Contrast-enhanced CT based radiomics for prediction of 3-year disease progression in NSCLC patients after anti-PD-1-targeted therapy☆ Authors: Wen Huang, Min-jia Lin, Yi-ting He, Hong Zhou, Yi Chen, Min Zong Published: 2026 ABSTRACT AbstractPurpose:This study aimed to develop and validate a radiomics-based model derived from contrast-enhanced computed tomography (CE-CT) to predict 3-year disease progression in patients with non-small-cell lung cancer (NSCLC) receiving anti-PD1 immunotherapy.MethodsA total of 173 patients with NSCLC undergoing anti-PD1 immunotherapy were retrospectively enrolled. We developed a integrated model based on Radscore by selecting radiomics features from target lesions (TL) and clinical features derived from pretreatment CE-CT images. The receiver operating characteristic (ROC) curve was used to evaluate the predictive performance of different models. Model interpretability was enhanced via Shapley Additive Explanations (SHAP).ResultsChronic obstructive pulmonary disease (COPD) and tumor stage were identified as significant clinical predictors of 3-year disease progression. The radiomics model achieved area under the curve (AUC) values of 0.758 (95% CI: 0.663–0.852) and 0.815 (95% CI: 0.618–1.000) in training and testing cohorts, respectively. The integrated model showed improved performance, with AUCs of 0.802 (95% CI: 0.721–0.884) and 0.836 (95% CI: 0.663–1.000), respectively. The nomogram exhibited superior net clinical benefit compared to radiomics- or clinical-only models. SHAP analysis identified shape Sphericity, gldm Small Dependence High Gray Level Emphasis, glszm Gray Level Variance, glszm Small Area High Gray Level Emphasis as key imaging features associated with 3-year disease progression.Conclusions:We developed an integrated clinical-radiomics model that effectively identifies NSCLC patients most likely to benefit from anti-PD-1 therapy. Using SHAP-based explainability, we clarified the contribution of imaging features, enabling more personalized treatment strategies.Keywords: Radiomics, Contrast enhanced CT, Non-small cell Lung Cancer, PD-1 targeted therapy, SHAP Key points: •COPD and tumor stage were identified as significant prognostic factors for 3-year disease progression in patients with NSCLC following anti-PD1 therapy. COPD and tumor stage were identified as significant prognostic factors for 3-year disease progression in patients with NSCLC following anti-PD1 therapy. •CECT based radiomics have good performance in predicting 3-year disease progression. CECT based radiomics have good performance in predicting 3-year disease progression. •The model integrates COPD status, tumor stage, and CECT radiomics signature, all of which can be easily obtained, can offer better prognostic information on 3-year disease progression status. The model integrates COPD status, tumor stage, and CECT radiomics signature, all of which can be easily obtained, can offer better prognostic information on 3-year disease progression status. Introduction Non-small cell lung cancer (NSCLC) is the most common type of lung cancer, accounting for about 85% of all lung cancer cases globally [1], [2]. Despite significant advancements in treatment modalities, NSCLC still poses a major challenge in oncology due to its high recurrence rate and poor prognosis [3], [4], [5]. Immunotherapy, especially the use of programmed cell death protein 1 (PD−1) inhibitors, has emerged as a revolutionary treatment approach for NSCLC. PD−1 inhibitors work by blocking the PD−1 pathway, which in turn unleashes the immune system to recognize and attack cancer cells [6]. This treatment has shown remarkable efficacy in a subset of NSCLC patients, leading to improved survival and better quality of life. However, the response to PD−1 therapy varies greatly among patients. Some patients experience long-term remission, while others show little or no response [7], [8]. Predicting which patients will benefit from PD−1 treatment remains a significant challenge in clinical practice [9]. Traditional methods for predicting treatment response and survival in NSCLC patients mainly rely on clinical factors such as TNM staging, histological subtypes, and biomarker status (e.g., PD-L1 expression). Although these factors are useful, they have limitations in accurately predicting the complex biological behavior of tumors and individual patient responses. For example, PD-L1 expression, which is widely used as a biomarker for PD−1 therapy, does not always accurately predict treatment response. There