Construction and validation of a recurrence prediction model for lung adenocarcinoma after treatment integrating radiomics features and clinical indicators Authors: Ruomin Liao, Yi Zhang, Gang Li Published: 2025 ABSTRACT ObjectiveTo investigate the feasibility and clinical value of constructing a predictive model for post-treatment recurrence risk in lung adenocarcinoma based on radiomics features (including texture and morphological features extracted from pretreatment chest CT images) and clinical indicators (age, gender, smoking history, pathological type, tumor-node-metastasis [TNM] stage, and tumor markers: carcinoembryonic antigen [CEA] and cytokeratin 19 fragment [CYFRA21-1]).MethodsA total of 350 lung adenocarcinoma patients who underwent standardized treatment from January 2020 to September 2024 were retrospectively enrolled and randomly divided into a training set (n = 245) and a validation set (n = 105) in a 7:3 ratio. Clinical data, pretreatment chest CT radiomics features, and tumor marker levels were collected. Univariate and multivariate logistic regression analyses were performed to identify independent risk factors, followed by the construction of a nomogram model. The model’s performance was evaluated and validated.ResultsThe recurrence rates were 33.11% (74/245) in the training set and 33.33% (35/105) in the validation set. Multivariate regression analysis showed that radiomics features (entropy and long axis ratio), TNM stage (III-IV), CEA level (≥ 5 ng/mL), and smoking history were independent risk factors (P < 0.05). The nomogram model demonstrated C-index values of 0.901 and 0.831 in the training and validation sets, respectively. Receiver operating characteristic curve analysis showed that the area under the curve (AUC) values for predicting post-treatment recurrence were 0.901 (95% CI: 0.854–0.948) and 0.831 (95% CI: 0.732–0.930) in the training and validation sets, with sensitivities of 0.846 and 0.682, specificities of 0.860 and 0.840, precision-recall AUCs of 0.886 and 0.812, positive predictive values of 0.821 and 0.793, and negative predictive values of 0.879 and 0.785, respectively. Calibration curves indicated good agreement between predicted and observed recurrence, supported by Hosmer-Lemeshow test P-values of 0.743 and 0.709.ConclusionThe nomogram model incorporating radiomics features and clinical indicators effectively predicts the risk of post-treatment recurrence in lung adenocarcinoma, offering a practical tool for personalized follow-up and intervention strategies.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-025-15427-8. Introduction Lung adenocarcinoma, the most prevalent histological subtype of lung cancer, accounts for approximately 40%–50% of all lung cancer cases, with its incidence steadily increasing annually [1]. Despite advances in multimodal treatments including surgery, chemotherapy, and targeted therapy, patient prognosis remains unsatisfactory, largely due to post-treatment recurrence, which contributes significantly to therapeutic failure and reduced survival [2]. Studies indicate that the postoperative recurrence rate of lung adenocarcinoma can reach 30%–50%, and the 5-year survival rate of recurrent patients significantly declines to only 10%–20%. Therefore, the early and accurate prediction of recurrence risk following treatment is of critical clinical importance for optimizing follow-up strategies and timely interventions to improve patient outcomes [3, 4]. Currently, clinical indicators commonly used to assess recurrence risk in lung adenocarcinoma include the Tumor-Node-Metastasis (TNM) staging system, tumor markers, and pathological characteristics. However, their predictive performance is limited. For instance, while TNM staging reflects tumor progression, patients within the same stage exhibit significant individual variations in recurrence risk [3]. Similarly, tumor markers lack sufficient