Early prediction of severe Omicron pneumonia using a multimodal a.i. model integrating delta CT radiomics and laboratory indicators Authors: Xiaoxian Ye, Xuhao Dai, Shengping Gong, Yingying Zhou, Ting Zhu, Binbin Song, Jiming Yang, Xiaoqin Ge, Jiangping Ren, Cong Shi, Yijian Cao Published: 2026 ABSTRACT AbstractEarly identification of patients at risk of severe pneumonia during Omicron SARS-CoV-2 infection is critical for optimizing care and allocating resources. While clinical markers provide insights, imaging-derived radiomics features may enhance prognostic accuracy. We developed a multimodal predictive model combining Delta Radiomics features from serial chest CT scans with clinical data, including blood biochemical markers and lymphocyte subsets. The primary prediction target was severe/critical Omicron pneumonia during hospitalization. Mild and moderate cases were grouped as non-severe disease, whereas severe and critical cases were defined as the severe class for binary classification. The model was trained on 91 patients from the first center, internally validated on 23 patients, and externally tested on 32 patients from a second center. Machine learning algorithms including Logistic Regression, Random Forest, and MLP were applied, and a nomogram was constructed for individualized risk prediction. The combined model showed high discrimination in the training cohort and maintained favorable performance in the internal validation and independent external test cohorts, achieving AUCs of 0.885 and 0.875, respectively. The Delta Radiomics signature, particularly with MLP, showed comparatively stable predictive performance. These findings suggest the added value of temporal CT-derived radiomics when integrated with clinical biomarkers, although further validation in larger prospective cohorts is required. Integrating temporal imaging features with clinical data offers a non-invasive method for early prediction of severe/critical Omicron pneumonia, supporting individualized triage and more efficient allocation of medical resources.Keywords: Omicron COVID-19, Delta radiomics, Clinical biomarkers, Prognostic risk stratification, Machine learning, NomogramSubject terms: Biomarkers, Computational biology and bioinformatics, Diseases, Health care, Mathematics and computing, Medical research Introduction Since the outbreak of coronavirus disease 2019 (COVID-19), infection caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) has imposed a sustained burden on global healthcare systems1. Although the virulence of emerging variants has evolved over time, a subset of patients still experience rapid clinical deterioration due to dysregulated inflammatory responses and immune imbalance2,3. Previous studies have demonstrated that elevated inflammatory markers, abnormal blood biochemical parameters, and alterations in lymphocyte subsets are closely associated with disease severity and adverse outcomes4–6. These findings highlight the critical role of host immune-inflammatory status in the progression of COVID-19 and suggest potential biomarkers for early risk stratification. Chest computed tomography (CT) plays a central role in assessing pulmonary involvement and monitoring disease evolution in COVID-197. Typical imaging manifestations, including ground-glass opacities, consolidation, and interstitial changes, reflect the extent of inflammatory infiltration and lung injury. With the advancement of quantitative imaging analysis, radiomics enables the extraction of high-dimensional features that characterize tissue heterogeneity beyond visual assessment and has been widely applied in disease stratification and prognostic prediction8. Several studies have demonstrated that models based on single time-point CT radiomics features can effectively predict disease severity and clinical outcomes in patients with COVID-199–12. Furthermore, integrating imaging features with clinical variables has been shown to improve predictive performance compared with single-modality models10,12. These findings provide a foundation for the application of quantitative imaging in COVID-19 risk assessment. Prior AI-based studies, including the multimodal CT analysis framework reported by Xu et al., have also supported the feasibility of CT-based rapid triage in COVID-19 patients, providing an important methodological context for the present study36. However, COVID-19 is inherently a time-dependent inflammatory disease. Previous imaging studies have shown that pulmonary lesions may undergo substantial changes within only a few days after symptom onset, exhibiting a dynamic evolutionary pattern7,13. A single CT examination reflects only one stage of the disease process and may fail to fully capture the rate of inflammatory progression and lesion evolution, thereby limiting the robustness and generalizability of predictive models. In early stages, subtle changes in lesion burden may already indicate divergent clinical trajectories, yet static imaging features may not adequately detect such dynamics. Delta radiomics, an extension of conventional radiomics, quantifies temporal changes in imaging features between sequential examinations and has emerged as a promising strategy for evaluating disease progression14. In oncology and other inflammatory conditions, delta radiomics has demonstrated superior predictive performance compared with static features by more sensitively reflecting underlying biological changes14,15. Nevertheless, in the context of COVID-19, studies investigating short-interval CT–derived delta radiomics features for early risk prediction remain limited, and the independent contribution of dynamic imaging features within multimodal predictive frameworks has not been systematically evaluated. Given the close relationship between pulmonary imaging alterations and inflammatory responses in Omicron variant COVID-19, dynamic quantitative imaging features