Development and validation of a radiomics-dosiomics model for predicting radiotherapy response in locally advanced, unresectable non-small cell lung cancer: a multi-center study Authors: Xun Wang, Xiufen Sun, Yueqin Chen, Guqing Zhang, Shucheng Ye, Aiping Zhang, Huipeng Yang, Zhanguo Sun, Shuang Ge Published: 2026 ABSTRACT AbstractEarly assessment of treatment response is essential for optimizing cancer management, as it allows timely interventions during the course of therapy, potentially improving cancer control and clinical outcomes. In this study, we aimed to develop and validate a machine learning model integrating radiomics, dosiomics, and clinical characteristics to predict response to radiotherapy in patients with locally advanced, unresectable non-small cell lung cancer (NSCLC). A total of 222 patients across multiple centers who received radiotherapy for NSCLC were enrolled and divided into training (n = 110), internal validation (n = 28), and external validation (n = 84) cohorts. Objective response rate (ORR) and progression-free survival (PFS) were predicted using models based on radiomics, dosiomics, and clinical characteristics. Model performance was evaluated using receiver operating characteristic (ROC) curves, DeLong test, decision curve analysis (DCA), Kaplan-Meier survival analysis, and Integrated Brier Score (IBS). The clinical models, CORR and CPFS (both based on planning target volume [PTV] and lymphocyte count) were compared with combined radiomics–dosiomics–clinical models (RDCORR and RDCPFS). For ORR prediction, RDCORR achieved AUCs of 0.901, 0.894, and 0.869 in the training, internal validation, and external validation cohorts, outperforming CORR (AUCs of 0.723, 0.606, and 0.723), with p < 0.05. DCA indicated that RDCORR outperformed CORR, providing a higher overall net benefit. For PFS prediction, RDCPFS yielded higher concordance indices (0.805, 0.730, and 0.743 in the training, internal validation, and external validation cohorts) than CPFS (0.679, 0.699, and 0.640, respectively). RDCPFS showed the lowest IBS across all cohorts (0.092, 0.107, and 0.093, respectively) compared with CPFS and the reference model, indicating better predictive accuracy. The combined model integrating radiomics, dosiomics, and clinical characteristics enhances the prediction of radiotherapy response in locally advanced, unresectable NSCLC, facilitating improved patient monitoring and more effective adjuvant clinical trial design.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-56201-5.Keywords: Radiomics, Dosiomics, Radiotherapy, Objective response rate, Progression-free survival, Non-small cell lung cancerSubject terms: Cancer, Oncology Introduction Non-small cell lung cancer (NSCLC) accounts for approximately 75%–80% of all primary lung cancer cases and remains a leading cause of cancer-related death worldwide, with a 5-year overall survival rate of around 20%1,2. Radiotherapy plays a crucial role in the management of NSCLC, especially for patients with locally advanced or unresectable diseases3–5. However, despite standardized treatment regimens, local failure occurs in up to 50% of patients, and insufficient local control remains a major cause of poor survival outcomes6,7. For patients with NSCLC who receive radiotherapy, attention should be paid to the radiation response of the tumor, such as near-term local control rate, objective response rate (ORR), and progression-free survival (PFS). Tumor radiosensitivity and treatment response vary considerably among individuals8. Early identification of patients who may benefit from radiotherapy enables timely adjustments to the treatment plan, thereby improving clinical efficacy. Conventional clinical factors, such as age, smoking status, tumor stage, or mutation status, hematological parameters, systemic immune-inflammation index, tumor consolidation and necrosis, etc., offer limited predictive value9–13, highlighting the necessity of more reliable biomarkers. Artificial intelligence (AI), particularly machine learning, has created new frontiers in oncologic imaging. Among these approaches, radiomics, which extracts high-dimensional quantitative features from medical images has shown great potential in the field of tumor imaging and has been widely applied to predict treatment responses and prognoses of various cancers, including lung cancer treated with chemotherapy or radiotherapy14–17. Radiotherapy efficacy is also closely related to dose distribution. Although features based on dose–volume histogram (DVH) have been used to predict PFS in locally advanced nasopharyngeal carcinoma, pathological complete response in esophageal cancer18,19, and radiation pneumonitis in NSCLC20, they capture only two-dimensional dose distribution. In contrast, the emerging dosiomics method can describe the three-dimensional dose distribution in terms of intensity, texture, shape and other dose characteristics, thereby