Interpretable machine learning model integrating delta-radiomics enhances postoperative recurrence prediction in early-stage lung adenocarcinoma

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide (1). The widespread implementation of lung cancer screening has led to the increased detection of early-stage lung cancers (2). Among all lung cancer cases, lung adenocarcinoma (LUAD) is the most common histological subtype, constituting approximately 40% (3). Although patients with early-stage LUAD are candidates for curative resection and generally have a favorable prognosis, significant heterogeneity in outcomes exists even among tumors of the same stage (4,5). For instance, the 5-year recurrence-free survival (RFS) for stage IA LUAD ranges from 63% to 81% (6), with most recurrences occurring within the first 3 years post-surgery (7,8). Therefore, the accurate identification of patients at high risk of recurrence, enabling intensified postoperative surveillance and timely intervention, may lead to improved clinical outcomes.

Computed tomography (CT) screening and follow-up play a pivotal role in the management of pulmonary nodules (9). Derived from serial CT imaging, the volume doubling time (VDT) has demonstrated additional prognostic value for disease-free survival in LUAD (10). However, VDT relies solely on changes in tumor volume and fails to capture the evolution of subtle intratumoral textural or morphological characteristics. Currently, radiomics is widely applied in medical imaging, enabling the high-throughput extraction and quantitative analysis of image features (11,12). As an extension, delta radiomics quantifies the temporal changes in radiomic features, offering dynamic insights into disease evolution (13). This approach has achieved significant progress in differentiating benign from malignant pulmonary nodules (14), predicting the invasiveness of early-stage LUAD (15), and evaluating treatment response in advanced lung cancer (16-18). Given its capacity to quantify tumor evolution, delta radiomics is particularly suitable for characterizing the phenotypic dynamics associated with aggressive tumor behavior. However, to date, no prognostic biomarker based on preoperative serial CT radiomics has been specifically developed for predicting early postoperative recurrence in stage IA LUAD.

This study aims to develop and evaluate an interpretable multimodal radiomics model that integrates clinical factors, conventional radiological features, preoperative radiomics features, and delta-radiomics features derived from serial CT data to predict early postoperative recurrence in patients with stage IA LUAD and to assess its incremental value over traditional criteria. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0812/rc).

Methods

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and approved by the Institutional Ethics Committee of The First Medical Center of the Chinese PLA General Hospital (Approval No. S2025-035-01). The requirement for informed consent was waived due to the retrospective nature of the study.

Patients

This retrospective cohort study enrolled patients diagnosed with early-stage LUAD at The First Medical Center of the Chinese PLA General Hospital between January 2011 and June 2023. The inclusion criteria were as follows: (I) patients underwent curative-intent surgery; (II) pathologically confirmed stage IA LUAD; (III) patients either experienced recurrence within 36 months after surgery or were followed up for at least 36 months in the absence of recurrence; (IV) availability of both baseline chest CT and at least one preoperative follow-up CT scan, with an interval greater than 20 days but less than 1 year between the two scans. Exclusion criteria included: (I) incomplete clinical or imaging data; (II) an interval exceeding 30 days between the last follow-up CT and surgery; (III) receipt of any neoadjuvant antitumor therapy; (IV) a history of other malignancies or multiple primary lung cancers.

Clinical data collected encompassed sex, age, smoking history, history of chronic obstructive pulmonary disease (COPD), family history of malignancy, surgical procedure (wedge resection, segmentectomy, or lobectomy), and pathological subtype. Tumors with a high-grade histological pattern (solid, micropapillary, or complex glandular) accounting for ≥5% were categorized into the high-grade component group; the remainder were classified as the non-high-grade group. Follow-up information was obtained through medical records and telephone interviews, with the last follow-up conducted in March 2025. The primary outcome was early postoperative recurrence, defined as any recurrence occurring within 36 months after curative surgery and confirmed by imaging or pathology. The 36-month cutoff was selected because prior studies have demonstrated that most recurrences in stage IA LUAD occur within the first 3 years post-surgery, and 3-year RFS is a standard early endpoint in resected early-stage lung cancer (7,8). RFS and overall survival (OS) served as secondary outcomes. RFS was defined as the time from surgery to recurrence, death, or last follow-up, with deaths without prior recurrence censored at the date of death. OS was defined as the time from surgery to death or last follow-up.