sensitivity and specificity to serve as reliable standalone predictors. In recent years, radiomics has emerged as a novel technique capable of quantifying tumor heterogeneity by extracting high-throughput texture and morphological features from medical images, thereby providing new biomarkers for prognostic evaluation [5, 6]. Evidence suggests that radiomic features are closely associated with tumor invasion, metastasis, and recurrence in lung cancer, complementing conventional clinical indicators [7]. However, the predictive value of a single modality remains limited, and the integration of multimodal data has become essential for enhancing model performance. A nomogram, as a visual predictive tool, consolidates multiple independent risk factors into an intuitive scoring system, facilitating quantitative risk assessment for individual patients. Thus, this study aimed to combine radiomic features with clinical indicators to construct a nomogram for predicting post-treatment recurrence risk in lung adenocarcinoma, followed by internal validation to evaluate its efficacy. The findings are expected to provide a reliable basis for personalized follow-up and intervention strategies in clinical practice. Materials and methods Study population This retrospective cohort study included lung adenocarcinoma patients who underwent standardized treatment (surgery, chemotherapy, targeted therapy, etc.) at our institution between January 2020 and September 2024. Inclusion criteria: Pathologically confirmed lung adenocarcinoma (including pure adenocarcinoma and adenosquamous carcinoma). Completion of pretreatment chest CT scans with complete and analyzable imaging data. Availability of complete clinical data (age, sex, smoking history, TNM stage, etc.) and tumor marker (CEA, CYFRA21-1) results. A minimum follow-up period of 6 months with complete records (the median follow-up duration was 28.6 months (range: 6–57 months), and 74.3% of patients had a follow-up time ≥ 12 months). Exclusion criteria: Concurrent malignancies. Pretreatment distant metastasis. CT images with significant artifacts or motion interference, preventing radiomic feature extraction. Missing clinical, imaging, or follow-up data. Sample size estimation Sample size estimation was based on the primary outcome of post-treatment recurrence. Considering previous studies and institutional clinical data, the expected recurrence rate was set at 30%–35% (consistent with the observed recurrence rate of 33.11%–33.33% in this study). The calculation was conducted using the “Sample Size Calculation” module in SPSS 26.0, with verification via the “pwr” package in R 4.2.3. The significance level (α) was set at 0.05 (two-tailed), with a statistical power (1-β) of 80%. Accounting for a 15% dropout rate (due to factors such as incomplete imaging or missing follow-up), the minimum required sample size was determined to be 310. Accordingly, 350 consecutive lung adenocarcinoma patients who received standardized treatment were enrolled. The actual sample size exceeded the minimum requirement, and statistical power (1-β > 80%) was confirmed via SPSS 26.0, fulfilling the prerequisites for multivariate logistic regression analysis and model construction. Data collection Clinical data: Baseline data were extracted from electronic medical records, including demographics such as age, gender, smoking history (defined as ≥ 100 cigarettes/year; smoking duration was calculated). Clinical scores: Modified Medical Research Council (mMRC) dyspnea scale (0–4) [8] and Visual Analog Scale (VAS) for pain (0–10) [9]. Pathological and staging data: Histological type (adenocarcinoma/adenosquamous carcinoma), TNM stage. Tumor markers: Pretreatment fasting venous blood levels of carcinoembryonic antigen (CEA) and cytokeratin 19 fragment antigen 21 − 1 (CYFRA21-1). Treatment information: Detailed treatment modalities (surgery alone, chemotherapy alone, targeted