providing more accurate and comprehensive spatial characterization of radiotherapy dose patterns21,22. In this study, we used multi-institution data to build and validate a machine learning model based on radiomics, dosiomics, and clinical characteristics to predict radiotherapy response in patients with NSCLC. Materials and methods Patient characteristics This was a retrospective multi-center study approved by the Ethics Committee of the Affiliated Hospital of Jining Medical University (Jining Medical University Ethics Committee for Human Research), China, and the requirement for informed consent was waived. This study was conducted according to the Declaration of Helsinki. This retrospective study included 138 patients with NSCLC who received intensity-modulated radiation therapy (IMRT) or volumetric-modulated arc therapy (VMAT) at the participating center between June 2016 and December 2022. Radiotherapy was delivered using conventional linear accelerators, including Varian Clinac 23EX and Varian Clinac IX systems (Varian Medical Systems, Palo Alto, CA, USA), and the XHLA1400 linear accelerator (Shinva Medical Instrument Co., Ltd., China). All treatments were delivered under free-breathing conditions, without the use of respiratory gating or breath-hold techniques. Patients were positioned and immobilized according to standard clinical protocols. The study cohort was randomly stratified into a training cohort (n = 110) and an internal validation cohort (n = 28) at a ratio of 8:2. Additionally, 84 patients from two other hospitals (49 from Hospital 1 and 35 from Hospital 2) were collected as the external validation cohort, with data between January 2019 and December 2022. Radiotherapy was delivered once daily at 1.8–3.0 Gy per fraction, five fractions per week, across 20–30 fractions, for a total dose of 40–60 Gy. In the PACIFIC era, following the establishment of consolidation immunotherapy after chemoradiotherapy in unresectable NSCLC, immunotherapy has become a key prognostic factor23; therefore, patients were stratified into immunotherapy (IO) and non-immunotherapy (non-IO) groups according to treatment regimens. The inclusion criteria were as follows: (i) Pathological diagnosis of NSCLC. (ii) Patients with locally advanced, unresectable NSCLC—defined as those not amenable to surgical resection due to tumor extent (e.g., T4 disease, N3 nodal involvement, or invasion of critical mediastinal structures). All patients had an Eastern Cooperative Oncology Group performance status (ECOG PS) of ≤ 2 to ensure tolerance to definitive radiotherapy. (iii) Patients who completed radiotherapy and had comprehensive clinical data. (iv) All patients underwent chest CT examination at 1, 3, 6, and 12 months after radiotherapy, or had the endpoint. The exclusion criteria were as follows: (i) Patients whose primary lesion was surgically removed and thus did not undergo radiotherapy. (ii) Recurrent chest radiotherapy or patients treated with stereotactic body radiation therapy. (iii) Patients with other concomitant malignancies. (iv) Patients with mediastinal primary tumors. (v) Patients who did not complete radiotherapy due to poor tolerance, severe complications, or personal/economic reasons, and thus lacked complete follow-up data. The primary endpoint of this study was PFS, and the secondary endpoint was the ORR of the primary tumor, assessed at 3 months after completion of radiotherapy. PFS was defined as the time from the initiation of radiotherapy to the first occurrence of disease progression or death from any cause, whichever occurred first. Tumor response and disease progression were assessed according to the Response Evaluation Criteria in Solid Tumors (RECIST) version 1.1. Imaging evaluations were performed at 3 months after radiotherapy and subsequently every 3 months within the first year of follow-up. All evaluations were independently performed by one experienced radiologist and one experienced oncologist, both blinded to the patients’ clinical information and model outputs. Discrepancies between the two evaluators were resolved by consensus. The flow chart of this study is shown in Fig. 1. The flow chart of this study. The clinical characteristics of the patients were recorded, including age, gender, smoking history, pathological type, T stage, treatment regimens, and baseline hematological parameters (obtained within 2 weeks before radiotherapy initiation), including lymphocyte count, neutrophil count, monocyte count, neutrophil-to-lymphocyte ratio (NLR), and lymphocyte-to-monocyte ratio (LMR). Radiotherapy-related parameters were also collected, including biological effective dose (BED), planning target volume (PTV), mean dose to PTV (PTV-Dmean), dose received by 98% of the PTV (PTV-D98). The BED, which accounts for both the total dose and fractionation schedule, was calculated using the linear-quadratic model to better reflect the biological