CT examination and image analysis

All subjects underwent non-contrast CT scans at both baseline and follow-up time points. Examinations were performed on a Philips Brilliance iCT 256 system using a standard pulmonary protocol: patients were scanned in the supine position during end-inspiratory breath-hold, covering the entire lung field from the apex to the costophrenic angles. The technical parameters were uniformly set as follows: tube voltage of 120 kV, automatic tube current modulation, a pitch of 0.758, a reconstruction matrix of 512×512, and a slice thickness of 1.00 mm.

Radiological feature extraction was independently performed by two experienced chest radiologists (F.Z. and W.L., with 6 and 8 years of experience in thoracic imaging, respectively). The readers were blinded to clinical data to ensure objectivity. The evaluated features included tumor location, maximum diameter, consolidation-to-tumor ratio (CTR), morphology, tumor-lung interface, presence of cavitation, lobulation, spiculation, air bronchogram, vascular convergence, and pleural retraction. Discrepancies were resolved through consensus after joint review and discussion.

Tumor segmentation and radiomic feature extraction

Tumor segmentation was performed using the 3D Slicer platform (version 5.2.1). To ensure comparability of features, all baseline and follow-up CT images were first resampled to an isotropic voxel size of 1.000×1.000×1.000 mm³. A radiologist (F.Z.) manually delineated the regions of interest (ROIs) slice-by-slice under lung window settings (window width, 1,500 HU; window level, −600 HU), carefully avoiding adjacent non-tumor structures such as great vessels, bronchi, and the chest wall. ROI delineation for baseline and follow-up scans was performed independently. The radiologist was blinded to recurrence status and the exact follow-up interval during segmentation.

A total of 1,037 radiomic features were extracted from each ROI using the PyRadiomics module within 3D Slicer. These comprised first-order statistics, shape features, and texture features derived from matrices including the Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level Size Zone Matrix (GLSZM), Gray-Level Dependence Matrix (GLDM), and Neighboring Gray-Tone Difference Matrix (NGTDM). To eliminate the influence of different scales, all features were normalized using min-max scaling.

Delta-radiomics features, which quantify the dynamic evolution of the tumor phenotype, were constructed based on a previously reported framework (19). The calculation formula was:

Indexdelta=(Ifollow-up−Ibaseline)(Tfollow-up−Tbaseline)[1]

Where I _baseline and I _follow-up represent the radiomic feature values at baseline and follow-up scans, respectively, and T _follow-up−T _baseline is the time interval. This index represents the linear slope of feature change over time, reflecting the dynamic evolution rate of the tumor imaging phenotype.

Reproducibility was assessed by another radiologist (W.L.), who independently segmented a random subset of 25 nodules. Features with an intraclass correlation coefficient (ICC) greater than 0.75 were considered to have good inter-observer agreement and were retained for subsequent analysis.

Feature selection and model construction

Eligible patients were randomly allocated to a training cohort and a validation cohort in a 7:3 ratio. To mitigate the data imbalance caused by the relatively low recurrence rate in early-stage lung cancer, the Synthetic Minority Oversampling Technique (SMOTE) (20) was applied to the training cohort, balancing the positive-to-negative case ratio to 1:1. Figure 1 illustrates the data distribution in the original dataset and after SMOTE balancing.

Figure 1 Distribution of the original dataset (A) and the SMOTE-balanced dataset (B). Green and orange dots represent the non-recurrence and recurrence groups, respectively. SMOTE, Synthetic Minority Oversampling Technique.

In the SMOTE-balanced training set, univariate logistic regression analysis was performed on clinical and radiological variables. Variables with statistical significance were subsequently entered into multivariate analysis to identify independent risk factors for constructing a clinical model. For radiomics and delta-radiomics features, selection was conducted in the training cohort. Features were initially filtered to retain those with high reproducibility (ICC >0.75), followed by significance analysis and correlation analysis. Subsequently, the Random Forest algorithm with 5-fold cross-validation was employed to select the ten most important features from both the radiomics and delta-radiomics pools. Based on the optimal algorithm results, a radiomics score (Rad-score) and a Delta Rad-score were constructed.