therapy alone) were retrospectively extracted from electronic medical records to assess treatment heterogeneity. Detailed treatment information was collected and categorized as: (1) Surgery (n = 218, radical resection of primary tumor); (2) Chemotherapy (n = 76, platinum-based doublet chemotherapy); (3) Targeted therapy (n = 56, tyrosine kinase inhibitor [TKI] therapy based on driver gene mutations); (4) Combined therapy (n = 32, combination of two or more of the above modalities). Radiomic feature extraction: Pretreatment chest CT images (slice thickness ≤ 1 mm, pitch ≤ 1.0) were collected. Two radiologists with ≥ 5 years of thoracic imaging experience independently performed manual segmentation of tumor regions (region of interest, ROI) using ITK-SNAP, encompassing all solid components of the primary tumor. The interobserver intraclass correlation coefficient (ICC) was calculated using a two-way random effects model (absolute agreement) via SPSS 26.0, based on tumor volume and maximum diameter of the segmented ROIs. If ICC < 0.85, a third senior radiologist performed re-segmentation for arbitration; if ICC ≥ 0.85, the ROI from the radiologist with 8 years of thoracic imaging experience was used for feature extraction. Detailed radiomics methodology parameters were provided in Supplemental Table 1. PyRadiomics was employed to extract [10]: Morphological features: Maximum tumor diameter (mm), tumor volume (mm³), elongation ratio (short-axis/long-axis), sphericity. Texture features: Gray-level co-occurrence matrix (e.g., contrast), gray-level run-length matrix (e.g., long-run emphasis), standard deviation, entropy, energy. Intensity features: Mean CT value (HU). Only features with ICC ≥ 0.85 were retained for stability. Outcome definition Recurrence was adjudicated per the International Association for the Study of Lung Cancer (IASLC) Staging Manual (8th Edition) [9] and the Chinese Society of Clinical Oncology (CSCO) Non-Small Cell Lung Cancer Guidelines (2024) [11]. Patients were classified into “recurrence” or “non-recurrence” groups based on ≥ 6 months of follow-up. Recurrence group: Any of the following post-treatment: New tumor lesions at the primary site or elsewhere in the lungs, pathologically confirmed as lung adenocarcinoma (consistent with the primary). Regional lymph node (hilar/mediastinal) enlargement, confirmed as metastatic via biopsy or PET-CT. Distant metastases (brain, bone, liver, adrenal, etc.) confirmed via imaging (CT/MRI/PET-CT) or pathology. Persistent elevation of tumor markers (CEA, CYFRA21-1) unexplained by benign conditions, with imaging suggesting progression. Two clinicians (one medical oncologist and one radiologist, each with ≥ 5 years of experience) independently verified recurrence. Discrepancies were resolved by a third senior clinician. Non-recurrence group: No evidence of recurrence or metastasis during follow-up, with stable imaging and tumor marker findings. Statistical analysis Data were analyzed using SPSS 26.0 and R 4.2.3. Normally distributed continuous variables were expressed as mean ± standard deviation (SD), compared using independent t-tests. Categorical variables were presented as n (%), analyzed using χ² or Fisher’s exact tests. Risk factor selection: A stepwise approach was employed: Univariate logistic regression (P < 0.05 for inclusion) identified potential predictors in the training set. Multivariate logistic regression (backward stepwise, P < 0.05) determined independent influencing factors, with odds ratios (ORs) and 95% confidence intervals (CIs) calculated. As a supplementary analysis to address the limitations of stepwise regression, Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression was performed. Model development and validation: The “rms” package in R constructed a nomogram. Internal validation used bootstrap resampling (1000 iterations). The receiver operating characteristic (ROC) curve