are patients with low PD-L1 expression who still respond well to PD−1 inhibitors, and vice versa [10], [11]. In recent years, radiomics has emerged as a promising tool in the field of oncology. Radiomics involves the extraction of a large number of quantitative features from medical images, such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) [12]. These features can capture the underlying biological characteristics of tumors, including tumor heterogeneity, angiogenesis, and immune-microenvironment, which are not easily detected by visual inspection of images [13], [14]. In the context of NSCLC treated with PD−1 inhibitors, pretreatment lung CT-based radiomics has the potential to play a crucial role. By analyzing the radiomic features of lung tumors before treatment, it may be possible to identify patterns that are associated with treatment response and survival. These features can potentially provide additional information beyond traditional clinical factors, enabling more accurate prediction of patient outcomes. For instance, certain texture-based radiomic features may be related to the tumor's immune-infiltrating status, which could influence the response to PD−1 therapy [15], [16]. The lack of understanding of predictions made by the models led to skepticism about its clinical application. In recent years, much effort has been made in explanatory machine learning to improve the interpretability of traditional complex black box machine learning models [17], [18]. In a recent study, Lundberg et al. explained predictions made by a clinical machine learning model by using the Shapley additive explanations (SHAP) framework, potentially increasing understanding and usability of their model. In this study, we aimed to explore the value of pretreatment lung CT-based radiomics in predicting the 3-year disease progression of NSCLC patients receiving PD−1 therapy. We hypothesized that radiomic features can serve as a powerful tool to complement traditional clinical factors and improve the accuracy of survival prediction in this specific patient population when modeled with machine learning and SHAP could yield accurate and explainable prediction of disease progression in patients with NSCLC. This research may contribute to the development of more personalized treatment strategies for NSCLC patients undergoing PD−1 immunotherapy. Materials and methods The study protocol was approved by the local ethics committee, approval number (2025-SR−694). Owing to the retrospective study design, the requirement for informed consent from the participants was waived. Patients and Clinical Data Data from patients histopathological confirmed of NSCLC between January 2018 and May 2022 were included in the analysis. The inclusion criteria were: (a) available baseline CT acquired just before antiPD−1/PD-L1 initiation (treatment-naïve imaging or imaging performed after other therapies but just before anti-PD−1/PD-L1 initiation); (b) presence of at least one measurable tumoral lesion according to the Immune-related Response Evaluation Criteria in Solid Tumors (iRECIST) version 1.1 on the baseline CT; (c) presence of at least one follow-up CT after 2 cycles of immunotherapy to assess response; (d) negative driver gene mutations. The exclusion criteria were: (a) the presence of other types of malignancy; (b) lack of radiomic or clinical signatures; (c) patients who received intercurrent radiotherapy or surgery on target lesions; (d) lost to follow-up; (e) discontinued treatment due to severe immune-related adverse events. Eventually, 173 participants were selected. All participants in the study cohort were stratified (7:3) into the training (n = 122) and testing (n = 51) groups by stratified random sampling, ensuring that the proportion of each class label is approximately maintained in both subsets. Information on baseline clinical characteristics, including age, sex, body mass index (BMI), smoking history, COPD status, Eastern Cooperative Oncology Group performance status (ECOG PS), histologic subtype of NSCLC, clinical stage and concurrent diseases. Pre-treatment peripheral blood biomarkers, such as absolute neutrophil count (ANC), absolute lymphocyte count (ALC), absolute eosinophil count (AEC), and platelet count (PLT), were also extracted from electronic medicalrecords. COPD is defined according to GOLD: post-bronchodilator forced expiratory volume in one second/forced vital capacity (FEV1/FVC) < 0.7 on spirometry. Follow-up All patients were consistently followed up 2 cycles