For fusion model construction, the Rad-score and Delta Rad-score were treated as two numerical predictors and combined with the selected clinical-radiological variables to form the final input feature matrix. Each patient therefore had a structured feature vector consisting of the selected clinical-radiological variables, the Rad-score, and the Delta Rad-score. No manual weighting was assigned to the clinical-radiological variables, Rad-score, or Delta Rad-score. Instead, the relative contribution of each predictor was learned automatically by the machine learning algorithm during model training. The validation cohort was not used for feature selection or model construction and was reserved for model evaluation. Multiple machine learning algorithms, including logistic regression, decision tree, Random Forest, support vector machine (SVM), and multilayer perceptron (MLP), were employed to develop fusion models. These models integrated the selected clinical-radiological features, the Rad-score, and the Delta Rad-score. The best-performing algorithm was chosen to construct the final fusion model. Calibration curves were used to assess the agreement between predicted probabilities and actual observed outcomes. The Shapley additive explanations (SHAP) method (21) was applied to interpret the machine learning model results, quantifying and visualizing feature importance.

Model performance was evaluated in both the training and validation cohorts using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. The DeLong test was used to compare the diagnostic performance of the fusion model against other models. Decision curve analysis (DCA) was performed to assess the clinical utility of the models.

The optimal cutoff value for the fusion model was determined using the Youden index in the training cohort, stratifying patients into high-risk and low-risk groups. Kaplan-Meier analysis was conducted to compare RFS and OS between the high-risk and low-risk groups, with differences assessed by the log-rank test. To demonstrate the incremental value of the fusion model over existing standards, patients were stratified into high-risk (T1c) and low-risk (T1a/T1b) groups according to the tumor-node-metastasis (TNM) classification (22). Kaplan-Meier analysis was then used to evaluate the prognostic performance of this traditional standard.

Statistical analysis

Statistical analyses were performed using SPSS (version 26.0), R (version 3.6.3), and Python (version 3.11.5). Continuous variables were presented as medians with interquartile ranges (IQRs) and compared using either the independent t-test or the Mann-Whitney U test, as appropriate. Categorical variables were expressed as frequencies and percentages and compared using the Chi-square test or Fisher’s exact test. A two-tailed P value <0.05 was considered statistically significant.

Results

Baseline characteristics of the study cohort

A total of 463 patients were ultimately included in the study, comprising 156 males and 307 females. The median age of all patients was 57 years. Baseline characteristics of the entire cohort are summarized in Table 1. The median interval between baseline and follow-up CT was 79 days (IQR, 58–113 days). The training and validation cohorts consisted of 324 and 139 patients, respectively. The median follow-up duration was 61 months. During the 3-year postoperative follow-up period, 23 patients (5.0%) experienced early disease recurrence (16 in the training cohort and 7 in the validation cohort), and 11 patients (2.4%) died. After SMOTE balancing, the training set comprised 616 patients, with 308 patients in each of the early recurrence and non-recurrence groups.

Variable selection and model development

Univariate analysis of the training cohort identified several variables significantly associated with early postoperative recurrence in stage IA LUAD, including sex, age, smoking history, history of COPD, family history of malignancy, high-grade histological components, location, lesion type, maximum tumor diameter, CTR, presence of cavitation, spiculation, air bronchogram sign, vascular convergence, and pleural retraction (all P<0.05). Multivariate analysis further revealed that sex [odds ratio (OR): 4.058; 95% confidence interval (CI): 1.480–11.129; P=0.007], history of COPD (OR: 11.715; 95% CI: 1.201–114.277; P=0.03), high-grade components (OR: 3.362; 95% CI: 1.621–6.972; P=0.001), location (OR: 2.989; 95% CI: 1.359–6.571; P=0.007), maximum tumor diameter (OR: 1.150; 95% CI: 1.062–1.247; P<0.001), CTR (OR: 13.611; 95% CI: 1.963–94.390; P=0.008), air bronchogram sign (OR: 5.635; 95% CI: 2.054–15.459; P<0.001), vascular convergence (OR: 0.169; 95% CI: 0.044–0.649; P=0.01), and pleural retraction (OR: 158.146; 95% CI: 26.537–942.455; P<0.001) were independently associated with early recurrence (Table 2). A clinical model was developed by integrating the above variables.