was plotted, and the area under the curve (AUC) value was calculated (AUC > 0.9: excellent; 0.7–0.9: good). A calibration curve was plotted and evaluated using the Hosmer - Lemeshow goodness - of - fit test (P > 0.05 indicated good fit). Decision curve analysis (DCA) assessed net benefit across threshold probabilities. Model stability: Adhering to the “10 events per variable” rule (5 predictors required ≥ 25 events), with variance inflation factor (VIF) < 2 confirming low multicollinearity. Significance was set at α = 0.05. Results Comparison of baseline characteristics in the training and validation sets Among 350 enrolled patients, a 7:3 ratio allocated 245 to the training set (74 recurrences, 33.11%) and 105 to the validation set (35 recurrences, 33.33%) randomly. No significant differences in baseline characteristics were observed (P > 0.05) (Table 1). Table 1Comparison of baseline characteristics between training and validation setsVariablesTraining cohort (n = 245)Validation cohort (n = 105)t/χ² P Age (years)63.52 ± 7.8962.98 ± 7.560.6820.495Gender (Male/Female)138/10758/470.0930.761Smoking history (years)18.64 ± 10.3217.95 ± 9.870.6150.539mMRC (score)1.89 ± 0.931.80 ± 0.900.7110.478VAS (0–10 points)4.25 ± 1.754.20 ± 1.440.2330.516Pathological type (Adenocarcinoma/Adenosquamous carcinoma)230/1598/70.1030.748TNM stage (I-II/III-IV)147/9862/430.0280.868CEA(ng/mL)5.89 ± 2.125.76 ± 1.980.6320.528CYFRA21-1(ng/mL)3.67 ± 1.343.59 ± 1.250.5680.570Maximum tumor diameter (mm)29.34 ± 8.5628.76 ± 8.120.6530.514Tumor volume (mm³)1285.67 ± 432.581262.34 ± 418.750.5720.567Long-axis to short-axis ratio0.68 ± 0.150.67 ± 0.140.5920.592Sphericity0.72 ± 0.110.71 ± 0.100.6430.521Radiomics energy value125.67 ± 35.42123.89 ± 34.150.5180.605Gray-level co-occurrence matrix contrast23.56 ± 8.7222.98 ± 8.350.6150.539Gray-level run-length matrixlong-run emphasis0.32 ± 0.120.31 ± 0.110.6780.498Gray-level standard deviation18.76 ± 5.3218.35 ± 5.110.6820.495Radiomics entropy4.23 ± 0.894.18 ± 0.850.5470.585Mean gray-level value (CT value, HU)45.67 ± 10.3444.98 ± 9.870.6120.541 Table 1Comparison of baseline characteristics between training and validation setsVariablesTraining cohort (n = 245)Validation cohort (n = 105)t/χ² P Age (years)63.52 ± 7.8962.98 ± 7.560.6820.495Gender (Male/Female)138/10758/470.0930.761Smoking history (years)18.64 ± 10.3217.95 ± 9.870.6150.539mMRC (score)1.89 ± 0.931.80 ± 0.900.7110.478VAS (0–10 points)4.25 ± 1.754.20 ± 1.440.2330.516Pathological type (Adenocarcinoma/Adenosquamous carcinoma)230/1598/70.1030.748TNM stage (I-II/III-IV)147/9862/430.0280.868CEA(ng/mL)5.89 ± 2.125.76 ± 1.980.6320.528CYFRA21-1(ng/mL)3.67 ± 1.343.59 ± 1.250.5680.570Maximum tumor diameter (mm)29.34 ± 8.5628.76 ± 8.120.6530.514Tumor volume (mm³)1285.67 ± 432.581262.34 ± 418.750.5720.567Long-axis to short-axis ratio0.68 ± 0.150.67 ± 0.140.5920.592Sphericity0.72 ± 0.110.71 ± 0.100.6430.521Radiomics energy value125.67 ± 35.42123.89 ± 34.150.5180.605Gray-level co-occurrence matrix contrast23.56 ± 8.7222.98 ± 8.350.6150.539Gray-level run-length matrixlong-run emphasis0.32 ± 0.120.31 ± 0.110.6780.498Gray-level standard deviation18.76 ± 5.3218.35 ± 5.110.6820.495Radiomics entropy4.23 ± 0.894.18 ± 0.850.5470.585Mean gray-level value (CT value, HU)45.67 ± 10.3444.98 ± 9.870.6120.541 Comparison of baseline characteristics between training and validation sets Gray-level run-length matrix long-run emphasis Univariate analysis of factors influencing post-treatment recurrence in lung adenocarcinoma In the training set, univariate analysis showed statistically significant differences between the recurrence and non-recurrence groups regarding smoking history, TNM stage (III–IV), CEA levels, long-axis-to-short-axis ratio, and radiomic entropy (all P < 0.05) (Table 2). Table 2Univariate analysis of factors influencing post-treatment recurrence in lung adenocarcinomaVariablesRecurrence group (n = 74)Non-Recurrence