effect of the delivered radiation. Univariable and multivariable regression analyses were performed to identify independent clinical predictors of ORR and PFS, using logistic regression for ORR and Cox proportional hazards regression for PFS. Variables with P < 0.05 in the univariable analysis were included in the corresponding multivariable models. The clinical models (CORR and CPFS) were constructed based on the independent predictors identified in the multivariable analyses. CT acquisition and image segmentation All patients underwent tumour localization using large-bore CT simulators (PHILIPS Brilliance CT Big Bore, and Canon Aquilion Prime TSX-303 A). Scanning was performed with a tube voltage of 120 kV. Tube current was set at approximately 250mAs according to institutional protocols, and automatic exposure control was applied in some cases depending on the scanner configuration. The helical pitch was 1, and slice thickness was 5 mm. CT images were reconstructed using a standard soft-tissue convolution kernel with a matrix size of 512 × 512, resulting in an in-plane voxel size of approximately 0.98 × 0.98 mm². Experienced radiation oncologists delineated the gross tumor volume (GTV) of the primary lung tumor using the Eclipse or Monaco treatment planning system (TPS), based on contrast-enhanced CT images. The clinical target volume (CTV) included the GTV and regions at risk of microscopic tumor spread, with expansion margins determined according to tumor location, anatomical boundaries, and institutional protocols. The planning target volume (PTV) was generated by further expanding the CTV to account for setup uncertainties and internal organ motion, with the margin size tailored to individual patient conditions and immobilization techniques. The prescribed dose was designed to cover at least 95% of PTV. Features extraction and selection, models building In this study, different regions of interest (ROIs) were adopted for radiomics and dosiomics feature extraction, considering their distinct biological and clinical implications. Radiomics features were extracted from GTV to ensure that the analysis primarily reflected intratumoral heterogeneity, which is widely recognized as the standard approach in radiomics research. In contrast, dosiomics features were derived from the PTV, as dose distribution in radiotherapy is planned and prescribed based on this volume, making PTV-based dosiomics more representative of the actual delivered dose. No additional manual segmentation was performed. The planning CT images, ROIs, and corresponding dosimetric images were exported from TPS and imported into 3D Slicer (version 5.1.0, https://www.slicer.org). All radiomics and dosiomics features were then extracted in compliance with the Image Biomarker Standardization Initiative (IBSI) using the SlicerRadiomics extension based on the Pyradiomics library. The original dose calculation grid size in the TPS was 2.5 mm for all patients. To ensure spatial consistency among CT images, dose distributions, and ROIs, both CT and dosimetric images were resampled to an isotropic voxel size of 1 mm prior to feature extraction using nearest-neighbor interpolation. Voxel intensities were discretized with a fixed bin width of 25 HU for CT and 0.1 Gy for dose distributions. These preprocessing steps partially mitigated variability among different scanners. Dosiomics features were extracted from the 3D dose maps within the corresponding ROI (PTV) using the same feature categories and extraction settings as those used for CT radiomics. A total of 1,834 radiomics and 1,834 dosiomics features were extracted from each ROI, including 216 first-order features, 876 texture features, 14 shape features and 728 wavelet features, respectively. To enhance feature robustness, all features were Z-score normalized, and highly correlated features (|r| > 0.9) were removed. Feature selection was performed in two steps: first, features were ranked using the minimum redundancy maximum relevance (mRMR) algorithm, followed by least absolute shrinkage and selection operator (LASSO) regression for final feature selection. A logistic regression (LR) model was fitted using the selected features to obtain regression coefficients for each feature, and the radiomics signature (RS) and dosiomics signature (DS) for ORR (RSORR, DSORR) were calculated for each patient. For PFS analysis, radiomics and dosiomics features were selected using the LASSO-Cox model with tenfold cross-validation, and the corresponding RS and DS (RSPFS, DSPFS) were calculated for each patient. The signatures were calculated as follows: where Xi represents the feature value, βi the corresponding regression coefficient, and β0 the intercept. RSORR and DSORR were used to construct the radiomics model (RORR) and dosiomics model (DORR). Finally, a combined radiomics–dosiomics–clinical model (RDCORR) was developed by incorporating