For each nodule, 1,037 radiomic features were initially extracted, of which 1,016 demonstrated good reproducibility (ICC >0.75). Significance analysis identified 873 radiomic features and 598 delta-radiomic features that were statistically associated with early recurrence of LUAD. After Pearson correlation analysis, 110 radiomic features and 153 delta-radiomic features were retained. Subsequently, the Random Forest algorithm with five-fold cross-validation was used to select the ten most important features from each pool, which were then used to construct a radiomics model (Rad-score) and a delta-radiomics model (Delta Rad-score), respectively. Random Forest feature importance analysis revealed the ranking contributions of predictive features in both the radiomics and delta-radiomics models (Figure 2).

Figure 2 Feature importance ranking in the radiomics and delta-radiomics models. (A) Top 10 most important features selected from the radiomics model based on the Random Forest algorithm with five-fold cross-validation. (B) Top 10 most important features selected from the delta-radiomics model.

Among the five machine learning models developed using the training set (Table 3), the model based on the Random Forest algorithm exhibited the best performance, achieving an area under the curve (AUC) of 0.881 in the validation cohort, and was therefore selected to construct the final fusion model. SHAP analysis (Figure 3) demonstrated that Rad-score and Delta Rad-score were the most important predictors of early recurrence in LUAD, with higher values of both scores positively correlated with early recurrence. Among clinical-radiological variables, pleural retraction and CTR showed substantial contributions.

Figure 3 SHAP analysis for model interpretation. (A) Summary plot showing the contribution of each feature to the final fusion model based on SHAP values. (B) SHAP dependence plot illustrating the impact of individual features on model predictions. Positive SHAP values are associated with an increased risk of early recurrence. COPD, chronic obstructive pulmonary disease; CTR, consolidation-to-tumor ratio; RD, radiomics; RF, Random Forest; SHAP, Shapley additive explanations.

Model performance

Table 4 summarizes the performance comparisons of the clinical model, radiomics model, delta-radiomics model, and fusion model. In the validation cohort, the fusion model demonstrated excellent discriminatory ability, achieving an AUC of 0.881 (95% CI: 0.799–0.952), which was statistically higher than that of the clinical model (AUC = 0.725; 95% CI: 0.565–0.876; P=0.01), but not significantly different from the radiomics model (AUC = 0.845; 95% CI: 0.723–0.942; P=0.99) and the delta-radiomics model (AUC = 0.827; 95% CI: 0.699–0.942; P=0.43) (Figure 4). Figure 4 shows that the fusion model achieved superior discrimination in the validation cohort (AUC 0.881), indicating its ability to correctly classify high-risk patients. In the validation set, the fusion model achieved an accuracy of 85.6%, a sensitivity of 71.4%, and a specificity of 86.4%. The calibration plot (Figure S1) showed good agreement between the recurrence probabilities predicted by the fusion model and the observed event frequencies, with a Brier score of 0.021. Quantitative calibration metrics were also calculated: the calibration intercept was 0.5205, the calibration slope was 2.1928, and the Hosmer-Lemeshow test indicated no significant evidence of poor fit (P=0.14). Calibration assessment was not performed in the validation cohort because the small number of recurrence events (n=7) would render the estimates statistically unstable. DCA (Figure 5) further confirmed that the fusion model provided the highest net clinical benefit across most threshold probabilities—meaning that using this model to guide postoperative surveillance or adjuvant therapy would result in more beneficial decisions and fewer unnecessary interventions compared to treating all patients as high-risk or low-risk. In the absence of SMOTE, the fusion model’s validation AUC dropped to 0.832 with a sensitivity of only14.3%, indicating that the model failed to identify most recurrence cases without resampling (Table S1).