of immunotherapy on imaging examinations for the first two years and then every 4–6 months thereafter. Tumour recurrence was determined based on radiologic evidence of new tumours. Three-year progression-free-Survival was defined as progressed at 3 years after immunotherapy. The 3-year disease progression status was the primary outcome. It was defined as a binary outcome (occurrence of disease progression or death from any cause within 3 years after initiation of therapy vs. remaining progression-free at the 3-year landmark time point). Pseudoprogression was excluded by performance status, symptoms and tumor markers such as CEA. If Pseudoprogression was suspected, CT was repeated 4 weeks later for confirmation. CT images acquisition and segmentation Patients underwent contrast enhancement chest CT. 3 ml/s (1.3 ml/kg of body weight, total of 60–110 ml) nonionic contrast medium (Iodamine, Iohexol, iodixanol) was injected through a high-pressure syringe with 45 s scan delay. A 64-detector row (SOMATOM Definition; Siemens) or 128-detector row (SOMATOM Definition AS+, Siemens) CT scanner was used for chest enhancement CT scanning. The scanning parameters were as follows: scanning type, helical full; tube current, 180 mA; voltage, 120kv; reconstruction matrix, 512*512; layer thickness, 5 mm; spacing, 5 mm. The scanning ranged from the thyroid cartilage’s lower edge to the liver’s lower edge. The arterial phase CT images were then uploaded to the PACS system. All CT images were independently reviewed by two experienced radiologists (M.Z. and H.L.L with 18 and 10 years of experience) to identify target lesions in the thorax and assess response at each time point according to iRECIST 1.1. All discrepancies between the two readers were reviewed until a consensus was reached. A radiologist (T.Z., with 8 years of experience) segmented the tumors on each slice manually in lung window setting (level = −600 HU; width = 1200 HU) using ITK-SNAP (http://www.itksnap.org) to obtain a three-dimensional volume of interest (VOI) of the whole tumor, which included both solid components and groundglass opacity of the tumor. A senior radiologist (M.Z., with 18 years of experience in thoracic imaging) was responsible for reviewing the segmentation of the tumors, and any discrepancies were addressed with further corrections. Feature selection and radiomics signature construction The study workflow is summarized in Fig. 1. A total of 107 handcrafted radiomic features—including shape, first-order statistics, and texture features—were extracted from all segmented lesions on the original CT images using PyRadiomics (version 2.3.0; Oncoradiomics SA, Belgium). Prior to feature extraction, all images were resampled to an isotropic voxel size of 1 × 1 × 1 mm³ to ensure consistency in voxel spacing. Workflow of the radiomic analysis in this study. The process commences with Tumor Segmentation, where regions of interest are manually annotated on computed tomography (CT) images. Subsequently, Feature Extraction is performed on the segmented tumor volumes, encompassing shape, texture, and first-order statistics, after resampling images to a uniform resolution of 1x1x1 mm. In the Feature Selection stage, robust features are identified based on resampling reproducibility (Pearson Correlation Coefficient > 0.90) and further refined using LASSO regression, with the potential integration of clinical risk factors. The selected features are then fed into logistic regression model to predict patient outcomes. Model performance is rigorously assessed through Model Validation, utilizing metrics such as concordance index, calibration curves, and decision curve analysis. Finally, Model Explanation is achieved using SHAP (SHapley Additive explanations) values to interpret feature contributions and illuminate the rationale behind the model's predictions. The study individually extracted radiomics features from the annotation of the lesion. Feature collinearity was assessed among radiomics features using pairwise correlation coefficients, and highly collinear features Pearson’scorrelation coefficient > 0.90) were excluded. LASSO regression (Least Absolute Shrinkage and Selection Operator) with 10-fold cross-validation was performed on the training set to select optimal radiomic features predictive of patient outcomes. Regularization parameter (λ) selection was guided by both minimum criteria and the one standard error rule. According to the minimum criteria Lasso model, we calculated Radscore for each candidate feature