Development and internal validation of a CT radiomics-based model for severity classification in HIV-associated Pneumocystis jirovecii pneumonia Authors: Yu Wang, Wei Wang, Caopei Zheng, Yuqing Sun, Budong Chen, Yulin Zhang Published: 2026 ABSTRACT AbstractBackgroundSeverity assessment of human immunodeficiency virus (HIV)-associated Pneumocystis jirovecii pneumonia (PJP) is clinically important because arterial blood gas indices are standard but may not fully capture the heterogeneity of lung involvement on chest computed tomography (CT). We aimed to develop and internally validate a CT radiomics model for classifying severity in adults with confirmed HIV-associated PJP.MethodsThis retrospective single-center study included 96 adult patients with confirmed HIV-associated PJP who underwent chest CT at presentation. Disease severity was classified as mild or moderate-to-severe according to room-air arterial blood gas criteria, with moderate-to-severe disease defined as arterial partial pressure of oxygen (PaO2) <70 mmHg or an alveolar-arterial oxygen gradient (A-aDO2) ≥35 mmHg. Clinical variables were retrospectively collected from medical records. Patients were randomly divided into training and test cohorts at a ratio of 7:3. Radiomics features were extracted from the bilateral lung parenchyma. After least absolute shrinkage and selection operator regression, features with non-zero coefficients were included in the final radiomics model. In parallel, a clinical logistic model incorporating serum lactate dehydrogenase, β-D-glucan, and CD4 count was developed in the training cohort and tested in the test cohort. Model performance was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals (CIs), calibration analysis, decision curve analysis, and pairwise DeLong test comparisons.ResultsOf the 96 patients, 38 were classified as mild and 58 as moderate-to-severe. Patients with moderate-to-severe disease had a higher frequency of dyspnea, higher levels of inflammatory markers, and lower CD4 count. The final radiomics model included 10 features. In the training cohort, the radiomics model achieved an AUC of 0.92 (95% CI: 0.85–0.97), compared with 0.65 (95% CI: 0.51–0.78) for the clinical logistic model. In the test cohort, the radiomics model showed a numerically higher AUC of 0.89 (95% CI: 0.72–1.00), followed by the clinical logistic model at 0.84 (95% CI: 0.68–0.97). Using their respective classification thresholds, the radiomics model yielded a sensitivity of 0.778 (95% CI: 0.548–0.910) and a specificity of 0.818 (95% CI: 0.523–0.949) in the test cohort, while the clinical logistic model yielded a sensitivity of 0.722 (95% CI: 0.491–0.875) and a specificity of 0.909 (95% CI: 0.623–0.984). Pairwise DeLong tests in the test cohort showed no statistically significant difference between the radiomics model and the clinical logistic model.ConclusionsIn this small single-center study, the CT radiomics model showed promising discrimination for severity classification in HIV-associated PJP, but these findings are preliminary and require external multicenter validation before clinical use.Keywords: Human immunodeficiency virus (HIV), Pneumocystis jirovecii pneumonia (PJP), chest computed tomography (chest CT), radiomics, machine learning Highlight box Key findings • In this study, a chest computed tomography (CT)-based radiomics model combined with machine learning was developed to stratify disease severity in patients with human immunodeficiency virus (HIV)-associated Pneumocystis jirovecii pneumonia (PJP). The model demonstrated good discriminative performance in both the training and independent test cohorts and showed numerically better discriminative performance than models based on clinical variables alone. These findings indicate that quantitative imaging features can capture severity-related heterogeneity that is not fully reflected by routine clinical indicators. What is known and what is new? • Previous studies have shown that severity assessment of HIV-associated PJP mainly relies on clinical symptoms, laboratory indicators, and visual interpretation of chest CT images. However, these approaches are largely developed for community-acquired pneumonia and often perform poorly in immunocompromised patients, limiting their ability to accurately identify moderate-to-severe PJP at presentation. • This study adds evidence that quantitative radiomics features extracted from chest CT images can capture severity-related heterogeneity of pulmonary involvement in HIV-associated PJP. By integrating these features with machine learning, our model provides additional discriminatory value beyond conventional clinical variables for identifying patients with moderate-to-severe disease. What is the implication, and what should change now? • The results suggest that radiomics-based analysis of chest CT images may serve as a useful adjunct