RSORR, DSORR, and independent clinical characteristics into a multivariable logistic regression model, and a nomogram was drawn to visualize the predicted probability of ORR. Similarly, a combined radiomics-dosiomics-clinical model (RDCPFS) was developed by incorporating RSPFS, DSPFS, and independent clinical characteristics, and a nomogram was drawn to estimate the risk of disease progression. Performance comparison between different models For ORR, model performance (CORR, RORR, DORR, and RDCORR) was evaluated using the area under the receiver operating characteristic curve (AUC). The optimum cut-off value was determined using the Youden index, and the sensitivity, specificity, and accuracy were calculated in the training, internal validation, and external validation cohorts. DeLong test was used to compare the AUCs among models. Decision curve analysis (DCA) was performed to assess the clinical utility. For PFS, risk scores (linear predictors) were calculated from the Cox proportional hazards models using the “predict” function in R software. The optimal cutoff values for CPFS and RDCPFS were determined using the “survcutpoint” function, and patients were subsequently stratified into high-risk and low-risk groups. Survival differences between groups were assessed using the log-rank test. Model discrimination was evaluated using the concordance index (C-index) derived from the Cox models. Additionally, predictive accuracy over time was assessed using the Integrated Brier Score (IBS) integrated up to 1 year, with lower IBS values indicating better model performance. Statistics Statistical analyses were performed using R software (version 4.4.1, https://www.r-project.org). A two-sided P value < 0.05 was considered statistically significant. Results Clinical characteristics The baseline clinical characteristics and corresponding statistical comparisons for the ORR and PFS cohorts are presented in Supplementary Table S1 and S2. The ORR was 72.7% (80/110), 71.4% (20/28), and 75.0% (63/84) in the training, internal validation, and external validation cohorts, respectively. Univariable logistic regression analysis identified PTV, PTV-Dmean, PTV-D98, BED, and lymphocyte count as significant predictors of ORR in the training cohort. Multivariable logistic regression analysis identified PTV (p = 0.024) and lymphocyte count (p = 0.012) as independent predictors of ORR (Table 1). The median follow-up time was 398 days (95% CI: 365–553 days) for the overall cohort, and 513 days (95% CI: 420–707 days), 657 days (95% CI: 398 days to not reached [NR]), and 340 days (95% CI: 325–365 days) for the training, internal validation, and external validation cohorts, respectively. At the last follow-up, 53/110 (48.2%), 13/28 (46.4%), and 39/84 (46.4%) patients remained progression-free in the training, internal validation, and external validation cohorts. Univariable Cox regression analysis identified PTV, lymphocyte count, and NLR as significant predictors of PFS in the training cohort. Multivariable Cox regression analysis identified PTV (p = 0.001), and lymphocyte count (p = 0.004) as independent predictors of PFS (Table 2). Table 1Regression analyses of ORR in training cohort.Clinical characteristicsTraining cohort(n = 110)ULRMLR p OR (95%CI) p OR (95%CI)Responders (n = 80)Non-responders (n = 30)PTV (cm3)317.9 ± 120.49385.0 ± 137.760.0181.004 (1.001–1.008)0.0241.005 (1.001–1.009)Lymphocytea1.8 ± 0.621.5 ± 0.460.0150.342 (0.144–0.813)0.0120.283 (0.106–0.758)BED (Gy)69.9 ± 4.9266.6 ± 5.610.0060.898 (0.831–0.969)0.1500.804 (0.598–1.082)PTV-Dmean (Gy)60.7 ± 4.0258.0 ± 4.850.0070.999 (0.998-1.000)0.5220.999 (0.996–1.002)PTV-D98 (Gy)55.0 ± 4.8852.8 ± 4.710.0370.999 (0.998-1.000)0.0631.003 (1.000-1.005)ULR Univariable logistic regression, MLR multivariable logistic regression.aBaseline values before radiotherapy, ×10^9/L. Regression analyses of ORR in training cohort. ULR Univariable logistic regression, MLR multivariable logistic regression. aBaseline values before radiotherapy, ×10^9/L. Table 2Regression analyses of PFS in training cohort.Clinical characteristicsTraining cohort(n = 110)UCRMCR p HR (95%CI) p HR (95%CI)Progressors (n = 57)Non-progressors (n = 53)PTV (cm3)365.3 ± 125.63302.8 ± 127.070.0061.003 (1.001–1.005)0.0011.004 (1.002–1.016)Lymphocyte1.6 ± 0.551.8 ± 0.520.0050.443 (0.252–0.781)0.0040.399 (0.215–0.742)NLR2.5 ± 1.482.1 ± 1.470.0471.163 (1.002–1.349)0.3801.081 (0.909–1.285)UCR Univariable Cox regression, MCR multivariable Cox regression. Regression analyses of PFS in training cohort. UCR Univariable Cox regression, MCR multivariable Cox regression. Models building and validation PTV and lymphocyte count were used to construct the CORR. The AUCs of the CORR for predicting the ORR in NSCLC were 0.723 (95% confidence interval [CI]: 0.611–0.835), 0.606 (95% CI: 0.347–0.865) and 0.723 (95% CI: 0.595–0.850) in the training, internal validation, and external validation