for building regression models. Where S denotes the set of selected features and xij is the standardized feature value for i. Rad-scores were calculated for both the training and test sets and visualized with ranked bar plots. Model development and validation Three logistic regression models were established: (1) Radiomics-only model, (2) Clinical-only model, incorporating COPD and tumor stage, and (3) Combined model integrating radioscore and clinical variables. For each model, predictive accuracy was evaluated in both the training and validation cohorts using receiver operating characteristic (ROC) analysis. The area under the ROC curve (AUC), sensitivity, specificity, and optimal threshold values were calculated. Model calibration was assessed using calibration curves and the Hosmer–Lemeshow goodness-of-fit test. SHAP evaluation To interpret the decision rules learned by the Logistic regression–based radiomic model, Shapley Additive Explanations (SHAP) were computed for each prediction. SHAP, a game theory–based approach, quantifies the contribution of each feature to the predicted probability of a given outcome—in this case, the risk of early disease progression. Feature contributions were estimated by perturbing input variables and modeling their influence on the prediction using a linear approximation, such as logistic regression, enabling intuitive interpretation of individual feature effects. Both the magnitude of feature impact and feature interactions were systematically analyzed. Statistical analysis Statistical analysis was performed using R software (version 4.3.3, http://www.R-project.org/) and SPSS software (version 20.0, IBM, Armonk, NY, USA). Categorical data were analyzed using the chi-square test or Fisher's exact test and presented as percentages. Based on previous literature, we selected 3-year disease progression status as the study endpoint. Patients were classified as progressors if disease progression (local or distant recurrence) or death from any cause occurred within 3 years; otherwise, they were classified as non-progressors. Logistic regression model was used to determine preoperative clinical risk factors associated with 3-year disease progession status. Univariate and multivariate logistic regression analyses were used to evaluate the relationship between various factors and prognosis. A p-value < 0.05 was considered statistically significant. Results Clinical characteristics A total of 173 patients with NSCLC were analysed in this study. 140 patients (80.9%) were males, and were treated with antiviral therapy regularly. 91 patients (52.6%) had a history of COPD. The median progression-free survival (PFS) time was 631.14 days (range 17–2314 days); 61 patients received immunotherapy as first-line therapy. 130 patients experienced tumour progression within three years. According to their 3-year follow up, all patients were divided into progression (n = 130) and non-progression (n = 43) groups. The baseline characteristics of the two groups are summarized in Table 1. History of COPD and PD-L1 expression among patients who accepted the gene test was significant different between the two groups. Logistic regression model was adapted to determine clinical risk factors associated with 3-year disease progression. Tumor stage and history of COPD were found to be independent clinical risk factors in multivariable logistic regression analysis (Table 2). Table 1Baseline characteristics of progression and non-progression groups.CharacteristicsProgression (n = 130)Non Progession (n = 43)P valueAge (years)61.22 ± 10.3163.53 ± 7.330.174Gender(male)105350.928BMI23.27 ± 3.1023.1 ± 2.720.642Pathology0.178Adenocacinoma7220Squamous cell carcinoma5117Undifferentiated carcinoma42Adenosquamous carcinoma20Sarcomatoid carcinoma11large cell carcinoma03Smoke79290.436Tumor stage0.062III2715IV10328Number of involved organs (tumor stage IV)0.25315017>15311Laboratory parameters*Hemoglobin124.08 ± 18.50126.53 ± 21.090.499Platelets(*109/L)249.72 ± 197.57226.14 ± 108.350.457Neutrophils(*109/L)5.11 ± 3.175.03 ± 2.560.885Eosinophils(*109/L)0.18 ± 0.290.20 ± 0.320.689FT34.02 ± 0.903.92 ± 0.870.549FT418.29 ± 12.9415.80 ± 2.970.213TSH3.94 ± 9.322.39 ± 1.720.282PD-L10.016<1%251–49%75≥ 50%185Not performed10328Line of therapy0.961144172 and more8626Concurrent diseaseHypertension82280.811Diabetes Mellitus2670.593Cirrhosis400.325COPD61300.009Radscore (median, interquartile range)−1.153 (−1.871, −0.623)−0.264(−0.941, −0.006)<0.001 Baseline characteristics of progression and