group (n = 171)t/χ² P Age (years)64.85 ± 7.9063.21 ± 7.401.5600.12Gender (Male/Female)39/3599/720.5660.451Smoking history (years)22.84 ± 10.5516.24 ± 10.014.6610.001mMRC (score)1.91 ± 0.931.88 ± 0.930.2320.817VAS (0–10 points)4.26 ± 1.754.24 ± 1.720.0830.934Pathological type (Adenocarcinoma/Adenosquamous carcinoma)68/6162/90.3000.573TNM stage (I-II/III-IV)35/39112/597.1280.007CEA(ng/mL)6.84 ± 2.325.50 ± 2.014.5690.001CYFRA21-1(ng/mL)3.80 ± 1.403.60 ± 1.011.2590.209Maximum tumor diameter (mm)29.44 ± 8.5829.21 ± 8.320.1970.844Tumor volume (mm³)1288.67 ± 432.551286.67 ± 432.500.0330.973Long-axis to short-axis ratio0.48 ± 0.110.69 ± 0.208.4880.001Sphericity0.73 ± 0.100.71 ± 0.111.3420.181Radiomics energy value125.87 ± 35.48125.43 ± 35.320.0890.929Gray-level co-occurrence matrix contrast23.66 ± 8.7823.26 ± 8.710.3290.742Gray-level run-length matrixlong-run emphasis0.33 ± 0.140.30 ± 0.111.7800.073Gray-level standard deviation18.86 ± 5.3818.56 ± 5.200.4100.681Radiomics entropy4.53 ± 0.924.01 ± 0.774.5690.001Mean gray-level value (CT value, HU)45.77 ± 10.3445.57 ± 10.010.1420.887 Table 2Univariate analysis of factors influencing post-treatment recurrence in lung adenocarcinomaVariablesRecurrence group (n = 74)Non-Recurrence group (n = 171)t/χ² P Age (years)64.85 ± 7.9063.21 ± 7.401.5600.12Gender (Male/Female)39/3599/720.5660.451Smoking history (years)22.84 ± 10.5516.24 ± 10.014.6610.001mMRC (score)1.91 ± 0.931.88 ± 0.930.2320.817VAS (0–10 points)4.26 ± 1.754.24 ± 1.720.0830.934Pathological type (Adenocarcinoma/Adenosquamous carcinoma)68/6162/90.3000.573TNM stage (I-II/III-IV)35/39112/597.1280.007CEA(ng/mL)6.84 ± 2.325.50 ± 2.014.5690.001CYFRA21-1(ng/mL)3.80 ± 1.403.60 ± 1.011.2590.209Maximum tumor diameter (mm)29.44 ± 8.5829.21 ± 8.320.1970.844Tumor volume (mm³)1288.67 ± 432.551286.67 ± 432.500.0330.973Long-axis to short-axis ratio0.48 ± 0.110.69 ± 0.208.4880.001Sphericity0.73 ± 0.100.71 ± 0.111.3420.181Radiomics energy value125.87 ± 35.48125.43 ± 35.320.0890.929Gray-level co-occurrence matrix contrast23.66 ± 8.7823.26 ± 8.710.3290.742Gray-level run-length matrixlong-run emphasis0.33 ± 0.140.30 ± 0.111.7800.073Gray-level standard deviation18.86 ± 5.3818.56 ± 5.200.4100.681Radiomics entropy4.53 ± 0.924.01 ± 0.774.5690.001Mean gray-level value (CT value, HU)45.77 ± 10.3445.57 ± 10.010.1420.887 Univariate analysis of factors influencing post-treatment recurrence in lung adenocarcinoma Gray-level run-length matrix long-run emphasis Multivariate logistic regression analysis of factors influencing post-treatment recurrence in lung adenocarcinoma Recurrence status (non-recurrence = 0, recurrence = 1) was used as the dependent variable, and relevant factors were incorporated into a multivariate logistic regression model (Supplemental Table 2). The results demonstrated that smoking history, CEA levels, TNM stage, and radiomic entropy were independent risk factors for post-treatment recurrence (all P < 0.05). Conversely, the long-axis-to-short-axis ratio was identified as an independent protective factor (all P < 0.05) (Table 3). Table 3The result of multivariate logistic regression analysisVariablesβSEWald P OR95%CISmoking history0.0590.01810.9360.0011.0601.024–1.098TNM stage0.9380.3746.2940.0122.5551.188–4.815CEA0.2700.0879.6080.0021.3101.104–1.553Long-axis-to-short-axis ratio−7.0681.24032.5110.0010.0010.001–0.010Radiomic entropy0.7680.23111.0280.0012.1561.370–3.392Constant−2.3451.3463.0330.0820.096 Table 3The result of multivariate logistic regression analysisVariablesβSEWald P OR95%CISmoking history0.0590.01810.9360.0011.0601.024–1.098TNM stage0.9380.3746.2940.0122.5551.188–4.815CEA0.2700.0879.6080.0021.3101.104–1.553Long-axis-to-short-axis ratio−7.0681.24032.5110.0010.0010.001–0.010Radiomic entropy0.7680.23111.0280.0012.1561.370–3.392Constant−2.3451.3463.0330.0820.096 The result of multivariate logistic regression analysis Construction of the nomogram prediction model