cohorts, respectively (Fig. 2A-C; Table 3). The corresponding sensitivity, specificity, and accuracy are summarized in Table 4. Predictive performance of models for objective response rate (ORR) after radiotherapy in NSCLC. (A–C) Receiver operating characteristic (ROC) curves of the models in the training, internal validation, and external validation cohorts, respectively. (D–F) Decision curve analysis (DCA) of the models in the corresponding cohorts. Table 3AUCs of different models for predicting ORR.Training cohortInternal validation cohortExternal validation cohortAUC95%CIAUC95%CIAUC95%CICORR0.7230.611–0.8350.6060.347–0.8650.7230.595–0.850RORR0.8840.814–0.9540.7940.559–0.9530.7980.696–0.900DORR0.7560.664–0.8480.7500.517–0.9830.7660.639–0.894RDCORR0.9010.844–0.9590.8940.766–1.0000.8690.783–0.955 AUCs of different models for predicting ORR. Table 4Predictive performance of ORR models.Training cohortInternal validation cohortExternal validation cohortSensitivitySpecificityAccuracySensitivitySpecificityAccuracysensitivityspecificityaccuracyCORR0.7000.7630.7450.6250.6500.6430.6190.7780.738RORR0.9330.7130.7730.8750.7000.7500.8100.7620.774DORR0.8670.6130.6820.7500.8500.8210.8100.7300.750RDCORR0.9000.7880.8180.7500.9000.8570.7140.8730.833 Predictive performance of ORR models. From 1,834 radiomics features and 1,834 dosiomics features, 19 radiomics features and 8 dosiomics features (Supplementary A1-2) were selected for ORR prediction. The ROC curves and the corresponding AUCs, sensitivity, specificity, and accuracy of the RORR, DORR, and RDCORR for predicting ORR after radiotherapy in NSCLC are presented in Fig. 2A–C, and Tables 3 and 4, respectively. According to the DeLong test, the performance of the RDCORR was significantly better than that of the CORR in the training, internal validation, and external cohorts (p = 0.0003, 0.032, and 0.018, respectively). DCA demonstrated that the RDCORR provided a higher overall net benefit than the CORR (Fig. 2D-F). A nomogram combining PTV, lymphocyte count, RSORR and DSORR for predicting ORR is shown in Fig. 3A. Nomograms for ORR and PFS. (A) Nomogram predicting ORR. (B) Nomogram predicting PFS. PTV and lymphocyte count were used to establish CPFS. From 1,834 radiomics features and 1,834 dosiomics features, 20 radiomics features and 14 dosiomics features (Supplementary A3-4) were selected for PFS prediction. RDCPFS was developed by combining PTV, lymphocyte count, RSPFS and DSPFS. Kaplan–Meier survival curves for PFS in NSCLC patients after radiotherapy, stratified by CPFS and RDCPFS are shown in Fig. 4A–F. Significant differences in PFS were observed between the low- and high-risk groups identified by the two models (log-rank test, p < 0.05). The C-index values of CPFS and RDCPFS were 0.679 and 0.805 in the training cohort, 0.699 and 0.730 in the internal validation cohort, and 0.640 and 0.743 in the external validation cohort, respectively. RDCPFS showed the lowest IBS across all cohorts (0.092, 0.107, 0.093 for the training, internal validation, and external validation cohorts, respectively) compared with CPFS and the reference model, indicating superior predictive performance (Supplementary Table S3). A nomogram combining PTV, lymphocyte count, RSPFS and DSPFS for predicting PFS is shown in Fig. 3B. Kaplan–Meier survival curves for progression-free survival (PFS) stratified by CPFS and RDCPFS in NSCLC. (A) CPFS in the training cohort. (B) RDCPFS in the training cohort. (C) CPFS in the internal validation cohort. (D) RDCPFS in the internal validation cohort. (E) CPFS in the external validation cohort. (F) RDCPFS in the external validation cohort. Discussion In this multicenter study, we evaluated the ability of a combined model incorporating radiomics, dosiomics, and independent clinical characteristics to predict radiotherapy-associated ORR and PFS in patients with locally advanced unresectable NSCLC. For ORR prediction, the combined model (RDCORR) achieved AUCs of 0.901, 0.894, and 0.869 in the training, internal validation, and external validation cohorts, respectively, which were higher than those of the CORR (AUCs of 0.723, 0.606, and 0.723, respectively). For PFS prediction, the combined model (RDCPFS) achieved C-indices of 0.805, 0.730, and 0.743 in the training, internal validation, and external validation cohorts, which were also higher than those of the CPFS (0.679, 0.699, and 0.640, respectively). These results demonstrated that the combined models outperformed the clinical models alone, indicating the incremental value of radiomics and dosiomics features beyond traditional clinical factors in predicting radiotherapy outcomes in NSCLC. Our study showed that PTV and lymphocyte count were independent predictors of ORR and PFS, emphasizing the importance of tumor burden and immune status in treatment outcomes. Consistent with previous studies, larger tumor volume has been associated with poorer local control