non-progression groups. Table 2Univariate and multivariate logistic analyses of risk factors for progression.VariablesUnivariate analysisMultivariate analysisOR95%CIP valueOR95%CIP-valueAge(years)0.97(0.93–1.01)0.174Gender(male)1.03(0.50–2.12)0.928BMI1.02(0.90–1.16)0.742Pathology1.03(0.51–2.08)0.929Smoke0.78(0.39–1.54)0.473Tumor stage1.92(0.92–4.02)0.0832.65(1.09–6.45)0.032Number of involved organs (stage IV)1.63(0.70–3.81)0.256Hemoglobin0.99(0.98–1.01)0.499Platelets1.00(1.00–1.00)0.457Neutrophils(*109/L)1.01(0.90–1.13)0.885Eosinophils(*109/L)0.79(0.26–2.41)0.689FT31.13(0.76–1.67)0.549FT41.06(0.97–1.16)0.213TSH1.09(0.94–1.27)0.282Line of therapy1.02(0.51–2.04)0.961Hypertension1.07(0.53–2.16)0.851Diabetes Mellitus1.27(0.51–3.19)0.611Cirrhosis0.62(0.19–1.98)0.422COPD0.39(0.19–0.79)0.0090.32(0.14–0.74)0.008PD-L1 expression(<50%VS>50%)0.80(0.19–3.36)0.762 Univariate and multivariate logistic analyses of risk factors for progression. Feature selection and radiomics signature construction A total of 107 invalid features were removed, including 51 highly correlated values. To reduce dimensionality and prevent overfitting, LASSO regularized logistic regression was implemented. The optimal regularization parameter λ was chosen by 10-fold cross-validation based on the minimum mean deviance. Variables with non-zero coefficients at the selected λ were selected. Finally, the 4 most progression-related features were retained to construct a Logistic Regression model, including shape Sphericity, gldm Small Dependence High Gray Level Emphasis, glszm Gray Level Variance, glszm Small Area High Gray Level Emphasis from the original image. Development and performance of different models The clinical-only model, which incorporated COPD status and tumor stage, demonstrated a 3-year disease progression prediction with a mean AUC of 0.673 (95% CI, 0.576–0.770) in the training set and 0.620 (95% CI, 0.389–0.851) in the validation set. An individualized radiomics model showed better discrimination, with a mean AUC of 0.758 (95% Cl, 0.663–0.852) in the training set, and 0.815 (95% Cl, 0.618–1.000) in the validation set. The best performance was found in the model trained with both radiomic and clinical features, which achieved an AUC of 0.802 in the training cohort (95% bootstrap confidence interval: 0.721–0.884) and AUC of 0.836 in the testing cohort (95% bootstrap confidence interval: 0.663–1.000). The diagnostic performance of the different models is summarized in Table 3. The nomogram for 3-year disease progression is presented as Fig. 2. The calibration curve demonstrated good agreement between predicted and actual 3-year disease progression status in the training and test (Fig. 3) sets in different models. The decision curve showed good performance of the radiomics model in terms of clinical application, which added more benefit than either a treat-all or treat-none scheme (Fig. 4). Table 3Performance of radiomic and clinical models in the prediction of 3-year disease progression.ModelAUCSensitivitySpecificityPPVNPVCombined modelTraining cohort0.8020.7610.7650.8930.553Testing cohort0.8360.8810.7780.9490.583*Radiomic modelTraining cohort0.7580.8300.5880.8390.571Testing cohort0.8150.7860.8890.9710.471†Clinical modelTraining cohort0.6730.5570.7350.8450.391Testing cohort0.6200.8330.4440.8750.364Note. —AUC = area under the receiver operating characteristic curve, NPV = negative predictive value, PPV = positive predictive value. * Features used for the radiomic-only model are Radscore, † Features used for the clinical-only model are COPD and Tumor stage Performance of radiomic and clinical models in the prediction of 3-year disease progression. Note. —AUC = area under the receiver operating characteristic curve, NPV = negative predictive value, PPV = positive predictive value. * Features used for the radiomic-only model are Radscore, † Features used for the clinical-only model are COPD and Tumor stage Nomogram incorporating the Radscore with all independent clinical predictors for predicting 3-year disease progression in the traing set. Calibration curves for a 3-year disease progression model developed on the training set (A) and evaluated on the test set (B) in different models. The dashed diagonal line in each plot represents ideal calibration, where predicted probabilities perfectly align with observed event rates. The solid curve illustrates the actual calibration performance of the model, with circular markers indicating the average predicted probability for different risk strata and their