to conventional clinical assessment for identifying patients with moderate-to-severe PJP. Incorporation of such quantitative imaging approaches into clinical workflows may help support timely severity assessment and more informed treatment decisions in HIV-infected patients. Introduction Background Pneumocystis jirovecii pneumonia (PJP) is a severe and potentially fatal pneumonia caused by Pneumocystis jirovecii, mainly affecting immunocompromised individuals (1). In the era prior to widespread use of antiretroviral therapy, approximately 70–80% of individuals infected with human immunodeficiency virus (HIV) developed PJP during the course of their disease, with reported mortality rates ranging from 20% to 40%. Moreover, the recurrence rate approached 50% within 11 months after an initial episode (2). With the broad implementation of measures suggested by the Centers for Disease Control and Prevention (CDC), PJP-related mortality has declined. Nevertheless, PJP remains highly prevalent in patients with newly diagnosed HIV infection, delayed diagnosis, or profound immunosuppression, and continues to be associated with a substantial mortality rate of approximately 14.6% (3,4). In clinical practice, therapeutic strategies differ markedly between patients with mild PJP and those with moderate-to-severe disease. Patients with moderate-to-severe PJP often require more aggressive management, including adjunctive systemic corticosteroids, intensified oxygen therapy, and close clinical monitoring, whereas mild cases are typically managed with standard antimicrobial therapy alone. Moderate-to-severe PJP is currently assessed mainly on the basis of arterial blood gas (ABG) analysis, particularly an arterial partial pressure of oxygen (PaO2) <70 mmHg or an alveolar-arterial oxygen gradient (A-aDO2) ≥35 mmHg, as these indices directly reflect the degree of gas exchange impairment and serve as important criteria for clinical stratification and the initiation of adjunctive corticosteroid therapy (5). However, ABG provides a physiological measure of gas-exchange impairment rather than a direct imaging characterization of the spatial extent, distribution pattern, and tissue heterogeneity of pulmonary lesions. Previous chest computed tomography (CT) radiomics studies in pneumonia have shown that quantitative imaging features can capture whole-lung lesion burden and lesion heterogeneity beyond visual assessment alone, and are associated with oxygenation impairment and disease severity (6). Therefore, the value of CT radiomics in HIV-associated PJP does not lie in replacing ABG as the physiological reference standard, but in providing quantitative descriptors of whole-lung disease burden and imaging heterogeneity that may complement ABG-based severity classification at presentation. Rationale and knowledge gap Currently, commonly used pneumonia severity assessment tools, such as the Pneumonia Severity Index and the Confusion, Urea, Respiratory rate, Blood pressure, age ≥65 years (CURB-65) score, are primarily based on inflammatory markers, vital signs, and organ function parameters, and are originally developed for patients with community-acquired pneumonia (7). However, HIV-infected individuals exhibit distinct characteristics in terms of immune suppression, inflammatory response patterns, and pathogen spectrum, which substantially limit the accuracy of these scoring systems for risk stratification of moderate-to-severe disease in patients with PJP (8). At present, clinical severity stratification of PJP still relies primarily on ABG analysis, particularly PaO2 and the A-aDO2, which serve as the reference standard for identifying moderate-to-severe disease at presentation. By contrast, commonly reported laboratory variables such as lactate dehydrogenase (LDH), CD4 count, and β-D-glucan (BDG) reflect different dimensions of disease biology, including host immune status, nonspecific tissue injury, fungal burden, or risk of adverse outcomes, but do not provide a stable framework for classifying the current severity of PJP at presentation. Consequently, although these indicators may have adjunctive clinical value, they do not resolve the need for a quantitative approach to severity classification (2,9-11). To our knowledge, there is still a lack of CT-based quantitative imaging models specifically developed for HIV-associated PJP that use admission imaging features to classify ABG-defined disease severity at presentation. In recent years, radiomics has emerged as a novel analytical approach that enables high-throughput extraction of large numbers of quantitative features from medical images. Radiomics has demonstrated considerable potential in imaging-based severity characterization of infectious lung diseases (12-14). By extracting features that reflect lesion extent, attenuation changes, texture heterogeneity, and pulmonary lesion