may more accurately reflect inflammatory burden and disease evolution. Therefore, in this study, we extracted delta radiomics features derived from baseline chest CT and a follow-up CT scan conducted 5 to 9 days later as the primary imaging variables. These features were integrated with blood biochemical parameters and lymphocyte subset profiles to develop a comprehensive binary classification model for early prediction of severe/critical Omicron pneumonia. We hypothesized that, compared with models based solely on single time-point radiomics features, delta radiomics features could more sensitively capture early disease progression and, when combined with clinical biomarkers, further enhance risk stratification performance. This study aimed to evaluate the independent predictive value of delta radiomics and to explore its association with inflammatory and immune-related indicators, thereby providing quantitative evidence to support early individualized triage in Omicron COVID-19. An overview of the entire research pipeline is illustrated in Fig. 1. Workflow of this study. Method Data acquisition This multicenter retrospective study included patients with laboratory-confirmed SARS-CoV-2 infection admitted between November 29, 2022, and January 15, 2023, at the First Affiliated Hospital of Ningbo University (n = 114) and the First Hospital of Jiaxing (n = 32). Patients had no documented prior SARS-CoV-2 infection and had not received prior symptomatic treatment before admission; all patients were included during the Omicron-predominant wave according to local epidemiological surveillance. Clinical data, chest CT images, hematological tests, biochemical parameters, and complete in-hospital clinical course data were collected. The primary prediction endpoint was severe/critical Omicron pneumonia during hospitalization, rather than all-cause mortality. Inclusion criteria were as follows: laboratory-confirmed SARS-CoV-2 infection during the Omicron-predominant period; availability of baseline and follow-up chest CT images; completion of routine blood, biochemical, and immune-related tests; and complete clinical and follow-up information. Exclusion criteria were as follows: lack of follow-up CT scans taken 5 to 9 days after the initial chest CT; severe immune disorders; or long-term immunosuppressive therapy that could interfere with immune evaluation. Patients were initially classified into low-, intermediate-, and high-risk groups according to respiratory symptoms, oxygen requirement, evidence of respiratory distress or ARDS, ICU-related clinical deterioration, and the extent of pulmonary involvement on CT. For model development, low- and intermediate-risk patients were combined as the non-severe class, whereas high-risk patients were defined as the severe/critical class. The Ningbo cohort was randomly split into training and internal validation sets at an 8:2 ratio, with the Jiaxing cohort serving as an independent external test cohort. The study was conducted in accordance with the Declaration of Helsinki and was approved by the ethics committees of both institutions (Approval Nos. 2026003RS and 2026-LP-019), with informed consent waived. Table 1 presents the baseline clinical characteristics of the training, internal validation, and external test cohorts, including age, inflammatory markers, immune cell counts, and other clinical parameters. Table 1Baseline clinical characteristics of the training, internal validation, and independent external test cohorts.Feature_nameOverallExternal test cohortTraining cohortInternal validation cohortp-valueAGE73.61 ± 13.1874.62 ± 15.2873.59 ± 11.9772.26 ± 15.030.808CRP_mg_L70.66 ± 65.1064.69 ± 54.0972.52 ± 68.9571.61 ± 65.500.842WBC9.48 ± 6.7710.46 ± 4.759.40 ± 7.508.43 ± 6.100.541NE7.71 ± 4.848.98 ± 4.447.43 ± 4.707.02 ± 5.740.225L#1.21 ± 4.790.88 ± 0.731.43 ± 6.050.82 ± 0.501.000RBC3.84 ± 0.723.70 ± 0.623.86 ± 0.723.94 ± 0.860.441HB_g_L119.45 ± 21.22116.22 ± 19.11119.55 ± 21.35123.57 ± 23.580.450PLT187.14 ± 79.18201.12 ± 78.97179.77 ± 74.12196.83 ± 97.220.347ApoA10.87 ± 0.220.79 ± 0.150.90 ± 0.250.84 ± 0.140.034ApoB0.72 ± 0.270.71 ± 0.240.74 ± 0.300.68 ± 0.180.609B2M5.23 ± 6.365.05 ± 5.475.63 ± 7.273.90 ± 2.390.500IL_680.62 ± 308.7056.93 ± 129.30102.99 ± 382.3025.04 ± 22.600.497IL_1015.79 ± 20.0912.38 ± 9.4318.10 ± 24.5311.35 ± 3.660.197SAA457.76 ± 493.80381.85 ± 454.96498.39 ± 532.83402.62 ± 368.210.439Apo_E50.84 ± 21.1953.22 ± 21.7751.17 ± 21.3846.23 ± 19.810.473CD360.24 ± 15.7862.83 ± 15.9959.48 ± 16.1459.66 ± 14.230.578CD433.35 ± 13.1635.43 ± 13.2632.32 ± 13.1834.52 ± 13.120.466CD824.41 ± 12.8425.50 ± 12.6024.22 ± 13.1323.69 ± 12.470.8534 + 8+0.54 ± 0.440.72 ± 0.580.47 ± 0.330.57 ± 0.531.000CD4_CD8_Ratio2.08 ± 2.741.86 ± 1.322.05 ± 2.732.52 ± 4.020.672CD5624.85 ± 15.2921.09 ± 13.1626.00 ± 16.0025.57 ± 14.920.288CD1913.86 ± 9.4714.82 ± 10.7413.51 ± 9.2313.89 ± 8.810.799GENDER0.229049 (33.56)14 (43.75)30 (32.97)5 (21.74)197 (66.44)18 (56.25)61 (67.03)18 (78.26)Vaccine0.348091 (62.33)21 (65.62)53 (58.24)17 (73.91)155 (37.67)11 (34.38)38 (41.76)6 (26.09)PMH0.693032 (21.92)6 (18.75)22 (24.18)4 (17.39)1114 (78.08)26 (81.25)69 (75.82)19 (82.61)Hypertension0.303054 (36.99)9 (28.12)38 (41.76)7 (30.43)192 (63.01)23 (71.88)53 (58.24)16 (69.57)Cardiovascular_Disease0.2950140 (95.89)30 (93.75)89 (97.80)21 (91.30)16 (4.11)2 (6.25)2 (2.20)2 (8.70)Diabetes_Mellitus0.2010110 (75.34)21 (65.62)73 (80.22)16 (69.57)136 (24.66)11 (34.38)18 (19.78)7 (30.43)Cohort0.594Non-severe80 (54.79)15 (46.88)52 (57.14)13 (56.52)Severe/critical66 (45.21)17 (53.12)39 (42.86)10(43.48)AGE: Patient age; CRP_: C-reactive protein, inflammation marker; WBC: White blood cell count; NE: Neutrophil count; ALC: Lymphocyte count; RBC: Red blood cell count; HB_: Hemoglobin level; ApoA1: Apolipoprotein A1, HDL component; ApoB: Apolipoprotein B, LDL component; B2M: Beta-2 microglobulin, immune activation marker; IL_6: Interleukin-6, inflammatory cytokine; IL_10: Interleukin-10, anti-inflammatory cytokine; SAA: Serum amyloid A, acute-phase protein; ApoE: Apolipoprotein E, lipid metabolism; CD3: T-cell marker; CD4: CD4 T lymphocytes; CD8: CD8 T lymphocytes; 4 + 8+: T-cells expressing both CD4 and CD8; CD4_CD8_Ratio: CD4/CD8 T-cell ratio; CD56: NK cell marker; CD19: B cell marker; GENDER: Gender (0 = male, 1 = female); Vaccine: Vaccination status (0 = unvaccinated, 1 = vaccinated); PMH: Past medical history (0 = no, 1 = yes); Hypertension: Hypertension status (0 = no, 1 = yes); Cardiovascular_Disease: Cardiovascular disease status (0 = no, 1 = yes); Diabetes_Mellitus: Diabetes status (0 = no, 1 = yes). In this table, the column previously labeled as “test” refers to the independent external test cohort from Jiaxing. Baseline clinical characteristics of the training, internal validation, and independent external test cohorts. AGE: Patient age; CRP_: C-reactive protein, inflammation marker; WBC: White blood cell count; NE: Neutrophil count; ALC: Lymphocyte count; RBC: Red blood cell count; HB_: Hemoglobin level; ApoA1: Apolipoprotein A1, HDL component; ApoB: Apolipoprotein B, LDL component; B2M: Beta-2 microglobulin, immune activation marker; IL_6: Interleukin-6, inflammatory cytokine; IL_10: Interleukin-10, anti-inflammatory cytokine; SAA: Serum amyloid A, acute-phase protein; ApoE: Apolipoprotein E, lipid metabolism; CD3: T-cell marker; CD4: CD4 T lymphocytes; CD8: CD8 T lymphocytes; 4 + 8+: T-cells expressing both CD4 and CD8; CD4_CD8_Ratio: CD4/CD8 T-cell ratio; CD56: NK cell marker; CD19: B cell marker; GENDER: Gender (0 = male, 1 = female); Vaccine: Vaccination status (0 = unvaccinated, 1 = vaccinated); PMH: Past medical history (0 = no, 1 = yes); Hypertension: Hypertension status (0 = no, 1 = yes); Cardiovascular_Disease: Cardiovascular disease status (0 = no, 1 = yes); Diabetes_Mellitus: Diabetes status (0 = no, 1 = yes). In this table, the column previously labeled as “test” refers to the independent external test cohort from Jiaxing. Data preprocessing Chest non-contrast CT scans were performed using multi-detector spiral CT scanners from Philips (Netherlands) and Siemens (Germany). The typical acquisition parameters were as follows: tube current of 80–250 mA with automatic exposure control; tube voltage of 100–120 kV; pitch of 0.9–1.2; field of view of 320–380 mm; matrix of 512 × 512; and slice thickness/reconstruction interval of 1.0–1.5 mm. All images were exported in DICOM format for further analysis. To minimize the impact of outliers, all images were standardized using fixed window settings to optimize visualization of pulmonary inflammatory lesions. Spatial normalization was performed by resampling the images to a fixed isotropic resolution to reduce variability in voxel spacing across scanners. Two board-certified radiologists with over 15 years of thoracic imaging experience, blinded to clinical outcomes, independently performed manual segmentation using ITK-SNAP software (version 3.8.0). The ROI was defined as the bilateral lung parenchyma mask including pneumonia-related abnormalities, rather than a tumor ROI. The trachea, main bronchi, pleural effusion, and extrapulmonary structures were excluded. Discrepancies between the two radiologists were resolved by consensus. Interobserver reproducibility was evaluated using intraclass correlation coefficients (ICCs) derived from independently segmented cases, and features with ICCs below 0.70 were excluded before model development. A schematic illustration of the workflow is shown in Fig. 2. Study participant selection flowchart. Feature extraction The extracted features were divided into three main categories: (I) geometry, (II) intensity, and (III) texture. Geometry features described the three-dimensional shape characteristics of the lung ROI and pneumonia-related abnormalities, offering a spatial depiction of pulmonary involvement. Intensity features represented the first-order statistical properties of voxel intensities within the ROI, providing insights into internal uniformity or variability. Texture features captured more complex patterns by detailing the second-order and higher-order spatial distributions of intensities, reflecting the relationships between neighboring voxels. These texture features were derived using methods such as the gray-level co-occurrence matrix (GLCM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM). In total, 1403 handcrafted features were extracted: 23 geometry features, 340 intensity features, and 1040 texture features. Radiomics feature extraction was performed using a standardized analysis workflow based on the open-source PyRadiomics package (http://pyradiomics.readthedocs.io). Feature early fusion: Features from baseline CT (features_pre) and follow-up CT (features_post) scans were combined using feature concatenation and subtraction techniques, resulting in a comprehensive feature vector that captured both static imaging information and short-interval temporal changes: Here, represents element-wise subtraction, and denotes the concatenation of features. Radiomic feature selection To ensure feature reproducibility, we initially excluded features with intra-class correlation coefficients (ICCs) below 0.70. Next, the U-test (p < 0.01) was applied within the training cohort to identify features showing significant differences between the non-severe group and the severe/critical group. Multicollinearity was assessed using Spearman correlation analysis, and features with correlation coefficients > = 0.90 or <=-0.90 were excluded, retaining representative features with robust diagnostic performance. To further reduce redundancy, hierarchical clustering was performed. Clusters with correlation coefficients greater than 0.95 were merged, and a single representative feature from each merged cluster was selected based on the maximum dynamic range. Feature