corresponding observed three-year event rates. Decision curve for the three models in the training set. Model explanation with SHAP To interpret the contribution of individual predictors in each model, Shapley additive explanations (SHAP) were applied to the radiomics, clinical, and combined models. Beeswarm plots were generated to visualize the overall distribution and directionality of SHAP values for each feature. In these plots, each point represents an individual case, with the x-axis showing the SHAP value (feature impact on model output) and the color gradient indicating the corresponding feature value (low to high). Features were ranked according to their mean absolute SHAP values, allowing identification of the most influential predictors within each model. This approach enabled a unified comparison of feature importance across the radiomics-only, clinical-only, and integrated models, highlighting the predictive contribution of both imaging-derived features and clinical variables. SHAP values of each prediction models are summarized in Fig. 5. SHAP beeswarm plots showing the contribution and value distribution of key features in the radiomics, clinical, and combined models. Discussion The 3-year disease progression status of NSCLC patients receiving anti PD−1 immunotherapy varies greatly among individuals and is influenced by multiple factors. In this study, three main findings were observed as follow: (1) COPD was the clinical factors in predicting 3-year disease progression in patients with NSCLC following anti-PD1 therapy. (2) CECT based radiomics have good performance in predicting 3-year disease progression. (3) The model integrates COPD status, tumor stage, and CECT radiomics signature, all of which can be easily obtained, and can offer better prognostic information on 3-year disease progression. To the best of our knowledge, this is the first study to combine common concurrent diseases, clinical factors, and CECT radiomics for predicting the 3-year disease progression of NSCLC patients receiving PD−1 therapy. PD−1/PD-L1 inhibitors have reshaped therapeutic strategies in NSCLC, but efficient response prediction biomarkers are still lacking. Chronic Obstructive Pulmonary Disease (COPD) is a heterogeneous lung condition characterized by chronic respiratory symptoms (dyspnea, cough, sputum production and/or exacerbations) due to abnormalities of the airways (bronchitis, bronchiolitis) and/or alveoli (emphysema) that cause persistent, often progressive, airflow obstruction [19]. COPD often coexists with other diseases, including NSCLC. Common molecular pathways between COPD and NSCLC, exhibiting overlapping molecular alterations, contribute to their coexistence or exacerbate their symptoms [20]. The pathways responsible for activating immune-related responses through PD−1 and PD-L1 appear dysregulated in COPD patients [21], [22], [23]. Riondino, S et al. summarized studies concerning the mechanism and the efficacy of immunotherapy in patients with NSCLC and COPD. NSCLC Patients complicated by COPD (or emphysema) achieved better prognosis when treated with immune checkpoint inhibitors. Patients affected by COPD achieved significantly better PFS [24]. Our study also proved a positive correlation between COPD and PD−1-targeted therapy in NSCLC patients, which is consistent with previous studies. This may indicate that COPD is associated with immune eacape machanisms developed by tumors and have a high sensitivity to PD−1 blockade in patients with COPD. Radiomics was proposed as a noninvasive complementary method to biopsy for evaluation of the tumor and the tumor immune microenvironment. Several previous studies have investigated the capability of handcrafted radiomics analysis in predicting lung cancer prognosis [25], [26], [27], [28]. Coroller TP et al. suggested that CT-based radiomics were well adequate for predicting the pathological response of primary tumors and lymph nodes in NSCLC treated with preoperative neoadjuvant chemoradiotherapy [25]. Tian et al. demonstrated that noncontrast CT radiomics can predict PD−1 expression level of NSCLC to infer clinical outcomes response to immunotherapy [26]. Chang et al. established the CT radiomics signature has the potential to predict the 3-year and 5-year PFS in NSCLC patients treated with chemotherapy [27]. Wang et al. developed an outcome prediction model based on Radscore for predicting the PFS of extensive-stage small cell lung cancer after chemotherapy with immunotherapy [28]. Tian et al. found that deep