burden from CT images, radiomics captures imaging information that often exceeds visual assessment by radiologists. This allows for a more refined characterization of pneumonia distribution, lesion heterogeneity, and imaging patterns of pulmonary involvement (15,16). In PJP, patients presenting with moderate-to-severe PJP typically exhibit more extensive ground-glass opacities, more pronounced ventilation-perfusion impairment, and more complex lesion texture patterns. Chest CT-based radiomics is well-suited to capture these imaging phenotypes associated with disease severity (17). Objective Accordingly, this study aimed to develop and internally validate a CT radiomics-based model for classifying ABG-defined mild versus moderate-to-severe HIV-associated PJP at presentation, and to compare its performance with clinical biomarker-based models. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0220/rc). Methods Study population The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Youan Hospital (No. LL-2024-172-K) and individual consent for this retrospective analysis was waived. Patients with HIV-associated PJP who were treated at Beijing Youan Hospital between January 2020 and August 2025 and had available chest CT images were screened for inclusion. The inclusion criteria were as follows: (I) age ≥18 years; (II) confirmed HIV infection with PJP meeting the diagnostic criteria for an acquired immunodeficiency syndrome (AIDS)-defining illness according to the CDC guidelines (18) defined as a CD4 count <200 cells/µL or the presence of at least one AIDS-related condition; and (III) diagnosis of PJP based on the criteria proposed by the Panel on Opportunistic Infections in Adults and Adolescents of the U.S. Department of Health and Human Services (5) including immunodeficiency (most commonly CD4 count <200 cells/µL), subacute onset of progressive dyspnea, fever, nonproductive cough, and chest discomfort with symptom progression over days to weeks, hypoxemia, elevated serum BDG and LDH levels, bilateral symmetric interstitial infiltrates with a perihilar “bat-wing” distribution on chest imaging (which may be normal in early stages), detection of Pneumocystis jirovecii organisms in lung tissue, bronchoalveolar lavage fluid (BALF), or induced sputum, or amplification of Pneumocystis jirovecii DNA in BALF or induced sputum using polymerase chain reaction. Disease severity was assessed using the A-aDO2 or PaO2 under room air conditions. According to the definition of moderate-to-severe PJP proposed by the U.S. Department of Health and Human Services expert panel, patients with PaO2 <70 mmHg or A-aDO2 ≥35 mmHg were classified as having moderate-to-severe disease, while the remaining patients were classified as having mild disease. Exclusion criteria included age <18 years, foreign nationality, pregnancy or lactation, and the presence of concomitant pulmonary infections caused by other pathogens. Clinical data collection A retrospective case-review approach was used to collect clinical data, including demographic characteristics (sex and age), underlying diseases and medical history, clinical manifestations (fever, cough, and dyspnea), and major laboratory parameters obtained at admission and during hospitalization. Laboratory variables included white blood cell count, neutrophil count, lymphocyte count, hemoglobin, platelet count, C-reactive protein (CRP), LDH, alanine aminotransferase (ALT), aspartate aminotransferase (AST), albumin, total bilirubin, blood urea, serum creatinine, plasma BDG levels, and CD4 count. ABG parameters and clinical outcomes, including length of hospital stay and in-hospital mortality, were also recorded. CT image acquisition Chest CT examinations were performed using a 64-slice CT scanner LightSpeed VCT (GE Healthcare, Waukesha, WI, USA). All patients were scanned in the supine position. The scanning parameters included a tube voltage of 120 kV with automatic tube current modulation, a matrix size of 512×512, a slice thickness of 5.0 mm, and a slice interval of 5.0 mm. The acquired CT images were transferred to a post-processing workstation for subsequent image processing and analysis. Image segmentation and radiomics feature extraction Chest CT images were obtained from patients with confirmed HIV-associated PJP. Image segmentation was performed using the 3D Slicer software. During segmentation, CT images were displayed using a fixed lung window setting [window level −500 Hounsfield units (HU), window width 1,500 HU] to facilitate identification of lung parenchymal lesion boundaries. The segmentation scope was restricted to the bilateral lung parenchyma, focusing on areas of parenchymal involvement while excluding the heart, large vessels, and other non-pulmonary structures. Regions of interest (ROIs) were delineated using a semi-automatic gray-level threshold-based approach by clinicians experienced