Figure 4 ROC curves of different models. (A) ROC curves of the clinical model, radiomics model, delta-radiomics model, and fusion model in the training cohort. (B) ROC curves of the four models in the validation cohort. AUC, area under the ROC curve; ROC, receiver operating characteristic.

Figure 5 DCA of the models in the training and validation cohorts. (A) DCA in the training cohort comparing the net clinical benefit of the fusion model, clinical model, radiomics model (Rad-score), and delta-radiomics model (Delta Rad-score) across a range of threshold probabilities. The fusion model consistently provided the highest net benefit across all threshold probabilities. (B) DCA in the validation cohort, showing that the fusion model provided the highest net benefit across most threshold probabilities compared with the other models. DCA, decision curve analysis.

To evaluate the prognostic value of the fusion model for survival outcomes, we compared the fusion model with traditional TNM staging in the validation cohort. Based on the optimal cutoff value of the fusion model derived from the training cohort, patients were stratified into high-risk and low-risk groups. Kaplan-Meier curves (Figure 6) showed that the risk groups stratified by the fusion model were significantly associated with both RFS and OS (log-rank test, both P<0.001). In contrast, the traditional TNM classification stratified patients into T1c and T1a/T1b groups, which showed a statistically significant association with RFS (log-rank test, P=0.02) but not with OS (log-rank test, P=0.32).

Figure 6 Kaplan-Meier survival curves for risk stratification. (A) RFS and (B) OS curves comparing high-risk and low-risk groups stratified by the fusion model using the optimal cutoff determined by the Youden index. Patients in the high-risk group had significantly worse RFS and OS compared with those in the low-risk group (log-rank test, both P<0.001). (C) RFS and (D) OS curves stratified by traditional TNM classification (T1a/T1b vs. T1c). The TNM classification showed a significant association with RFS (log-rank test, P=0.02) but not with OS (log-rank test, P=0.32). OS, overall survival; RFS, recurrence-free survival; TNM, tumor-node-metastasis.

Discussion

In this study, we developed and validated a machine learning model based on serial follow-up CT images that integrates clinical-radiological features, preoperative radiomics features, and delta-radiomics features to predict the risk of early postoperative recurrence in stage IA LUAD. Our findings indicate that delta-radiomics features can be used to predict recurrence in LUAD. The machine learning model that incorporated clinical-radiological features, preoperative radiomics features, and delta-radiomics features demonstrated superior predictive performance compared to conventional clinical models and preoperative radiomics models alone. Furthermore, the fusion model outperformed traditional TNM classification in stratifying the risk of RFS and OS.

Several previous studies have employed machine learning models to predict postoperative recurrence risk in patients with stage IA LUAD. In a retrospective study involving 628 patients with stage IA LUAD, Chen et al. combined the International Association for the Study of Lung Cancer (IASLC) grading system with radiomics to predict postoperative outcomes, achieving an AUC of 0.692 for 3-year OS and 0.689 for RFS (23). In another study, Wang et al. integrated radiomics features with spread through air spaces (STAS) to predict early recurrence in stage I LUAD, achieving an AUC of 0.817 in the validation cohort (24). Compared with these studies, our approach demonstrated better predictive performance for postoperative recurrence. This improvement may be attributed to our delta-radiomics model, which captures the dynamic changes of pulmonary nodules during growth. The delta-radiomics model consistently outperformed conventional clinical models in predicting early recurrence of LUAD (AUC in the training cohort: 0.994 vs. 0.962; AUC in the validation cohort: 0.827 vs. 0.725) and achieved comparable performance to the preoperative radiomics model (AUC in the training cohort: 0.994 vs. 0.945; AUC in the validation cohort: 0.827 vs. 0.845). After fusing multi-dimensional information using machine learning algorithms, the fusion model exhibited the highest predictive performance (AUC of 0.881 in the validation cohort) and the greatest clinical net benefit, further highlighting the added value of delta-radiomics features in predicting early recurrence of LUAD. Moreover, the use of SMOTE to address data imbalance enhanced the model’s ability to identify the minority positive cases.