and survival24–27. This may be attributed to increased tumor proliferation, hypoxic microenvironments, and potential dose heterogeneity in larger tumors, which can reduce treatment effectiveness26. Higher lymphocyte count was associated with improved ORR and PFS, which is consistent with previous studies showing that pretreatment lymphocytopenia is linked to higher mortality across multiple cancer types28–30. Higher lymphocyte levels may reflect a stronger antitumor immune response, supporting treatment efficacy and highlighting the role of the immune system in cancer progression28. In addition to these clinical factors, treatment-related variables may also influence outcomes. Systemic treatment regimens varied among patients, and given the prognostic relevance of immunotherapy in the PACIFIC trial era, patients were stratified into IO and non-IO groups. Subgroup analysis showed no statistically significant association between treatment strategy and ORR or PFS. However, this finding should be interpreted with caution because of the imbalance in sample size between groups and the relatively limited number of patients receiving immunotherapy, particularly during the earlier years of the study period (2016–2022), when immunotherapy was less commonly used. Radiation dose, represented by BED, has been associated with improved local control and survival in previous study31. In our multivariable analysis for ORR, only PTV and lymphocyte count remained independent predictors, whereas BED and other dosimetric parameters lost statistical significance, likely due to collinearity and shared information among radiation dose distribution metrics. Similarly, in the PFS analysis, although PTV, lymphocyte count, and NLR were significant in univariable analysis, only PTV and lymphocyte count remained independent predictors in multivariable model. NLR has been reported as a strong prognostic indicator in cancer32; however, its effect in this study may overlap with that of other immune-related variables, leading to its exclusion from the multivariable model. In our study, the integration of radiomics and dosiomics features improved the predictive performance for ORR and PFS. The combined models (RDCORR and RDCPFS) achieved better discrimination than conventional clinical models (CORR and CPFS), with higher AUCs and C-index values across training, internal validation, and external validation cohorts. Radiomics has been widely applied in AI-based models to predict oncologic outcomes such as local control and PFS by capturing tumor shape, intensity, and texture features that reflect intratumoral heterogeneity17,33. Dosiomics characterizes radiation dose and spatial dose distributions during radiotherapy and describes dose heterogeneity within tumors. Studies have shown that dosiomics can predict radiotherapy efficacy and survival outcomes in different tumors22,34–36. Radiomics and dosiomics provide complementary information reflecting tumor phenotype and dose distribution characteristics, respectively. In our study, for ORR prediction, although the radiomics-only (RORR) and dosiomics-only (DORR) models exhibited comparable performance, the integrated model (RDCORR) achieved relatively higher AUC values across the training, internal validation, and external validation cohorts. For PFS prediction, RDCPFS was further evaluated using the 1-year IBS and the C-index. RDCPFS showed relatively low IBS values across cohorts and maintained acceptable discriminative performance (C-index > 0.7), suggesting a potentially useful predictive signal for PFS. These suggest that combining complementary information from radiomics, dosiomics, and clinical characteristics may provide additional predictive value; however, the improvement in model performance was modest, and the findings should be interpreted with caution given the limited sample size and wide confidence intervals in the internal validation cohort. Notably, certain geometric features, such as LeastAxisLength, may help explain the observed predictive signal. Previous studies have shown that shape-related descriptors may be associated with treatment response37, suggesting that such features—whether derived from imaging or dose distributions—may reflect underlying tumor aggressiveness beyond simple volumetric measures. Additionally, there is overlap between the features predicting ORR and PFS, suggesting that these two prognostic endpoints may be influenced by similar biological mechanisms. A previous review summarized the correlation and regression analyses between different tumor response outcomes and PFS, primarily focusing on the relationship between ORR and PFS or OS38. Regarding the association between ORR and PFS, most reported correlation coefficients for absolute outcome values were > 0.60 (range − 0.72 to 0.96). For the correlation between treatment effects on ORR and PFS, most regression R2 values were > 0.40. One possible