in thoracic imaging, with manual correction applied when necessary to ensure consistency and accuracy of segmentation. After ROI delineation, all images were resampled to an isotropic voxel resolution of 1×1×1 mm3. Radiomics feature extraction was subsequently performed within the ROIs based on original CT attenuation values using the PyRadiomics package (version 3.1.0) implemented in Python. A total of 1,302 radiomics features were extracted from each ROI, including first-order statistical features and texture features derived from the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix, and gray-level dependence matrix. Higher-order texture features obtained after Laplacian of Gaussian filtering and wavelet transformation were also extracted. Given that PJP is characterized by diffuse parenchymal involvement, shape features, which are highly sensitive to segmentation boundaries and have limited cross-case comparability, were not included in the analysis. Model construction All samples were randomly divided into training and test cohorts at a ratio of 7:3 using stratified random sampling to ensure comparable proportions of moderate-to-severe and mild cases in both cohorts. Feature normalization, feature selection, and model training were performed exclusively in the training cohort. Radiomics features were standardized using Z-score normalization in the training cohort, and the same parameters were applied to the test cohort. Feature selection was conducted using least absolute shrinkage and selection operator (LASSO) regression with cross-validation in the training cohort. Five-fold stratified cross-validation was used to evaluate model error across different regularization parameters (λ). The final λ value was selected by considering both minimal cross-validation error and feature number constraints, retaining radiomics features with non-zero coefficients. The selected radiomics features were used to build an elastic-net logistic regression classifier. Model hyperparameters were optimized using repeated stratified five-fold cross-validation with random search. Model performance in the training set was assessed using out-of-fold (OOF) predicted probabilities generated from an outer stratified five-fold cross-validation procedure. After hyperparameter tuning, the final model was fitted on the full training set and then applied to the independent test set. The clinical model included three variables: serum LDH, BDG, and CD4 count, and was constructed using logistic regression. Continuous variables were standardized before analysis. Performance in the training set was evaluated using OOF predicted probabilities obtained from stratified five-fold cross-validation, and performance in the test set was evaluated using predictions from the independent test set. Model performance was evaluated separately in the training and test cohorts. Discriminative ability was quantified using the area under the receiver operating characteristic curve (AUC). The AUC in the training set was calculated from OOF predicted probabilities, while the AUC in the test set was calculated from predicted probabilities in the independent test set. The 95% confidence intervals (CIs) for the AUC were estimated using 1,000 stratified bootstrap resamples. The discriminative performance of the radiomics model and the clinical logistic regression model was compared using the AUC and its 95% CI, and the difference in AUC between the two models in the test set was further assessed with the DeLong test. For both the radiomics model and the clinical model, the classification threshold was determined from the OOF predicted probabilities in the training set by maximizing the Youden index (sensitivity + specificity − 1), and the same threshold was then applied to the test set. Confusion matrices were constructed accordingly, and sensitivity, specificity, positive predictive value, and negative predictive value were calculated. Model calibration was evaluated using calibration curves. In the training set, bootstrap correction was used to adjust for optimism in the calibration results, and agreement between predicted probabilities and observed outcomes was further assessed in the test set. Decision curve analysis (DCA) was also performed to evaluate the potential clinical net benefit of the model. Given the limited sample size and the high dimensionality of the initial radiomics feature set, LASSO feature selection, repeated cross-validation, bootstrap-based CI estimation, and calibration analysis were used to reduce the risk of overfitting. Results from the test set were reported as internal validation findings. Statistical analysis Statistical analyses were primarily performed using Python version 3.9 and IBM SPSS Statistics version 26.0, which were used for radiomics modeling, model score visualization, and clinical variable comparisons, respectively. Radiomics