Based on the multivariate logistic regression results, a nomogram was developed to predict post-treatment recurrence in lung adenocarcinoma. Each risk factor was assigned a specific point scale, and individual scores were calculated by mapping patient-specific values onto the corresponding scale. The total score was obtained by summing these points, with the final value representing the predicted recurrence probability (Fig. 1). Fig. 1Nomogram for Predicting Post-Treatment Recurrence in Lung Adenocarcinoma (X1: Smoking history, X2: TNM stage, X3: CEA, X4: Long-axis-to-short-axis ratio, X5: Radiomic entropy) Nomogram for Predicting Post-Treatment Recurrence in Lung Adenocarcinoma (X1: Smoking history, X2: TNM stage, X3: CEA, X4: Long-axis-to-short-axis ratio, X5: Radiomic entropy) Evaluation and validation of the nomogram model The nomogram demonstrated excellent calibration and discriminative ability in both training and validation sets (C-index: 0.901 and 0.831, respectively; mean absolute error: 0.115 and 0.153, respectively; Hosmer-Lemeshow test: χ² = 5.216, P = 0.743 and χ² = 5.445, P = 0.709, respectively) (Fig. 2). ROC curve analysis showed that the AUC values for predicting post-treatment recurrence were 0.901 (95% CI: 0.854–0.948) and 0.831 (95% CI: 0.732–0.930) in the training and validation sets, with sensitivities of 0.846 and 0.682, specificities of 0.860 and 0.840, precision-recall AUCs of 0.886 and 0.812, positive predictive values of 0.821 and 0.793, and negative predictive values of 0.879 and 0.785, respectively (Fig. 3). The result of LASSO logistic regression confirmed the selection of the same five independent risk factors with consistent model performance (training set AUC = 0.897, validation set AUC = 0.826). Fig. 2Calibration curves in the training set (A) and the validation set (B) Calibration curves in the training set (A) and the validation set (B) Fig. 3ROC curves in the training set (A) and the validation set (B) ROC curves in the training set (A) and the validation set (B) Sensitivity analysis To address treatment heterogeneity, sensitivity analysis was performed by subgrouping patients based on treatment modalities. The nomogram maintained good discriminative ability across all subgroups: surgery alone (AUC = 0.887, 95% CI: 0.821–0.953), chemotherapy alone (AUC = 0.862, 95% CI: 0.775–0.949), and targeted therapy alone (AUC = 0.875, 95% CI: 0.791–0.959). This confirms the robustness of the independent predictive factors across different treatment types. Performance comparison of multimodal models The combined model (integrating all five factors) demonstrated superior discriminative ability (Training set AUC: 0.901; Validation set AUC: 0.831) compared to the clinical-only model (including smoking history, TNM stage, CEA) (AUC: 0.723/0.689) and Radiomics-only model (including entropy, long-axis-to-short-axis ratio) (AUC: 0.815/0.762). This result confirmed the incremental value of integrating clinical and radiomics data. Decision curve analysis of the nomogram model DCA indicated that within a threshold probability range of 0.1–0.9, the nomogram provided greater net clinical benefit than strategies assuming either universal recurrence or non-recurrence (Supplemental Fig. 1). Discussion Accurate assessment of recurrence risk following treatment for lung adenocarcinoma is a crucial aspect of clinical prognosis management. The ability to predict recurrence risk using multidimensional indicators is of great significance for developing individualized follow-up strategies and initiating timely interventions to prolong patient survival [11]. This study integrated radiomics features and clinical indicators based on data from 350 lung adenocarcinoma patients. Through univariate and multivariate logistic regression analyses, five independent predictive factors were identified, and a nomogram prediction model was constructed. The model demonstrated