Development of a multifactorial prediction model for symptomatic radiation pneumonitis in lung cancer patients undergoing stereotactic body radiotherapy Authors: Xiong Yang, Quncai Wang, Changli Ruan, Luzhou Wang, Hongyun Gong, David Huang Published: 2026 ABSTRACT AbstractPurposeSymptomatic radiation pneumonitis (SRP) remains a clinically significant toxicity following lung stereotactic body radiotherapy (SBRT). However, existing prediction models have seen limited clinical adoption, as they tend to be either too complex for routine application or too simplistic for reliable risk stratification. This study aimed to develop a robust prediction model by integrating key dosimetric, volumetric, and clinical variables, and to translate this model into a practical risk score to support individualized pretreatment assessment.Materials and methodsThis retrospective analysis included 120 patients with primary lung cancer treated with SBRT between January 2019 and December 2024. A total of 24 demographic, clinical, volumetric, and dosimetric parameters were evaluated. The primary endpoint was the development of SRP. A three-step statistical framework was employed for model development. First, candidate predictors were identified using univariate analysis (Mann-Whitney U test for continuous variables and Chi-square test for categorical variables), with statistical significance defined as p < 0.05. Second, this set of variables was refined through multicollinearity diagnosis using Spearman correlation and variance inflation factor (VIF) analysis to ensure predictor independence. Third, an optimal multivariate logistic regression model was constructed, incorporating the synthetic minority over-sampling technique (SMOTE) to address class imbalance, and validated using 5-fold cross-validation and an independent test set. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), accuracy, calibration plots, and the Brier score. A practical risk stratification system was subsequently developed based on the model’s predicted probabilities.ResultsThe overall incidence of SRP was 43.3% (52/120). Univariate analysis identified 12 variables significantly associated with SRP (p < 0.05), encompassing all dose-volume histogram (DVH) parameters (V5-V30, MLD), target volumes, and other prior therapy (OT). Following multicollinearity analysis, four independent candidates—lung V5 (%), Lung-to-PTV Ratio (LPR), OT, and treatment planning system (TPS)—were retained for multivariate modeling. Multivariate logistic regression confirmed V5 (OR = 1.113 per %; p < 0.01), LPR (OR = 0.959 per unit; p < 0.01), and OT (OR = 2.672; p < 0.01) as significant independent predictors, with LPR exhibiting the strongest effect. The final model demonstrated excellent discriminative performance, achieving a cross-validated AUC of 0.875 and a test set AUC of 0.921, along with good calibration (Brier score = 0.13). Based on predicted probabilities, patients were stratified into three distinct risk groups (low: P < 0.3; medium: 0.3 ≤ P ≤ 0.7; high: P > 0.7), enabling effective differentiation of SRP risk for clinical application.ConclusionThe integration of a low-dose parameter (V5), a novel volumetric ratio (LPR), and other prior treatment history (OT) yields a robust and parsimonious model for predicting SRP after lung SBRT. The derived three-tiered risk score offers a practical, interpretable tool that may inform individualized patient counseling and assist clinicians during treatment planning, although further prospective validation is needed.Keywords: Radiation pneumonitis (RP), Stereotactic body radiotherapy (SBRT), Lung cancer, Lung-to-PTV ratio (LPR), Risk stratification Introduction Stereotactic body radiation therapy (SBRT) is an advanced technique that delivers ablative radiation doses to tumors with remarkable precision while sparing adjacent normal tissues [1–3]. It has revolutionized the management of early-stage, medically inoperable non-small cell lung cancer (NSCLC) and pulmonary oligo-metastases, achieving local control rates exceeding 90% with potentially better survival outcomes compared to conventional fractionated radiation therapy (CFRT) [4–6]. However, the unique radiobiology of hypofractionation—characterized by heightened vascular, stromal, and immunomodulatory effects—also predisposes patients to distinct toxicity patterns. Among these, radiation pneumonitis (RP) remains the most common clinically relevant complication [1, 7, 8]. The Common Terminology Criteria for Adverse