analyses and data visualization were conducted in the Python environment using key libraries including NumPy, Pandas, SciPy, and scikit-learn. Normality of continuous variables was assessed using the Shapiro-Wilk test. Normally distributed variables are presented as mean ± standard deviation and were compared using the independent-samples t-test. Non-normally distributed variables are presented as median (interquartile range) and were compared using the Mann-Whitney U test. Categorical variables are presented as counts and percentages and were compared using the chi-square test or Fisher’s exact test, as appropriate. In addition, the predicted scores/probabilities of the radiomics model and the clinical logistic model in the test set were visualized using boxplots and compared between the two outcome groups using the Mann-Whitney U test. A two-sided P value <0.05 was considered statistically significant. Results Patient characteristics A total of 96 patients with PJP were included in this study, comprising 38 patients in the mild group and 58 patients in the moderate-to-severe group. There were no significant differences between the two groups with respect to age or sex distribution. Major comorbidities, including cardiovascular and cerebrovascular diseases, diabetes mellitus, and liver disease, were similarly distributed between the two groups, with no statistically significant differences observed. Regarding clinical manifestations, the incidence of dyspnea was significantly higher in the moderate-to-severe group than in the mild group (P=0.02), whereas the frequencies of fever and cough did not differ significantly between the groups. Laboratory findings indicated more pronounced inflammatory responses and tissue injury in patients with moderate-to-severe disease. Compared with the mild group, patients in the moderate-to-severe group exhibited significantly higher white blood cell counts (P=0.003), neutrophil counts (P<0.001), platelet counts (P=0.02), CRP levels (P<0.001), AST levels (P=0.046), blood urea levels (P=0.01), and LDH levels (P=0.003). In contrast, serum albumin levels (P<0.001), serum creatinine levels (P=0.03), and CD4 counts (P=0.001) were significantly lower in the moderate-to-severe group. No significant differences were observed in the remaining laboratory parameters. With respect to in-hospital outcomes, patients in the moderate-to-severe group had a significantly longer length of hospital stay than those in the mild group (P=0.004), whereas in-hospital mortality did not differ significantly between the two groups (Table 1). Table 1Comparison of baseline characteristics, laboratory findings, and clinical outcomes between mild and moderate-to-severe HIV-associated PJP patientsCharacteristicsMild PJP (n=38)Moderate-to-severe PJP (n=58)P valueAge, years, median (IQR)36.50 (32.00–52.25)41.00 (32.00–49.25)0.66Sex, male, n (%)37 (97.37)55 (94.83)>0.99Comorbidities, n (%) Cardio-cerebrovascular disease, n (%)6 (15.79)7 (12.07)0.60 Diabetes mellitus, n (%)5 (13.16)9 (15.52)0.75 Liver disease, n (%)3 (7.89)8 (13.79)0.52Clinical symptoms Fever, n (%)26 (68.42)39 (67.24)0.90 Cough, n (%)23 (60.53)35 (60.34)0.99 Dyspnea, n (%)26 (68.42)51 (87.93)0.02Hematology White blood cell count, ×109/L, median (IQR)5.01 (3.55–6.45)6.57 (4.70–8.96)0.003 Neutrophil count, ×109/L, median (IQR)3.56 (2.11–4.34)5.00 (3.10–7.22)<0.001 Lymphocyte count, ×109/L, median (IQR)0.84 (0.53–1.40)0.81 (0.52–1.10)0.39 Hemoglobin, g/L, median (IQR)122.50 (115.25–134.50)128.00 (106.75–134.00)0.95 Platelet count, ×109/L, median (IQR)202.00 (173.50–287.75)263.00 (203.75–340.00)0.02Renal and liver function ALT, U/L, median (IQR)27.00 (19.75–55.75)24.50 (16.75–39.25)0.14 AST, U/L, median (IQR)32.50 (21.75–42.50)39.00 (28.50–50.25)0.046 Albumin, g/L, mean ± SD37.61±5.0730.41±7.24<0.001 Total bilirubin, μmol/L, median (IQR)8.40 (6.60–10.35)8.55 (6.38–11.33)0.89 Urea, mmol/L, median (IQR)4.69 (3.37–5.44)5.35 (4.43–7.24)0.01 Serum creatinine, μmol/L, median (IQR)72.50 (62.00–80.25)62.00 (55.00–78.00)0.03Inflammation and tissue injury CRP, mg/L, median (IQR)8.22 (3.65–16.98)31.44 (8.08–53.54)<0.001 LDH, U/L, mean ± SD253.40±56.61392.95±181.400.003Immune status and fungal biomarkers β-1,3-D-glucan, pg/mL, median (IQR)137.40 (48.60–231.00)132.95 (10.00–224.05)0.59 CD4 count, cells/μL, median (IQR)46.00 (26.75–94.75)24.00 (12.50–40.75)0.001Length of stay, days, median (IQR)11.00 (7.00–17.00)14.50 (11.00–18.25)0.004Mortality, n (%)0 (0.00)1 (1.72)>0.99P values were calculated using the independent-samples t-test for normally distributed continuous variables, the Mann-Whitney U test for non-normally distributed continuous variables, and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. All tests were two-sided. ALT, alanine aminotransferase; AST, aspartate aminotransferase; CRP, C-reactive protein; HIV, human immunodeficiency