refinement was then conducted using Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression. Cross-validation within the training cohort was employed to determine the optimal regularization parameter lambda by minimizing binomial deviance. Based on the selected features, a machine learning model was developed, and a radiomics score for each patient was calculated as a weighted linear combination of these features. All data-dependent feature selection steps were performed using the training cohort only; the internal validation and external test cohorts were not used for feature screening, LASSO fitting, or model training. Signature construction Delta Radiomics Signature: We developed a radiomics-based risk model by applying feature selection using LASSO and subsequently training multiple machine learning algorithms, including Logistic Regression (LR)16, Naive Bayes17, Support Vector Machine (SVM)18, k-Nearest Neighbors (KNN)19, Random Forest (RF)20, Extra Trees21, XGBoost22, LightGBM23, AdaBoost24, and Multi-Layer Perceptron (MLP)25. Comparative analyses were performed to evaluate the predictive performance of each model. To reduce potential information leakage, all model fitting and hyperparameter tuning were performed in the training cohort, and the learned preprocessing parameters and selected feature sets were then applied unchanged to the internal validation and independent external test cohorts. The final external test cohort was used only for model evaluation. Baseline and follow-up CT Radiomics Signatures: In parallel, radiomics features were extracted separately from baseline and follow-up CT images. These features underwent the same selection and dimensionality reduction procedures to construct baseline and follow-up radiomics signatures. Comparing these signatures with the Delta model allowed us to evaluate whether incorporating short-interval temporal changes increased model stability and predictive capability. Model performance evaluation and development of the comprehensive nomogram The predictive performance of the machine learning models was assessed across the training, internal validation, and independent external test cohorts. Receiver operating characteristic (ROC) curves were constructed to evaluate discriminative ability, with model calibration analyzed through calibration plots and further validated using the Hosmer-Lemeshow goodness-of-fit test. Decision curve analysis (DCA) was employed to examine clinical utility. Model performance metrics included the area under the ROC curve (AUC) with 95% confidence intervals (95% CIs), as well as accuracy, precision, recall, and F1-score. The optimal model was selected primarily according to its performance in the training and internal validation cohorts, whereas the independent external test cohort was reserved for final evaluation. Because of the relatively small sample size and high-dimensional radiomics feature space, we interpreted the training performance cautiously and emphasized external test performance and confidence intervals in the Results and Discussion. Results Clinical characteristics In the analysis, univariate logistic regression identified Apolipoprotein A1 (ApoA1) as a significant predictor, with an inverse association with severe/critical Omicron pneumonia (OR = 0.58, 95% CI: 0.395–0.852, P = 0.02), suggesting a potential protective association. Other clinical factors such as COVID-19 vaccination status, hypertension, diabetes, and immune cell markers (CD19, CD56) did not show significant associations (all P > 0.05). Similarly, IL-6, WBC, CRP, IL-10, and gender exhibited weak or non-significant associations with the outcome in this cohort. The detailed results of the univariate logistic regression analysis are summarized in Table 2. Table 2Univariable logistic regression analysis of clinical characteristics.Feature_nameOR_UNIOR lower 95%CI_UNIOR upper 95%CI_UNIp_value_UNIApoA10.5800.3950.8520.020Vaccine0.6520.3781.1250.198Hypertension0.6560.4141.0420.134PMH0.6830.4561.0220.120Diabetes_Mellitus0.8000.3671.7450.6384 + 8+0.8260.4511.5110.602ApoB0.8260.5341.2760.470RBC0.9370.8571.0230.226CD190.9780.9560.9990.092CD560.9880.9771.0000.103CD80.9920.9791.0040.275CD40.9940.9841.0040.321NE0.9940.9551.0340.785AGE0.9970.9921.0010.212CD30.9970.9911.0020.344B2M0.9970.9601.0350.887PLT0.9980.9961.0000.097HB_g_L0.9980.9951.0010.267Apo_E0.9980.9921.0040.586SAA1.0001.0001.0010.814IL_61.0010.9991.0020.349WBC1.0010.9721.0290.971CRP_mg_L1.0020.9981.0050.429IL_101.0030.9921.0150.662CD4_CD8_Ratio1.0070.9091.1150.908L#1.0190.9561.0870.631GENDER1.1790.7731.7990.523OR_UNI: Odds Ratio from the univariable analysis, representing the strength of association between the feature and the outcome. OR lower 95%CI_UNI: Lower bound of the 95% confidence interval for the univariable odds ratio. OR upper 95%CI_UNI: Upper bound of the 95% confidence interval for the univariable odds ratio. _value_UNI: P-value from the univariable analysis, indicating the statistical significance of the association. Univariable logistic regression analysis of clinical characteristics. OR_UNI: Odds Ratio from the univariable analysis, representing the strength of association between the feature and the outcome. OR lower 95%CI_UNI: Lower bound of the 95% confidence interval for the univariable odds ratio. OR upper 95%CI_UNI: Upper bound of the 95% confidence interval for the univariable odds ratio. _value_UNI: P-value from the univariable analysis, indicating the statistical significance of the association. Lasso feature selection By applying the LASSO logistic regression model, we selected features with non-zero coefficients to construct the radiomics score. The coefficients and the mean squared error (MSE) obtained through 10-fold cross-validation are displayed in Fig. 3. LASSO-Based Selection and Interpretation of Radiomics Features. (A) Coefficient profiles of radiomics features across different