Events (CTCAE) Version 5.0 grades RP from 1 to 5, defining symptomatic RP (SRP) as grade 2 or higher [9]. Symptoms such as cough, dyspnea, and chest pain can significantly impair quality of life and pose life-threatening risks [10–13]. This underscores the critical need to proactively predict SRP risk, which will facilitate timely treatment adjustment and improve post-treatment management. Therefore, identifying robust predictive factors enables individualized risk estimation before SBRT, allowing pre-emptive interventions to enhance efficacy and reduce complications. For CFRT, the mean lung dose (MLD) and V20 are established predictors of RP [14–16]. However, their relevance to SBRT—with its distinct, non-uniform dose distributions—is less clear. Some studies affirm their utility after SBRT, while others propose that low-dose metrics (e.g., V5, V10) may be more predictive, possibly due to heightened parenchymal sensitivity in hypofractionation [17]. Furthermore, clinical factors such as pre-existing pulmonary fibrosis, central tumor location, and prior systemic therapy are recognized to modulate SRP risk [18–20]. Currently, predictive models often succumb to a dichotomy: they are either overly complex, incorporating numerous correlated variables that invite overfitting, or oversimplified, relying on a single parameter of limited predictive power [21–24]. Moreover, most remain confined to academic literature without being translated into clinically practical tools. Consequently, their adoption is hampered in routine practice—such as during treatment planning optimization or patient consultation—which ultimately hinders real-time clinical decision-making. To address these gaps, this study was designed to integrate methodological rigor with clinical applicability. It hypothesized that integrating key dosimetric, novel volumetric, and readily available clinical factors would best predict SRP following lung SBRT. The primary objectives were to: (1) identify robust predictors while controlling for multicollinearity; (2) develop a parsimonious, interpretable multivariate logistic regression model; and (3) translate the final model into a practical clinical risk score to support routine decision-making. Materials and methods Study population This retrospective cohort study was conducted in accordance with the Declaration of Helsinki and approved by the Clinical Research Ethics Committee of Renmin Hospital of Wuhan University (approval number: 2024 K-K037), with a waiver for informed consent due to the retrospective nature. All patients were screened with primary or metastatic lung cancer treated with definitive SBRT between January 2019 and December 2024. Inclusion criteria were: (a) pathologically confirmed lung cancer; (b) completion of prescribed SBRT with a dose equal to or greater than 5 Gy per fraction with no prior or subsequent radiation therapy (RT); (c) availability of complete dosimetric data from the treatment planning system (TPS); and (d) minimum clinical and radiological follow-up of 12 months post-SBRT to adequately capture RP events. No patients were excluded based on pre-existing lung disease. Among 132 lung cancer patients who received SBRT, 120 were eligible for analysis. Radiotherapy treatment All patients underwent four-dimensional computed tomography (4D-CT) simulation in the supine position with appropriate immobilization. Gross tumor volumes (GTVs) were delineated on each individual phase of the 4D-CT dataset. These ten phase-specific GTVs were then combined to generate the internal target volume (ITV), which was subsequently stored on the average intensity projection (AVG) images for planning purposes. A 5-mm isotropic expansion was applied to the ITV to create the planning target volume (PTV) to account for setup uncertainties. All delineation and volume generation steps described above were performed by radiation oncologists using contouring software MIM (Version 7.3.7, MIM Software Inc., Cleveland, OH). Treatment plans were designed using either the Monaco (Elekta AB, version 5.40.04) or Pinnacle (Philips Radiation Oncology Systems, version 9.10) TPS, employing volumetric modulated arc therapy (VMAT) or intensity-modulated radiotherapy (IMRT) techniques with 6 MV photon beams. Prescriptions varied among oncologists, with fractions ranging from 3 to 14 and single doses ranging from 5 to 10 Gy, and the median biological effective dose (BED₁₀) for PTV was 96.5 Gy. All treatments were delivered with daily image-guided radiation