virus; IQR, interquartile range; LDH, lactate dehydrogenase; PJP, Pneumocystis jirovecii pneumonia; SD, standard deviation. Comparison of baseline characteristics, laboratory findings, and clinical outcomes between mild and moderate-to-severe HIV-associated PJP patients P values were calculated using the independent-samples t-test for normally distributed continuous variables, the Mann-Whitney U test for non-normally distributed continuous variables, and the chi-square test or Fisher’s exact test for categorical variables, as appropriate. All tests were two-sided. ALT, alanine aminotransferase; AST, aspartate aminotransferase; CRP, C-reactive protein; HIV, human immunodeficiency virus; IQR, interquartile range; LDH, lactate dehydrogenase; PJP, Pneumocystis jirovecii pneumonia; SD, standard deviation. Radiomics feature extraction and selection In terms of imaging manifestations, patients with moderate-to-severe PJP exhibited a markedly greater extent of lung parenchymal involvement than those with mild disease, characterized by more extensive bilateral diffuse ground-glass opacities and reticular changes. In contrast, mild PJP was predominantly associated with focal or relatively limited pulmonary involvement (Figure 1A,1B). Representative chest CT images and radiomics feature selection in HIV-associated PJP. Representative axial chest CT images of patients with moderate-to-severe disease (A) and mild disease (B). LASSO cross-validation and coefficient path plots illustrate radiomics feature selection (C,D). Radiomics features retained by LASSO regression and their coefficients are shown (E). CT, computed tomography; CV, cross-validation; HIV, human immunodeficiency virus; LASSO, least absolute shrinkage and selection operator; MSE, mean squared error; PJP, Pneumocystis jirovecii pneumonia; SD, standard deviation; SE, standard error. During radiomics feature selection, five-fold cross-validated LASSO regression was applied to reduce the dimensionality of high-dimensional imaging features. The cross-validation error curve demonstrated a progressive decrease in mean squared error with increasing regularization parameter (λ), reaching its minimum at approximately λ≈0.07 (Figure 1C). Based on the minimum error criterion, a λ value of 0.0688 was selected as the final regularization parameter. The LASSO coefficient path plot showed that, as λ increased, the coefficients of most features gradually shrank toward zero, whereas only a small number of features retained non-zero coefficients at higher λ values, indicating relatively stable contributions to model discrimination (Figure 1D). At the selected λ value, a total of 10 radiomics features were retained. These features were primarily derived from wavelet-transformed intensity distribution and texture characteristics, including first-order statistical features and GLCM-based features. Features with positive coefficients were associated with an increased risk of moderate-to-severe PJP, whereas features with negative coefficients were associated with lower disease severity (Figure 1E). Discriminative performance of the radiomics model The discriminative performance of the radiomics model and the clinical logistic regression model is shown in Figure 2. The radiomics model showed good discriminative performance in both the training and test sets. Its AUC was 0.92 (95% CI: 0.85–0.97) in the training set and 0.89 (95% CI: 0.72–1.00) in the test set. The sensitivity and specificity were 0.775 (95% CI: 0.625–0.877) and 0.926 (95% CI: 0.766–0.979), respectively, in the training set, and 0.778 (95% CI: 0.548–0.910) and 0.818 (95% CI: 0.523–0.949), respectively, in the test set. Comparison of discriminative performance between the radiomics model and the clinical logistic model for severity stratification in HIV-associated PJP. ROC curves of the radiomics model and the clinical logistic model in the training and test sets (A). Boxplots of predicted probabilities in the test set, illustrating the distribution of model scores in patients with mild disease (label 0) and moderate-to-severe disease (label 1) for the radiomics model and the clinical logistic model (B). Confusion matrices of the radiomics model in the training and test sets, based on the classification threshold derived from the training-set OOF predictions and then applied unchanged to the test set (C). AUC, area under the curve; CI, confidence interval; HIV, human immunodeficiency virus; OOF, out-of-fold; PJP, Pneumocystis jirovecii pneumonia; ROC, receiver operating characteristic. For the clinical logistic regression model, the AUCs were 0.65 (95% CI: 0.51–0.78) in the training set and 0.84 (95% CI: 0.68–0.97) in the test set. Its sensitivity and specificity were 0.550 (95% CI: 0.398–0.693) and 0.704 (95% CI: 0.515–0.841), respectively, in the training set, and 0.722 (95% CI: 0.491–0.875) and 0.909 (95% CI: 0.623–0.984), respectively, in