A major limitation of machine learning is the lack of interpretability, which may reduce clinicians’ willingness to adopt it. To address this, we applied SHAP analysis to interpret the machine learning model, quantifying the contribution and direction of clinical and radiomics features to the outcome. SHAP analysis based on the optimal Random Forest algorithm revealed that preoperative radiomics features and delta-radiomics features contributed the most to the model’s decision-making, and both were positively associated with the risk of early recurrence. This finding suggests that intratumoral microscopic heterogeneity and its temporal evolution play a more critical role in driving tumor recurrence than macroscopic imaging features. Notably, the strong predictive power of delta-radiomics may be related to the underlying tumor biological processes. The transition of tumors from an indolent to an invasive phenotype is often accompanied by pathophysiological changes such as angiogenesis, accelerated cell proliferation, and microenvironment remodeling, which manifest as dynamic alterations in radiomic features on CT images (13). Thus, delta-radiomics features may serve as imaging surrogate markers for these latent biological processes. Among the clinical-radiological features, pleural retraction and CTR showed moderate contributions. Previous studies have confirmed that pleural retraction is an independent risk factor for poor prognosis in early-stage non-small cell lung cancer (NSCLC). A large retrospective study by Xing et al. demonstrated that pleural retraction was independently associated with worse RFS and OS in patients with NSCLC (T ≤2 cm, N0, M0) [hazard ratio (HR) 6.57, 95% CI: 3.28–13.17, P<0.001 for RFS; HR 8.817, 95% CI: 3.51–22.16, P<0.001 for OS] (25). CTR, as an important imaging indicator of tumor invasiveness, has also been shown to have prognostic value. A study by Yun et al. based on 1,032 patients with clinical stage IA LUAD revealed a positive correlation between CTR and postoperative recurrence risk; the recurrence rate increased from 2.1% to 8.3% when CTR rose from 26–50% to 51–75% (26). In contrast, features such as maximum tumor diameter, high-grade components, air bronchogram sign, vascular convergence, location, history of COPD, and sex contributed only modestly. Those clinical features’ estimates should be interpreted with extreme caution and are presented only for exploratory purposes. Although SHAP analysis indicated that Rad-score and Delta Rad-score contributed most to model predictions, the fusion model outperformed both sub-models in net clinical benefit as shown by DCA. This suggests that clinical-radiological variables, despite their lower individual SHAP values, add incremental clinical utility by improving calibration and risk stratification at the population level.

From a clinical translation perspective, the fusion model developed in this study has potential practical utility. First, the model is based on preoperative serial CT imaging obtained during routine clinical practice and does not require additional invasive tests or biological specimens, making it highly accessible. Second, the model provides individualized recurrence probabilities and, through SHAP analysis, offers feature-level explanations that help clinicians understand the basis of the predictions, thereby enhancing trust in the model. Based on model-based risk stratification, high-risk patients could receive more intensive postoperative surveillance or be considered for adjuvant therapy, whereas low-risk patients could be spared unnecessary medical intervention. Moreover, compared with traditional TNM staging, our model demonstrated superior prognostic value in stratifying both RFS and OS, suggesting that it may serve as a valuable complement to the current staging system. We acknowledge that the small number of recurrence events (n=23) relative to the number of predictors raises concern for model overfitting. The near-perfect training AUC (0.997) likely reflects the model’s capacity to memorize patterns in the SMOTE-augmented training data, and the drop to 0.881 in validation underscores the need for cautious interpretation. These findings should be considered hypothesis-generating rather than definitive.