learning CT radiomics can provide a noninvasive method to predict PD−1 expression of NSCLC and PD−1 expression level was associated with PFS [26]. In addition, Feng et al. also illustrated CT radiomics signature was independently associated with PFS in HCC patients [29]. In this study, we also demonstrated that the CECT radiomics signature, can offer prognostic information on 3-year disease progression, which is similar with the previous studies. This indicated that CECT radiomics signature was associated with 3-year disease progression in patients treated with anti PD−1 immunotherapy. However, we did not validate the PD-L1 expression characteristics with the CECT radiomics owing to our limited data. The nomogram is a common statistical tool that can accurately predict progression and survival in patients with different malignant tumors [27], [30]. Liu et al. found the nomogram integrated CT radiomics signature and clinical data could be useful in predicting treatment response and predict 3-year progression free survival in patients with locally advanced hypopharyngeal carcinoma [31]. Saad and his colleges [15] also developed a deep learning CT signature that successfully predicted survival in patients with NSCLC treated with immune checkpoint inhibitors using large real-world data from two cancer centers, and they also integrated deep learning signature and clinicopatholigcal and radiological risk factors into a joint model that achieved the best prediction. In our study, compared with only handcrafted radiomics approach, the nomogram model intergrated COPD and tumor stage performs better than the CECT-radiomics only, all of which can be easily obtained, to offer prognostic information on long-term survival, which is a key concern for patients. In addition, many studies have applied the nomogram to fit patient survival curves to guide clinical treatment, but most studies applied clinical factors combined with radiomic features are readily available and easy to interpret for patients, they are widely applied for survival analysis in clinical scenarios. Radiomic features allow quantitative information about a lesion on a CT image to simulate a physician’s diagnosis of a patient and are also widely used in survival analysis. This model demonstrated satisfactory discriminative ability, can be used to stratify patients with NSCLC into risk groups and guide treatment decisions. The recurrence patients in our prediction model are regarded as high-risk group who require close surveillance and may be candidates for clinical trials of therapies. One major concern with a radiomic-based prediction model is the interpretability of models. Radiomic models that rely on complex machine learning algorithms often have low clinical utility, as clinicians are unable to understand or explain certain predictions made by the model. To allow for interpretability, clinical models are often restricted to simple linear modeling methods, such as logistic regression, which often leads to lower accuracy. To address this issue, we applied the recently proposed SHAP values to interpret the model. The advantage of SHAP is that it can uncover patterns learned by complex prediction models without being restricted to simple modeling methods. The analysis identified shape sphericity as the most important factor in predicting 3-year disease progression. Sphericity is a measure of tumor compactness, with high sphericity indicating a compact spherical tumor. A previous study on head and neck cancer also yielded similar findings, in that tumor compactness measured with CT was prognostic of survival [32]. While there are no clear pathologic findings that relate to the compactness of a tumor, a possible explanation could be that complex nonspherical geometry could correspond to a more invasive and infiltrative tumor. Beyond shape characteristics, the remaining radiomics features selected in our model—SmallAreaHighGrayLevelEmphasis from GLSZM, SmallDependenceHighGrayLevelEmphasis from GLDM, and GrayLevelVariance—consistently reflect intratumoral textural heterogeneity. These features quantify small, high-intensity subregions and local gray-level fluctuations within the lesion, indicating the presence of heterogeneous tissue components such as variable cellular density, micro-necrosis, fibrosis, or uneven enhancement. T stage and overall stage were also found to be clear indicators of risk in the model, and these findings were consistent with conventional knowledge. In a previous study on advanced nasopharyngeal