the test set. However, the DeLong test showed no statistically significant difference in AUC between the radiomics model and the clinical logistic regression model in the test set (Z=0.508, P=0.61). The boxplots for the test set, however, showed a certain degree of score separation between the mild and moderate-to-severe groups for both models, with less overlap in the predicted probability distributions of the radiomics model. Calibration performance and clinical utility of the radiomics model Calibration analysis demonstrated that, in the training cohort, the apparent calibration curve of the radiomics model closely aligned with the ideal reference line. After bootstrap correction, the calibration curve further approached the ideal line, with only minor deviations observed in the high predicted probability range, indicating a low degree of overfitting and good internal calibration performance (Figure 3A). Calibration performance and clinical utility of the radiomics model for severity stratification in HIV-associated PJP. Calibration curves of the radiomics model in the training cohort (A) and the independent test cohort (B). Decision curve analysis of the radiomics model in the test cohort (C). HIV, human immunodeficiency virus; PJP, Pneumocystis jirovecii pneumonia. In the independent test cohort, predicted probabilities showed good agreement with observed outcomes, with observed event rates increasing linearly as predicted risk increased. The overall distribution closely approximated the ideal calibration line, suggesting stable classification performance of the model in the independent test cohort (Figure 3B). Decision curve analysis showed that, across a range of clinically relevant threshold probabilities (approximately 0.10–0.70), the radiomics model consistently yielded a higher net benefit than both the “treat-all” and “treat-none” strategies, and maintained an advantage at a disease prevalence of approximately 0.62, indicating potential clinical utility of the model for decision-making (Figure 3C). Discussion Key findings In this study, a chest CT-based radiomics model combined with machine learning was developed to classify disease severity in HIV-associated PJP. The results demonstrated that the model achieved good discriminative performance in both the training and independent test cohorts. Receiver operating characteristic analysis indicated that the model effectively distinguished between moderate-to-severe and mild PJP, while maintaining stable classification performance across different datasets. Clinical characteristic analyses showed that patients with moderate-to-severe PJP had significantly higher levels of inflammation- and tissue injury-related markers, including white blood cell count, neutrophil count, platelet count, CRP, AST, blood urea, and LDH, compared with patients with mild disease. In contrast, serum albumin levels, serum creatinine levels, and CD4 counts were markedly lower in the moderate-to-severe group, indicating more pronounced systemic inflammatory responses and immune dysfunction in these patients. However, the discriminative ability of individual clinical indicators for identifying severe disease was limited. Taken together, the radiomics-based classification model integrates imaging heterogeneity of pulmonary lesions and demonstrates favorable discriminative value for assessing disease severity in HIV-associated PJP at presentation, providing a potential auxiliary tool for identification of patients with moderate-to-severe disease at presentation in clinical practice. Strengths and limitations This study has several strengths. First, a CT-based radiomics framework was applied to quantitatively characterize whole-lung parenchymal involvement in HIV-associated PJP, enabling objective assessment of imaging heterogeneity beyond visual interpretation. Second, the radiomics model was developed using machine learning and validated in an independent test cohort, supporting its robustness and reproducibility for disease severity stratification. Third, the integration of imaging features with routinely available clinical information enhances the potential clinical applicability of the proposed approach. Despite these encouraging results, several limitations should be acknowledged. First, this was a single-center retrospective study with a relatively limited sample size. Second, only baseline CT images and selected routine clinical variables were included, and dynamic imaging data and longitudinal follow-up information were unavailable. As a result, the potential longitudinal value of radiomics features over the disease course could not be assessed. Comparison with similar research The incorporation of radiomics methodology improved the accuracy of disease severity classification in this study. Previous studies have shown that ground-glass opacities on chest CT are closely