.
18. Shiri I, Sorouri M, Geramifar P, Nazari M, Abdollahi M, Salimi Y, et al. Machine learning-based prognostic modelling using clinical data and quantitative radiomic features from chest CT images in COVID-19 patients. Comput Biol Med 2021; 132:104304. doi: 10.1016/j.compbiomed.2021.104304.
19. Chao H, Fang X, Zhang J, Homayounieh F, Arru CD, Digumarthy SR, et al. Integrative analysis for COVID-19 patient outcome prediction. Med Image Anal 2021; 67:101844. doi: 10.1016/j.media.2020.101844.
20. Zhang P, Li J, Liu H, Han N, Ju J, Kou Y, et al. Long-term bone and lung consequences associated with hospital-acquired severe acute respiratory syndrome: A 15-year follow-up from a prospective cohort study. Bone Res 2020; 8:8. doi: 10.1038/s41413-020-0084-5.