lambda values in the LASSO model; the vertical dashed line indicates the optimal lambda = 0.02221, where non-zero coefficients were selected. (B) Mean squared error (MSE) from 10-fold cross-validation, with the optimal lambda = 0.02221 marked by the dashed line. (C) Bar plot of the LASSO-selected radiomics features ranked by coefficient magnitude. Positive coefficients indicate features associated with a higher predicted probability of severe/critical Omicron pneumonia, whereas negative coefficients indicate features associated with a lower predicted probability. The selected features mainly included first-order intensity, wavelet-transformed texture, gray-level dependence, gray-level size-zone, and fractal-related features, reflecting short-interval changes in lesion density, heterogeneity, and spatial complexity. For clinical interpretability, features with larger absolute coefficients in Fig. 3C were considered to contribute more strongly to the Delta radiomics signature. The top contributing features were mainly related to first-order intensity statistics, texture heterogeneity, gray-level zone distribution, wavelet-transformed texture patterns, and fractal-related complexity. Changes in these feature categories may correspond to dynamic alterations in pulmonary inflammatory involvement, including increasing lesion density, ground-glass opacity or consolidation, spatial heterogeneity, and irregular lesion distribution. Models results Based on performance evaluation across cohorts, the Delta models demonstrated more favorable predictive ability and stability than baseline-only or follow-up-only radiomics models. Although several models achieved high AUCs on the training set (> 0.90), performance declines were observed on validation and test sets, indicating a potential risk of overfitting. Among the Delta models, the MLP achieved the best overall balance between training and test performance, with a test AUC of 0.884 and F1-score of 0.825, while maintaining strong training performance (AUC = 0.961, F1 = 0.909). Other Delta-based models, such as Random Forest and XGBoost, also performed well but slightly below MLP. These results suggest that incorporating temporal changes between baseline and follow-up CT features may improve generalizability, although further validation in larger cohorts remains necessary. Comparative metrics are summarized in Table 3, with ROC curves shown in Fig. 4. Table 3Comprehensive performance metrics of the radiomics models in training and test cohorts.TaskModelTrain AUC (95% CI)Test AUC (95% CI)ΔAUCTrain F1Test F1ΔF1PreLR0.900 (0.838–0.960)0.790 (0.669–0.901)0.1100.8310.6810.150PreNaiveBayes0.809 (0.721–0.937)0.825 (0.712–0.939)-0.0160.7210.766-0.045PreSVM0.949 (0.899–0.999)0.803 (0.683–0.923)0.1460.9250.7600.165PreKNN0.906 (0.847–0.964)0.728 (0.601–0.856)0.1780.8050.6270.178PreRandomForest0.957 (0.914–0.999)0.752 (0.623–0.881)0.2050.9160.7370.179PreExtraTrees0.917 (0.863–0.972)0.783 (0.663–0.902)0.1340.8110.6340.177PreXGBoost0.981 (0.955–0.999)0.762 (0.632–0.891)0.2190.9510.7170.234PreLightGBM0.909 (0.844–0.973)0.742 (0.611–0.873)0.1670.8750.6520.223PreAdaBoost0.962 (0.927–0.996)0.705 (0.564–0.846)0.2570.8970.7240.173PreMLP0.909 (0.848–0.968)0.835 (0.726–0.944)0.0740.8430.7500.093PostLR0.883 (0.816–0.948)0.748 (0.617–0.878)0.1350.7860.7460.040PostNaiveBayes0.806 (0.717–0.895)0.774 (0.648–0.901)0.0320.6870.720-0.033PostSVM0.943 (0.902–0.984)0.763 (0.635–0.889)0.1800.8570.7310.126PostKNN0.859 (0.783–0.934)0.724 (0.593–0.855)0.1350.8100.6790.131PostRandomForest0.964 (0.932–0.995)0.672 (0.527–0.816)0.2920.8920.7000.192PostExtraTrees0.901 (0.841–0.960)0.731 (0.597–0.865)0.1700.8000.7380.062PostXGBoost1.000 (0.998-1.000)0.733 (0.599–0.867)0.2670.9870.7040.283PostLightGBM0.950 (0.910–0.989)0.722 (0.587–0.856)0.2280.8710.5850.286PostAdaBoost0.973 (0.947–0.998)0.676 (0.532–0.819)0.2970.9090.6550.254PostMLP0.913 (0.859–0.967)0.715 (0.578–0.851)0.1980.8140.6090.205DeltaLR0.986 (0.970-1.000)0.791 (0.671–0.911)0.1950.9370.7310.206DeltaNaiveBayes0.849 (0.768–0.929)0.792 (0.666–0.918)0.0570.7910.800-0.009DeltaSVM0.990 (0.976-1.000)0.810 (0.693–0.925)0.1800.9490.7870.162DeltaKNN0.911 (0.854–0.967)0.841 (0.735–0.947)0.0700.8420.7940.048DeltaRandomForest0.985 (0.968-1.000)0.716 (0.579–0.852)0.2690.9270.7100.217DeltaExtraTrees0.947 (0.904–0.988)0.806 (0.691–0.919)0.1410.8890.6820.207DeltaXGBoost1.000 (1.000–1.000)0.791 (0.671–0.911)0.2091.0000.7550.245DeltaLightGBM0.947 (0.904–0.988)0.765 (0.637–0.891)0.1820.8640.7170.147DeltaAdaBoost0.991 (0.979-1.000)0.766 (0.641–0.891)0.2250.9330.7240.209DeltaMLP0.961 (0.926–0.996)0.884 (0.799–0.967)0.0770.9090.8250.084 Comprehensive performance metrics of the radiomics models in training and test cohorts. Evaluation and Inter-Model Comparison of Radiomics Signatures Across Cohorts. (A–C) Receiver operating characteristic (ROC) curves of the models in the training sets. (D–F) Receiver operating characteristic (ROC) curves of the models in the test sets. Clinical models result Clinical models were evaluated across multiple machine learning algorithms using AUC and accuracy metrics. Test set performance was moderate, with AUCs ranging from ~ 0.50 to 0.68, and Random Forest achieving the highest AUC (0.681). Training accuracy exceeded test accuracy across models, indicating some overfitting. Overall, clinical variables alone showed limited predictive robustness. Detailed results are presented in Fig. 5. Comparison of model performance. (A) Receiver operating characteristic (ROC) curves of different machine learning models. (B) Accuracy of each model on the training and test sets. Signature comparison and nomogram models In the comparison of different signatures, the Delta Radiomics and Combined models outperformed the Clinic, baseline radiomics, and follow-up radiomics signatures across most cohorts. In the training cohort, the Combined model achieved a high AUC of 0.997, followed