associated with the occurrence of PJP and represent one of the most diagnostically relevant imaging features (19). Patients with a prior history of PJP often develop restrictive ventilatory defects and impaired diffusing capacity, with more pronounced abnormalities observed in those with moderate-to-severe disease. Imaging findings in such patients frequently include persistent ground-glass opacities, bronchiectasis, and subpleural cystic changes, which may indicate chronic or irreversible lung tissue injury (20). From an immunological perspective, patients with CD4 counts below 100 cells/µL are more likely to exhibit diffuse pulmonary infiltrates, whereas pleural effusion, central lymphadenopathy, pneumothorax, and cystic lesions are relatively uncommon (21). In HIV-positive patients, PJP typically presents with diffuse bilateral ground-glass opacities, representing the classic radiologic pattern, and some patients may also show mediastinal lymphadenopathy or limited cystic changes (22,23). Explanations of findings In the present study, key radiomics features selected by LASSO regression were predominantly derived from wavelet-transformed gray-level distribution features and GLCM-based texture parameters. First-order statistical features, such as skewness, high-percentile values, and maximum attenuation, reflect asymmetry in gray-level distribution and increased high-density components within lesions, whereas GLCM-related features, including contrast, cluster tendency, and gray-level uniformity metrics, characterize spatial relationships and local intensity variations within lesions. These features jointly quantify pulmonary lesion heterogeneity from both intensity distribution and spatial texture perspectives, suggesting that patients with moderate-to-severe PJP have more extensive parenchymal involvement, greater inflammatory exudation, and more heterogeneous density distributions. Previous studies have demonstrated that GLCM-based texture features are highly sensitive to the spatial complexity of inflammatory lung lesions and can be used for severity stratification (24). The findings of the present study further support the potential value of radiomics texture features for identifying moderate-to-severe disease risk in HIV-associated PJP. Implications and actions needed The results of this study suggest that CT-based radiomics may serve as a complementary imaging tool for severity classification at presentation in HIV-associated PJP, alongside conventional clinical and laboratory assessment. By quantitatively characterizing the extent and heterogeneity of pulmonary involvement, the proposed model may provide additional information for distinguishing mild from moderate-to-severe disease at presentation. However, given the single-center retrospective design, limited sample size, and lack of external validation, the current findings should be considered preliminary and should not yet be interpreted as sufficient to support routine clinical implementation. From a research perspective, further validation in larger, multicenter cohorts is needed to confirm the robustness and generalizability of the model. Future studies incorporating external validation and, where appropriate, longitudinal imaging and clinical data may help further define the role of radiomics in disease assessment in HIV-associated PJP. Conclusions This study suggests that a chest CT-based radiomics model may provide complementary value for classifying disease severity at presentation in patients with HIV-associated PJP. Given the limited sample size and internal validation only, further external validation is needed before clinical application. Supplementary The article’s supplementary files as Acknowledgments None. Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Beijing Youan Hospital (No. LL-2024-172-K) and individual consent for this retrospective analysis was waived. Footnotes Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0220/rc Funding: This work was supported by the Beijing Research Ward Excellence Program (Nos. BRWEP2024W042180101 and BRWEP2024W042180110). Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0220/coif). Y.W., W.W., C.Z. and Y.S. report support for the present manuscript from the Beijing Research Ward Excellence Program. B.C. reports support for the p resent manuscript from the Beijing Research Ward Excellence Program (BRWEP2024W042180110). Y.Z. reports support for the present manuscript from the Beijing Research Ward Excellence Program (BRWEP2024W042180101). The authors have no other conflicts of interest to declare. Data Sharing Statement Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0220/dss References Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0220/prf Supplementary Materials The article’s supplementary files as Data Availability Statement Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0220/dss