excellent predictive performance in both the training and validation cohorts (AUCs of 0.901 and 0.831, respectively). Notably, radiomics entropy, long-axis-to-short-axis ratio, and CEA levels contributed significantly to the model, suggesting that multimodal data integration substantially enhances recurrence risk prediction accuracy, providing a new theoretical foundation and practical tool for clinical precision assessment. To mitigate the potential limitations of stepwise regression (e.g., model instability and inflated effect sizes in high-dimensional data), LASSO logistic regression was performed as a supplementary analysis, which validated the stability of the identified five predictors and yielded consistent model performance (training set AUC = 0.897, validation set AUC = 0.826), further confirming the reliability of our findings. Treatment heterogeneity was a potential confounding factor for recurrence prediction. Subgroup sensitivity analysis showed that the model performed consistently well across different treatment modalities, indicating that the identified independent predictors are not significantly affected by treatment type. This is because the core predictive factors (e.g., radiomics entropy reflecting tumor heterogeneity, long-axis-to-short-axis ratio representing morphological characteristics) are inherent biological properties of tumors, which are less likely to be altered by treatment and thus maintain stable predictive value across different therapeutic strategies. In this study, radiomics features emerged as independent predictors with core value, a mechanism that can be explained through the quantitative characterization of tumor heterogeneity. Radiomics entropy reflects the complexity of intratumoral grayscale distribution, with higher entropy indicating more disordered tumor cell proliferation and differentiation as well as greater genomic instability, which directly correlates with tumor aggressiveness and recurrence potential [12, 13]. Studies have confirmed that tumor heterogeneity is a key factor driving treatment resistance and recurrence. Elevated entropy may be accompanied by enhanced biological behaviors such as vasculogenic mimicry and epithelial-mesenchymal transition, promoting tumor cell escape from therapeutic elimination and the formation of metastatic lesions [14]. In this study, entropy demonstrated a strong predictive capacity for recurrence (OR = 2.156, 95% CI: 1.370–3.392), consistent with previous radiomics studies indicating that “high entropy predicts poor prognosis” in lung cancer. Long-axis-to-short-axis ratio, as a morphological feature, acted as a protective factor (OR = 0.001), indicating that a higher ratio (i.e., a more spherical tumor shape) was associated with a lower recurrence risk. This finding can be explained by tumor growth patterns: round or oval tumors (long-axis-to-short-axis ratio close to 1) typically exhibit well-defined margins and uniform growth rates, suggesting more homogeneous tumor cell differentiation and lower invasiveness. In contrast, irregularly shaped tumors (low long-axis-to-short-axis ratio) are more likely to breach the basement membrane and metastasize via stromal or vascular invasion [15]. Additionally, Long-axis-to-short-axis ratio is associated with microvessel density, with irregular tumors often exhibiting disorganized vascular networks that facilitate tumor cell dissemination [16]. The inclusion of long-axis-to-short-axis ratio addresses a limitation of traditional tumor size assessments (e.g., maximum diameter), which fail to account for morphological heterogeneity, thereby enhancing the model’s predictive value. Among clinical indicators, smoking history, CEA levels, and TNM stage exhibited predictive roles consistent with the known biological behavior of lung adenocarcinoma [17]. Smoking history was an independent risk factor (OR = 1.060), with mechanisms linked to DNA