by DeltaRad (0.961), PostRad (0.913), PreRad (0.909), and Clinic (0.911). However, because the training AUC may overestimate model performance in small high-dimensional datasets, the internal validation and independent external test cohorts were considered more clinically informative. In the validation cohort, DeltaRad demonstrated the highest predictive performance with an AUC of 0.938, while the Combined model achieved 0.885. In the independent external test cohort, DeltaRad showed an AUC of 0.855, and the Combined model achieved the highest AUC of 0.875. Overall, models incorporating Delta features or integrating clinical and radiomics data provided improved accuracy, sensitivity, specificity, and predictive values compared with clinical or single-timepoint radiomics alone. Detailed performance metrics are summarized in Table 4, and ROC curves, DCA, and DeLong test results are presented in Fig. 6. Table 4Performance metrics of clinic, pre radiomics, post radiomics, delta radiomics, and combined models.SignatureAUC (95% CI)AccuracySensitivitySpecificityPPVNPVCohortClinic0.911(0.856–0.965)0.8130.9490.7120.7120.949trainPreRad0.909(0.848–0.968)0.8570.8970.8270.7950.915trainPostRad0.913(0.859–0.967)0.8240.8970.7690.7450.909trainDeltaRad0.961(0.925–0.996)0.9230.8970.9420.9210.925trainCombined0.997(0.991-1.000)0.9780.9740.9810.9740.981trainClinic0.577(0.321–0.833)0.6520.8000.5380.5710.778valPreRad0.877(0.722-1.000)0.8260.8000.8460.8000.846valPostRad0.731(0.516–0.945)0.6961.0000.4620.5881.000valDeltaRad0.938(0.842-1.000)0.9130.8001.0001.0000.867valCombined0.885(0.739-1.000)0.8700.8000.9230.8890.857valClinic0.722(0.536–0.906)0.7190.7060.7330.7500.687testPreRad0.804(0.649–0.958)0.7500.5880.9330.9090.667testPostRad0.688(0.501–0.875)0.6560.3531.0001.0000.577testDeltaRad0.855(0.726–0.983)0.7810.6470.9330.9170.700testCombined0.875(0.754–0.994)0.8120.8820.7330.7890.846test Performance metrics of clinic, pre radiomics, post radiomics, delta radiomics, and combined models. Evaluation and Inter-Model Comparison of Different Signatures Across Cohorts. (A–C) Receiver operating characteristic (ROC) curves of the Clinic, Pre Radiomics, Post Radiomics, Delta Radiomics, and Combined Models in the training, validation, and testing sets. (D–F) The decision curve analysis (DCA) for both training and testing sets. (G–I) DeLong’s test results (p-values) for pairwise model comparison in the respective cohorts. A nomogram was developed to predict severe/critical pneumonia in patients with Omicron infection. This tool integrates key clinical and radiomics predictors, including Apolipoprotein A1 (ApoA1), clinical factors summarized under “Clinic” (e.g., age, comorbidities), and the Delta Radiomics (DeltaRad) signature capturing changes between baseline and follow-up CT scans. Each predictor is assigned points, and the sum generates a Total Points score, which maps to the Risk axis indicating the probability of severe/critical disease. Higher DeltaRad values or adverse clinical features increase total points, while protective factors such as higher ApoA1 reduce risk. This nomogram enables individualized risk stratification, early identification of high-risk patients, and integrated decision-making using both imaging and clinical data. Detailed information is shown in Fig. 7. Nomogram for severity prediction in Omicron pneumonia. Discussion In this multicenter study, we demonstrated that combining Delta Radiomics features from sequential chest CT scans with clinical biomarkers may improve early prediction of severe/critical pneumonia in Omicron COVID-19 patients. Models using Delta features generally outperformed baseline-only, follow-up-only, and clinical-only models across the training, internal validation, and independent external test cohorts, highlighting the added value of temporal imaging information. Importantly, given the relatively small sample size and high-dimensional feature space, we interpreted the near-perfect training performance cautiously and placed greater emphasis on validation and external test performance. Several clinical biomarkers have been reported to be associated with COVID-19 severity. Apolipoprotein A1 (ApoA1), Apolipoprotein E (ApoE), beta-2 microglobulin (B2M), C-reactive protein (CRP), and interleukin-6 (IL-6) are commonly linked to inflammatory response and immune dysregulation during infection26–29. In the present study, ApoA1 emerged as the only significant clinical predictor in univariate logistic regression. ApoA1 is a major component of high-density lipoprotein and may reflect anti-inflammatory, antioxidant, endothelial-protective, and immunomodulatory capacity. Lower ApoA1 levels may indicate poorer metabolic and inflammatory reserve in elderly patients with Omicron infection. By contrast, conventional inflammatory markers such as CRP, IL-6, and SAA may be strongly affected by sampling time, acute-phase fluctuation, and treatment heterogeneity, which may explain why they did not reach statistical significance in this relatively small cohort. Therefore, the association between ApoA1 and severe/critical Omicron pneumonia should be regarded as exploratory and requires further validation. Although these markers are useful for identifying high-risk patients, reliance on single biomarkers remains limited. Most studies report associations rather than causal mechanisms, and heterogeneity across patient populations may lead to inconsistent findings. Moreover, individual biomarkers cannot fully capture the complex and dynamic processes underlying disease progression. Artificial intelligence–based approaches have increasingly been used to predict COVID-19 severity. Many models rely on chest imaging, particularly CT, often combined with radiomics features, and have shown promising predictive performance30–35. However, most radiomics studies analyze images acquired at a single time point. Because pulmonary lesions in