This study has several limitations. First, as a retrospective study restricted to an East Asian population, the generalizability of the findings may be limited. We acknowledge that requiring documented recurrence or at least 36 months of follow-up may introduce selection bias, as patients who died early without recurrence or were lost to follow-up are underrepresented. This is a recognized limitation of retrospective recurrence prediction studies. The results should therefore be interpreted with appropriate caution. Second, although the sample size was adequate, the number of outcome events was relatively small. We applied SMOTE to augment the number of positive cases to enhance model robustness. The clinical multivariable model included nine predictors with only 16 training events, resulting in unstable estimates. We acknowledge that the clinical model should be viewed as an exploratory baseline rather than a robust standalone predictor. Furthermore, calibration assessment was limited to the training cohort because the validation cohort contained only seven recurrence events, precluding reliable estimation of calibration slope, intercept, or Hosmer-Lemeshow test in the independent split. This study should be considered preliminary, and further large-scale, multicenter, prospective studies are warranted. Claims regarding superiority to TNM staging should be interpreted as hypothesis-generating, and definitive conclusions require external validation in independent patient populations. Third, the validation performance was derived from a single random split of the data. Although five-fold cross-validation was used during feature selection, the reported AUC may still be influenced by the specific partition. Finally, nodule volume data were obtained through semi-automated segmentation; the development of fully automated algorithms could reduce the burden on clinicians and facilitate clinical application.

Conclusions

The interpretable machine learning model integrating clinical characteristics, conventional radiological features, preoperative radiomics, and delta-radiomics provides a precise tool for individualized prognostic stratification in patients with stage IA LUAD and outperforms traditional clinical models and TNM staging systems. By incorporating SHAP analysis, our study reveals the central role of dynamic imaging features in predicting recurrence and provides imaging-based evidence for understanding the relationship between tumor evolution and recurrence risk. This model holds promise for assisting clinicians in identifying patients at high risk of recurrence and making individualized postoperative management decisions, ultimately improving patient outcomes.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0812/rc

Data Sharing Statement: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0812/dss

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0812/prf

Funding: This study was supported by Capital’s Funds for Health Improvement and Research (CFH 2022-2-5023).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0812/coif). The authors have no conflicts of interest to declare.