damage induced by polycyclic aromatic hydrocarbons and nicotine, as well as chronic inflammatory microenvironments promoting residual tumor cell proliferation and metastasis [18, 19]. Elevated CEA levels (OR = 1.310), a classic tumor marker, reflect increased secretory activity of tumor cells and are closely associated with lymph node metastasis and pleural invasion in lung adenocarcinoma. Dynamic CEA monitoring has been validated as a critical tool for recurrence surveillance. TNM stage was identified as a risk factor in this study. Based on variable coding (stage I–II = 0, stage III–IV = 1), the results indicated that stage III–IV patients faced a higher recurrence risk, aligning with clinical observations—advanced tumors have greater primary tumor burden, higher lymph node metastasis rates, and thus increased likelihood of residual disease or micrometastases post-treatment [20, 21]. However, it should be noted that recurrence risk varies even within the same TNM stage (e.g., some stage I patients experience early recurrence due to tumor location or molecular characteristics). Therefore, TNM staging should be combined with radiomics features to improve individualized prediction accuracy [22]. The nomogram model developed in this study integrates these multidimensional indicators into a visual tool with three key advantages. First, it enables risk stratification through quantifiable scoring, allowing clinicians to rapidly estimate recurrence probability using CT imaging features, laboratory results, and staging information without requiring complex statistical knowledge. Second, by incorporating both radiomics and clinical indicators, the model accounts for microscopic tumor heterogeneity (entropy), morphological features (long-axis-to-short-axis ratio), and macroscopic clinical characteristics (stage, biomarkers), offering more comprehensive predictions than single-indicator models. Third, internal validation demonstrated strong stability (validation set AUC = 0.831), suggesting generalizability in similar populations. Clinically, this model can help identify high-risk patients (e.g., those with scores ≥ 70) who may benefit from intensified follow-up or early adjuvant therapy, while low-risk patients could undergo extended follow-up intervals, reducing healthcare burdens. This study has several limitations. First, its single-center retrospective design may introduce selection bias, and variations in treatment regimens (e.g., surgical methods, chemotherapy agents) could affect recurrence outcomes. Future multicenter prospective studies are needed to validate the model’s performance. Second, radiomics feature extraction relied on manual segmentation, which, despite being validated by intraclass correlation coefficient (ICC) analysis, may introduce subjective errors. Automated segmentation algorithms could improve consistency. Third, molecular pathological features, which are closely linked to recurrence risk, were not included; their integration might further enhance predictive accuracy. Additionally, the follow-up period was relatively short, and the model’s predictive value for long-term recurrence (e.g., 5-year recurrence) requires validation with extended data. In conclusion, the nomogram model integrating radiomics features (entropy, long-axis-to-short-axis ratio) and clinical indicators (smoking history, TNM stage, CEA) effectively predicts recurrence risk in lung adenocarcinoma patients after treatment. Radiomics features provide critical tumor heterogeneity information, enhancing the model’s precision. Clinical translation of this model may optimize follow-up strategies and support individualized intervention decisions. Further refinements incorporating molecular biomarkers and multicenter data may establish this tool as a cornerstone of lung adenocarcinoma prognosis management. Supplementary Information Supplementary Material 1. Supplementary Material 1.