COVID-19 evolve rapidly in distribution, density, and extent, single-time-point features may not adequately reflect the temporal dynamics of disease progression.The nomogram derived from the combined model offers a practical tool for clinicians, facilitating individualized risk stratification and supporting early intervention. This is particularly relevant in resource-limited settings, where timely identification of high-risk patients can optimize monitoring and treatment. Delta radiomics addresses this limitation by quantifying changes in imaging features between sequential scans. By capturing temporal variations in lesion characteristics, it provides additional information on disease evolution. Nevertheless, studies applying Delta radiomics to predict COVID-19 severity remain limited. In this multicenter study, we developed a model integrating Delta radiomics features from serial chest CT scans with clinical biomarkers. The combined model consistently outperformed models based on pre-treatment radiomics, post-treatment radiomics, or clinical variables alone across training, validation, and external test cohorts. These results suggest that temporal imaging information adds significant value in assessing disease progression. To facilitate clinical application, a nomogram based on the combined model was constructed for individualized risk estimation. By integrating imaging-derived temporal features with key clinical indicators, the nomogram provides a practical tool for early identification of patients at risk of severe pneumonia. Such risk stratification may support timely monitoring and intervention, particularly in settings with limited medical resources. The concept of CT-based AI triage has also been supported by previous work such as the study by Xu et al., whereas our study further focused on short-interval delta CT radiomics and its integration with laboratory and immune biomarkers in Omicron-infected patients36. To improve the interpretability of the Delta radiomics signature, we further examined the LASSO-selected features shown in Fig. 3C. Features with larger absolute coefficients contributed more strongly to the radiomics score. The selected features mainly represented first-order intensity statistics, texture heterogeneity, gray-level zone distribution, wavelet-transformed texture patterns, and fractal-related complexity. Positive coefficients were associated with a higher predicted probability of severe/critical Omicron pneumonia, whereas negative coefficients were associated with a lower predicted risk. Clinically, these feature categories may reflect meaningful short-interval CT changes during inflammatory progression. First-order features may indicate changes in lesion attenuation related to inflammatory exudation, ground-glass opacity, or consolidation. Texture and gray-level zone features may reflect increasing spatial heterogeneity and redistribution of pneumonia-related abnormalities. Wavelet and fractal features may capture subtle density changes and irregular lesion complexity beyond visual assessment. These findings suggest that Delta radiomics may provide quantitative imaging surrogates of inflammatory progression, although the biological meaning of individual features requires further validation. Despite the strengths of our study, several limitations should be considered. First, although external validation was performed, the external cohort remained relatively small. Second, the extraction of high-dimensional radiomics features from a limited sample size may increase the risk of overfitting, even though reproducibility filtering, stepwise dimensionality reduction, LASSO-based feature selection, and independent external testing were used. Third, the requirement for serial CT scans within a defined time window limited the number of eligible Delta radiomics cases and may introduce selection bias. Fourth, although AUCs were reported with 95% confidence intervals, larger prospective multicenter cohorts and standardized imaging protocols are needed to further confirm the robustness and generalizability of the model. In addition, only a limited number of clinical biomarkers were included. Other factors, such as treatment heterogeneity, viral load, dynamic oxygenation parameters, and additional inflammatory markers, were not fully incorporated and may improve future models. The requirement for two CT examinations within a defined time window is also an important limitation. Although the 5–9-day interval was selected to capture early and measurable pulmonary changes in symptomatic hospitalized patients, this criterion reduced the number of eligible patients and may have introduced selection bias. Therefore, the Delta radiomics framework should not be interpreted as recommending routine repeated CT for all patients with Omicron infection. It may be more applicable to higher-risk hospitalized patients who already require follow-up imaging because of worsening respiratory symptoms, oxygen desaturation, elevated inflammatory markers, comorbidities, or discordance between imaging findings and clinical status. In settings with limited CT resources, infection-control constraints, or radiation concerns, serial CT-based triage should be applied selectively. Future prospective studies should evaluate low-dose CT protocols, more flexible imaging intervals, and integration with dynamic clinical variables to improve clinical practicality. In summary, combining Delta radiomics features from serial chest CT scans with clinical biomarkers provides a non-invasive approach for early prediction of severe/critical pneumonia in Omicron COVID-19 patients. This strategy enables improved risk stratification and may assist clinicians in identifying high-risk patients at an early stage, but the findings should be considered preliminary until validated in larger prospective cohorts. Conclusion We developed and externally validated a