carcinoma, a radiomic model that combined contrast-enhanced T1-weighted MRI texture features with TNM staging achieved an AUC of 0.80 for predicting 3-year disease progression on a discovery and holdout test set, which was similar to our findings [33], despite involving a different tumor and imaging equipment. It is important to note that patterns explained by SHAP values do not directly explain the underlying sample characteristics but instead explain the pattern learned by the machine learning model. Outliers that deviate from the general trend of the data could not be explained. The results of this study demonstrate that explanatory machine learning techniques such as SHAP may offer value in understanding the predictions made by radiomic or clinical models that rely on machine learning. This study has several limitations. First, although we included a relatively large number of patients with NSCLC for prognosis prediction, this was a retrospective and single-territory center study with potential bias. Second, the DL model was developed and validated using data cohorts from the same hospital. Although there was no overlap between the internal and segmentectomy test sets, true external validation (using an external test set) was not carried out. Third, as a retrospective, real-world disease study, it might be biased by data heterogeneity in image protocol, therapy regimen, previous local treatment, line-of-treatment, outcome, and response metrics. Finally, many other variables could have influenced 3-year disease progression, such as pathological characteristics and retreatment after recurrence. To focus on the clinical question of long-term survival, we dichotomized survival time into a binary outcome. Although this approach enhances clinical interpretability, it may lead to a loss of statistical power and fails to capture the continuous relationship between variables and the risk of survival. Future studies could employ continuous time-to-event analysis methods, such as the Cox proportional hazards model, to further validate our findings. However, this pretherapeutic model showed favorable performance, suggesting good predictive value of the pretherapeutic radiomics features for 3-year disease progression. In the future, we will incorporate other survival-related factors into a predictive model. Conclusions We constructed a convenient and feasible radiomics model that integrated the CECT radiomics signature and clinical risk factors to predict 3-year disease progression in patients with NSCLC following anti-PD1 therapy. The ability of risk stratification in the pretherapeutic setting can identify patients for clinical trials and additional surveillance. Ethics The study protocol was approved by the local ethics committee, approval number (2025-SR−694). Owing to the retrospective study design, the requirement for informed consent from the participants was waived. Funding The author(s) received no financial support for the research, authorship, and/or publication of this article" or "This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. CRediT authorship contribution statement Min Zong: Writing – review & editing, Supervision, Investigation, Formal analysis, Conceptualization. Yi Chen: Writing – review & editing, Visualization, Supervision, Methodology, Investigation, Formal analysis, Conceptualization. Hong Zhou: Validation, Project administration, Methodology, Formal analysis, Data curation. Yi-ting He: Visualization, Methodology, Data curation. Min-jia Lin: Visualization, Validation, Software, Formal analysis, Data curation. Wen Huang: Writing – original draft, Software, Formal analysis, Data curation, Conceptualization. Declaration of Competing Interest We declare that we have no financial and personal relationships with otherpeople or organizations that can inappropriately influence our work, there is no professional or other personal interest of any nature or kind in any product, service and/or company that could be construed as influencing the position presented in.or the review of the manuscript entitled. Footnotes Summary statement: The CT-based radiomics nomogram incorporating tumor staging, COPD, and has preliminary clinical application prospects to predict efficacy in NSCLC patients. Contributor Information Wen Huang, Email: hw8708@126.com. Min-jia Lin, Email: 13616869291@163.com. Yi-ting He, Email: heyiting0119@163.com. Hong Zhou, Email: zhouhong@njmu.edu.cn. Yi Chen, Email: yi-chen.gzu@outlook.com. Min Zong, Email: mzong@njmu.edu.cn. References