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 approved by the Ethics Committee of the Chinese PLA General Hospital (Approval No. S2025-035-01). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. In addition, informed consent was waived due to the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Siegel RL, Kratzer TB, Giaquinto AN, et al. Cancer statistics, 2025. CA Cancer J Clin 2025;75:10-45. [Crossref] [PubMed]
2. Yotsukura M, Asamura H, Motoi N, et al. Long-Term Prognosis of Patients With Resected Adenocarcinoma In Situ and Minimally Invasive Adenocarcinoma of the Lung. J Thorac Oncol 2021;16:1312-20. [Crossref] [PubMed]
3. Thai AA, Solomon BJ, Sequist LV, et al. Lung cancer. Lancet 2021;398:535-54. [Crossref] [PubMed]
4. Deng J, Zhao M, Wang T, et al. A modified T categorization for part-solid lesions in Chinese patients with clinical stage I Non-small cell lung cancer. Lung Cancer 2020;145:33-9. [Crossref] [PubMed]
5. Murakami S, Ito H, Tsubokawa N, et al. Prognostic value of the new IASLC/ATS/ERS classification of clinical stage IA lung adenocarcinoma. Lung Cancer 2015;90:199-204. [Crossref] [PubMed]
6. Carr SR, Wang H, Hudlikar R, et al. A Unique Gene Signature Predicting Recurrence Free Survival in Stage IA Lung Adenocarcinoma. J Thorac Cardiovasc Surg 2023;165:1554-64. [Crossref] [PubMed]
7. Yu S, You C, Yan R, et al. Establishment and validation of a nomogram model for predicting postoperative recurrence-free survival in stage IA3 lung adenocarcinoma: a retrospective cohort study. Transl Lung Cancer Res 2022;11:2275-88. [Crossref] [PubMed]
8. Hendriksen BS, Ortiz BA, Wigle DA, et al. Risk factors and survival associated with lung cancer recurrence after curative-intent surgery: beyond TNM pathological staging. J Thorac Dis 2025;17:10285-97. [Crossref] [PubMed]
9. Zhong D, Sidorenkov G, Jacobs C, et al. Lung Nodule Management in Low-Dose CT Screening for Lung Cancer: Lessons from the NELSON Trial. Radiology 2024;313:e240535. [Crossref] [PubMed]
10. Park S, Lee SM, Kim S, et al. Volume Doubling Times of Lung Adenocarcinomas: Correlation with Predominant Histologic Subtypes and Prognosis. Radiology 2020;295:703-12. [Crossref] [PubMed]
11. El-Jammal M, Chung C. Paving a Path to Clinical Impact with Radiomics: Enabling Reproducibility and Reach. Cancer Res 2025;85:2353-5. [Crossref] [PubMed]
12. Qi YJ, Su GH, You C, et al. Radiomics in breast cancer: Current advances and future directions. Cell Rep Med 2024;5:101719. [Crossref] [PubMed]
13. Nardone V, Reginelli A, Grassi R, et al. Delta radiomics: a systematic review. Radiol Med 2021;126:1571-83. [Crossref] [PubMed]
14. Wang Z, Wang F, Yang Y, et al. Development of a nomogram-based model incorporating radiomic features from follow-up longitudinal lung CT images to distinguish invasive adenocarcinoma from benign lesions: a retrospective study. BMC Pulm Med 2024;24:534. [Crossref] [PubMed]
15. Ma Y, Ma W, Xu X, et al. How Does the Delta-Radiomics Better Differentiate Pre-Invasive GGNs From Invasive GGNs? Front Oncol 2020;10:1017. [Crossref] [PubMed]
16. Han X, Wang Y, Jia X, et al. Predictive value of delta-radiomic features for prognosis of advanced non-small cell lung cancer patients undergoing immune checkpoint inhibitor therapy. Transl Lung Cancer Res 2024;13:1247-63. [Crossref] [PubMed]
17. Balbi M, Mazzaschi G, Leo L, et al. Longitudinal Changes of CT-radiomic and Systemic Inflammatory Features Predict Survival in Advanced Non-Small Cell Lung Cancer Patients Treated With Immune Checkpoint Inhibitors. J Thorac Imaging 2025;40:e0801. [Crossref] [PubMed]
18. Wu Z, Hua P, Chen X, et al. A study on the prediction of targeted therapy efficacy in advanced lung adenocarcinoma patients with EGFR mutations using CT-based delta-radiomics model. Front Med (Lausanne) 2025;12:1599206. [Crossref] [PubMed]
19. Jin W, Shen L, Tian Y, et al. Improving the prediction of Spreading Through Air Spaces (STAS) in primary lung cancer with a dynamic dual-delta hybrid machine learning model: a multicenter cohort study. Biomark Res 2023;11:102. [Crossref] [PubMed]
20. Dablain D, Krawczyk B, Chawla NV. DeepSMOTE: Fusing Deep Learning and SMOTE for Imbalanced Data. IEEE Trans Neural Netw Learn Syst 2023;34:6390-404. [Crossref] [PubMed]
21. Qi X, Wang S, Fang C, et al. Machine learning and SHAP value interpretation for predicting comorbidity of cardiovascular disease and cancer with dietary antioxidants. Redox Biol 2025;79:103470. [Crossref] [PubMed]
22. Goldstraw P, Chansky K, Crowley J, et al. The IASLC Lung Cancer Staging Project: Proposals for Revision of the TNM Stage Groupings in the Forthcoming (Eighth) Edition of the TNM Classification for Lung Cancer. J Thorac Oncol 2016;11:39-51. [Crossref] [PubMed]
23. Chen Y, Wu J, You J, et al. Integrating IASLC grading and radiomics for predicting postoperative outcomes in stage IA invasive lung adenocarcinoma. Med Phys 2024;51:6513-24. [Crossref] [PubMed]
24. Wang Y, Ding Y, Liu X, et al. Preoperative CT-based radiomics combined with tumour spread through air spaces can accurately predict early recurrence of stage I lung adenocarcinoma: a multicentre retrospective cohort study. Cancer Imaging 2023;23:83. [Crossref] [PubMed]
25. Xing Y, Zhou J, Yang Z, et al. Visible Pleural Retraction is a Prognostic Marker in Early-Stage Non-Small Cell Lung Cancer: A Retrospective Cohort Study. Ann Surg Oncol 2025; Epub ahead of print. [Crossref]
26. Yun JK, Kim JY, Ahn Y, et al. Predicting Recurrence after Sublobar Resection in Patients with Lung Adenocarcinoma Using Preoperative Chest CT Scans. Radiology 2024;313:e233244. [Crossref] [PubMed]