explanation is that certain features, such as gene expression and tumor microenvironment, may affect both tumor response and growth dynamics, thereby affecting ORR and PFS. While the results are encouraging and suggest that integrating radiomics, dosiomics, and clinical features may improve model performance, several limitations should be acknowledged. The retrospective design and relatively small sample size in a high-dimensional setting may have reduced statistical power and introduced potential confounding; therefore, the robustness of the models should be interpreted with caution. Prospective multicenter studies with larger and more homogeneous cohorts are warranted to further validate the predictive performance and clinical applicability of these models. GTV and PTV delineation were performed manually, which is time-consuming and subject to inter-observer variability. The use of semi-automatic or automatic delineation methods in future studies could improve efficiency and consistency while enhancing model generalizability. Another limitation is that radiomics analysis was restricted to the GTV, without incorporating peritumoral regions. Emerging evidence suggests that peritumoral areas may provide complementary biological information related to tumor–microenvironment interactions, which may differ from intratumoral features30. Future studies incorporating multiple regions of interest, such as the GTV, peritumoral regions, and PTV, may better characterize tumor heterogeneity and treatment-related effects. In particular, further investigation into the distinct and temporal roles of intra- and peritumoral radiomics features during radiotherapy may improve the understanding of tumor behavior and potentially enhance prognostic modeling in NSCLC39. In addition, although standardized CT and dose acquisition protocols were used, including a uniform 2.5 mm dose grid, variations in exposure control and scanner settings may still have existed. CT and dose maps were resampled to a 1 mm isotropic grid for spatial alignment; however, interpolation during upsampling may introduce smoothing effects and influence feature stability. Conclusions This study suggests that combining radiomics and dosiomics features can provide complementary information beyond traditional clinical parameters in patients with locally advanced unresectable NSCLC. By improving predictive accuracy, these models may enhance risk stratification and facilitate better clinical decision support. Future research should focus on further refining these models and validating their applicability in real-world clinical settings. Supplementary Information Below is the link to the electronic supplementary material. Please delete the file of Supplementary Material 1(MOESM1), as it is a manuscript with revision marks. Supplementary Material 2 Supplementary Material 3 Acknowledgements Some of our experiments were carried out on the Onekey AI platform. Thank Onekey AI and its developers’ help in this scientific research work. Abbreviations Clinical model Radiomics model Dosiomics model Radiomics-dosiomics-clinical model Radiomics signature Dosiomics signature Neutrophil-to-lymphocyte ratio Lymphocyte-to-monocyte ratio Objective response rate Progression-free survival Area under the curve Receiver operating characteristic Region of interest Author contributions X.W., X.-F. S., Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Supervision, Writing – original draft, Writing – review & editing. G.-Q. Z., Data curation, Investigation, Supervision. Y.-Q. C., S.-C. Y., Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Supervision. A.-P. Z., H.-P. Y., Data curation, Investigation. Z.-G. S., S.G., Supervision. Project administration, Supervision, Funding acquisition, Resources. All authors reviewed the manuscript. Funding This work was supported by the Jining City of Science and Technology Bureau (Key research and development project: 2025YXNS018), Shandong Province Medical Health Science and Technology Development Plan Project (202409030221), Bethune · Excellence in Medicine — Research Program for Advanced Solid Tumors (STLKY0022). Data availability The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Ethical approvaland informed consent The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee (2023-12-C013) of the Affiliated Hospital of Jining Medical University, Jining Cancer Hospital, and Jining First People’s Hospital. Informed consent was waived due to the retrospective nature of this study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xun Wang, Xiufen Sun, Zhanguo Sun and Shuang Ge contributed equally to this work. Contributor Information Zhanguo Sun, Email: yingxiangszg@163.com. Shuang Ge, Email: geshuang616@163.com. References Associated Data Data Availability Statement The datasets used and/or analyzed during the current study available from the corresponding author on reasonable request.