multimodal model for early intelligent triage of Omicron pneumonia, integrating Delta Radiomics features derived from sequential chest CT scans with key clinical and laboratory biomarkers. The model predicts the binary outcome of non-severe versus severe/critical Omicron pneumonia, enabling individualized risk stratification of symptomatic patients in the early stages of disease progression. Supported by a clinically applicable nomogram, the combined model outperformed clinical-only and single-timepoint radiomics models in validation and external testing. Further large-scale prospective multicenter studies are needed to confirm its robustness and clinical utility. Abbreviations 95% confidence interval Adaptive Boosting Artificial intelligence Apolipoprotein A1 Acute Respiratory Distress Syndrome Area Under the Curve Computed tomography Coronavirus disease 2019 Decision curve analysis Temporal changes of radiomics features between serial images Digital Imaging and Communications in Medicine Extremely Randomized Trees Gray Level Co-occurrence Matrix Gray Level Run Length Matrix Gray Level Size Zone Matrix K-Nearest Neighbors Least Absolute Shrinkage and Selection Operator Light Gradient Boosting Machine Logarithmic Odds Ratio Logistic regression Multilayer Perceptron Mean squared error Naive Bayes Classifier Odds ratio Probability value, used in hypothesis testing for statistical significance Random Forest Receiver operating characteristic Region of Interest Support Vector Machine Extreme Gradient Boosting Author contributions All authors have made substantial contributions to this work, as detailed below: XY#, XD#, SG#: Conceptualization, Formal Analysis, Writing – original draft. YZ: Data curation, Formal Analysis, Writing – review & editing. JR: Data curation, Formal Analysis, Project administration, Writing – review & editing. BS: Data curation, Formal Analysis, Investigation, Writing – review & editing. JY: Funding acquisition, Project administration, Supervision, Writing – review & editing. XG: Funding acquisition, Data curation, Validation, Writing – review & editing, Formal Analysis. TZ: Conceptualization, Supervision, Writing – review & editing. CS: Project administration, Supervision, Writing – review & editing. YC*: Lead corresponding author, Funding acquisition, Methodology, Study design, Supervision, Writing – review & editing. Funding This work was supported by the Zhejiang Medical and Health Science and Technology Project (2024KY328) and the Ningbo Public Welfare Science and Technology Program Project (2023S046) and the National Clinical Key Specialty Construction Project (2023-GJZK-001). Data availability The data supporting this study’s findings are not publicly available due to privacy and confidentiality restrictions. However, the datasets can be obtained from the corresponding author upon reasonable request, subject to approval and compliance with relevant data protection regulations. Code availability No new proprietary algorithm, software, or bespoke computational tool was developed in this study. Radiomics feature extraction was performed using the open-source PyRadiomics package, and machine-learning analyses were conducted using commonly used open-source algorithms and packages. The main analysis workflow, feature selection procedures, model types, and parameter settings are described in the Methods section. Additional implementation details are available from the corresponding author upon reasonable request. Because of institutional and ethical restrictions, raw CT images and identifiable clinical data are not publicly available. Declarations Competing interests The authors declare no competing interests. Ethics approval This study was approved by the Ethics Committee of the First Affiliated Hospital of Ningbo University (Approval No. 2026003RS) and the Ethics Committee of the Affiliated Hospital of Jiaxing University (Approval No. 2026-LP-019), and was conducted in accordance with local and institutional requirements. Written informed consent for participation and publication was waived by the Ethics Committee of the First Affiliated Hospital of Ningbo University and the Ethics Committee of the Affiliated Hospital of Jiaxing University because this was a retrospective study using anonymized paired chest CT scans and laboratory data collected during routine clinical care. No data explicitly declined by participants were used. All data were strictly anonymized and were used solely for the development and validation of an AI prognostic model for Omicron COVID-19. This exemption did not adversely affect the rights or interests of the participants. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xiaoxian Ye, Xuhao Dai and Shengping Gong have contributed equally to this work. Contributor Information Jiangping Ren, Email: fyyrenjiangping@nbu.edu.cn. Cong Shi, Email: shicong5103@sina.com. Yijian Cao, Email: caoyj22@tsinghua.org.cn. References Associated Data Data Availability Statement The data supporting this study’s findings are not publicly available due to privacy and confidentiality restrictions. However, the datasets can be obtained from the corresponding author upon reasonable request, subject to approval and compliance with relevant data protection regulations. No new proprietary algorithm, software, or bespoke computational tool was developed in this study. Radiomics feature extraction was performed using the open-source PyRadiomics package, and machine-learning analyses were conducted using commonly used open-source algorithms and packages. The main analysis workflow, feature selection procedures, model types, and parameter settings are described in the Methods section. Additional implementation details are available from the corresponding author upon reasonable request. Because of institutional and ethical restrictions, raw CT images and identifiable clinical data are not publicly available.