therapy (IGRT) to ensure accurate patient positioning prior to beam delivery. Data collection and variable definitions For each patient, a comprehensive set of 24 parameters was extracted and categorized as follows: Demographic and clinical characteristics: Age, gender, primary tumor location (classified as right upper lobe [RUL], right lower lobe [RLL], left upper lobe [LUL], or left lower lobe [LLL]), and history of any anticancer treatment prior to SBRT (surgery, chemotherapy, immunotherapy, or targeted therapy), which was dichotomized as prior treatment (OT: 1 = yes, 0 = no). Baseline pulmonary comorbidities were not systematically available and thus were not included in the analysis; this limitation is addressed in the discussion section. Volumetric parameters: internal target volume (ITV, cc), planning target volume (PTV, cc), ipsilateral lung volume excluding the PTV (IL, cc), and the target volume receiving at least the prescription dose (TV, cc). A novel metric, the lung-to-PTV ratio (LPR), was calculated as IL(cc) / PTV(cc) to serve as a surrogate for functional lung reserve ipsilateral to the tumor. Other organs at risk, including the heart and esophagus, were contoured for treatment planning purposes to ensure adherence to dose constraints, but their dosimetric parameters were not analyzed in this study. Dosimetric parameters: Prescription dose (D, Gy), number of fractions (F), dose per fraction (d, Gy). The maximum and mean PTV doses, converted to equivalent doses in 2 Gy fractions (EQD₂) using an α/β ratio of 10 Gy for tumor effect, were denoted as Dmax2 and Davg2, respectively. For the ipsilateral lung, the mean lung dose (MLD) and the relative volumes receiving ≥ 5, ≥10, ≥ 20, ≥25, and ≥ 30 Gy (V₅–V₃₀) were recorded. Plan quality indices: conformity index (CI = TV / PTV), homogeneity index (HI = Dmax / Dprescription), and PTV coverage (PTVcov) by the prescription isodose (%). Outcome: The primary endpoint was the development of SRP, defined as the highest recorded RP grade of ≥ 2 (CTCAE v5.0) during follow-up. Diagnosis was based on clinical symptoms (cough, dyspnea, fever) and corresponding radiographic changes (CT imaging) within the radiation portal, corroborated by the treating radiation oncologist and a dedicated thoracic radiologist to exclude infectious or progressive disease etiologies. Statistical analysis and model development Statistical analyses were performed using Python (version 3.9) with key libraries including pandas, numpy, scipy, statsmodels, scikit-learn, matplotlib, and seaborn. A structured, three-step analytical framework was implemented to ensure model robustness and clinical applicability. Step 1: Enhanced Univariate Screening and Descriptive Statistics. The study cohort was stratified into two groups based on the primary outcome: the SRP group (RP grade ≥ 2) and the non-SRP group (RP grade < 2). Descriptive statistics were reported as median with interquartile range (IQR) for continuous variables and as frequency with percentage (%) for categorical variables. Group comparisons were performed using the non-parametric Mann-Whitney U test for continuous variables and the Chi-square test for categorical variables. Variables with a p-value < 0.05 were considered statistically significant and were retained as candidate predictors for subsequent analyses. Effect sizes for significant continuous variables were quantified using Cohen’s d, with |d| > 1 indicating a large effect. Step 2: Multicollinearity Diagnosis and Variable Reduction. To prevent model instability arising from high correlations among candidate predictors, a rigorous multicollinearity assessment was conducted on the statistically significant continuous variables identified in Step 1. First, a Spearman rank-order correlation matrix was computed, and variable pairs with an absolute correlation coefficient |r| > 0.8 were considered highly correlated. Second, variance inflation factor (VIF) analysis was performed to quantify the degree of multicollinearity. A standard threshold of VIF > 10 was used to indicate severe multicollinearity requiring correction. An iterative variable elimination procedure was then applied: the variable with the highest VIF exceeding 10 was removed, VIFs were recalculated for the remaining variables, and this process was repeated until all retained variables exhibited VIF < 10. This approach prioritized statistical independence over univariate significance, yielding a refined, non-redundant set of predictors suitable for multivariate modeling. Step 3: Multivariate Logistic Regression Modeling. Using the refined set of candidate variables obtained from Step 2, an initial multivariate logistic regression model was constructed to identify independent predictors of SRP. Subsequently, to ensure robustness and optimal predictive performance, a comprehensive modeling and validation framework was implemented. Both univariate and multivariate logistic regression models were developed for the final selected predictors. All continuous variables were standardized prior to modeling. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the training data. Models were trained and evaluated using 5-fold stratified cross-validation, with performance assessed based on the receiver operating characteristic (ROC) curve, the area under the ROC curve (AUC), accuracy, sensitivity, specificity, and F1-score. The final multivariate model was selected based on its superior and balanced performance across these metrics. The generalizability of the selected model was further validated on an independent test set comprising 20% of the total cohort, which was held out from all training and cross-validation procedures. Within the final model, predictor importance was quantified by the absolute magnitude of standardized regression coefficients. For clinical translation, a three-tiered risk stratification system (low, medium, and high risk) was established based on predefined probability thresholds derived from the model’s predicted probabilities. Model calibration was assessed using a calibration plot, the Brier score, and the calibration slope. Results Baseline characteristics and univariate analysis for predictor screening Of the 120 patients included, the median age was 66 years, with a male predominance (65.0%). Tumors were most frequently located in the LUL, and the median PTV was 43.01 cc. Treatment planning was predominantly performed using the Pinnacle system, with acceptable plan quality metrics (median CI: 1.03; HI: 1.16; PTV coverage: 95.24%). The incidence of SPR (RP ≥ grade 2) was 43.3%, though the majority of cases were grade 2, and no grade 4 or 5 events were observed. Detailed baseline and dosimetric characteristics are summarized in Table 1. Table 1Patient characteristics and univariate analysisVariableTotal(n = 120)Non-SRP(n = 68)SRP(n = 52)P-valueEffect SizeAge, years66.00[31.00–89.00]66.00[32.00–89.00]66.00[31.00–85.00]0.780.04SexM:78/F:42M:47/ F:21M:31/F:210.280.08Tumor locationLUL:50/RUL:33 /RLL:25/LL:12LUL:30/RUL:18 /RLL:13/LLL:7LUL:20/RUL:15 /RLL:12/LLL:50.910.07TPSPinnacle: 81/Monaco: 39Pinnacle: 51/ Monaco: 17Pinnacle: 30/ Monaco: 220.050.17OT1:65/0:550: 41/1:271:38/0:14< 0.0010.32D, Gy60.00[24.00–72.00]60.00[24.00–72.00]60.00[40.00–72.00]0.690.12F10/8/12/6/7/5/9/11/14/310/6/8/12/7/5/9/11/310/12/7/8/6/5/14/110.560.25d, Gy/F6.0/8.0/5.0/10.0/7.0/7.5/9.06.0/8.0/5.0/10.0/7.06.0/8.0/5.0/10.0/7.0/7.5/9.00.730.17ITV, cc13.53[0.59-138.43]7.46[0.64–45.59]20.17[0.59-138.43]< 0.0010.99PTV, cc43.01[7.03-211.64]28.89[7.03–96.38]64.89[17.48-211.64]< 0.0011.25TV, cc43.80[7.78-208.87]29.80[7.78–99.42]66.53[16.28-208.87]< 0.0011.2PTVcov, %95.24[74.54–100.00]95.23[81.44–99.47]95.28[74.54–100.00]0.980.22CI1.03[0.65–1.42]1.03[0.77–1.42]1.03[0.65–1.32]0.450.22HI1.16[1.05–2.13]1.15[1.08–1.47]1.171.05–2.13]0.140.28Dmax2, Gy105.28[55.05–166.30]103.43[55.05–166.30]108.12[70.19-143.91]0.190.2Davg2 Gy89.23[42.67-126.87]87.05[42.67-126.87]92.16[49.03-120.98]0.490.11MLD, Gy8.62[2.63–24.43]6.43[2.63–14.90]11.26[5.88–24.43]< 0.0011.51IL, cc1431.70[752.36-2910.99]1497.63[777.22-2910.99]1405.01[752.36-2329.11]0.130.31V₅, %29.56[6.97–63.14]23.09[6.97–42.76]32.62[18.23–63.14]< 0.0011.24V₁₀, %21.30[5.92–54.04]15.84[5.92–34.40]25.37[14.04–54.04]< 0.0011.34V₂₀, %13.16[1.47–35.09]9.76[3.11–21.96]17.77[1.47–35.09]< 0.0011.38V₂₅, %10.75[2.41–31.01]7.55[2.41–18.15]14.52[6.36–31.01]< 0.0011.55V₃₀, %8.95 [1.92–27.92]6.51[1.92–15.64]13.03[4.74–27.92]< 0.0011.58LPR35.89[6.50-202.69]49.48[8.26-202.69]20.71[6.50-93.05]< 0.0011.13 Patient characteristics and univariate analysis 66.00 [31.00–89.00] 66.00 [32.00–89.00] 66.00 [31.00–85.00] Pinnacle: 81/ Monaco: 39 60.00 [24.00–72.00] 60.00 [24.00–72.00] 60.00 [40.00–72.00] 10/8/12/6/7/5/ 9/11/14/3 10/6/8/12/7/5/ 9/11/3 10/12/7/8/6/5 /14/11 6.0/8.0/5.0/10.0/ 7.0/7.5/9.0 6.0/8.0/5.0/ 10.0/7.0 6.0/8.0/5.0/10.0/ 7.0/7.5/9.0 13.53 [0.59-138.43] 7.46 [0.64–45.59] 20.17 [0.59-138.43] 43.01 [7.03-211.64] 28.89 [7.03–96.38] 64.89 [17.48-211.64] 43.80 [7.78-208.87] 29.80 [7.78–99.42] 66.53 [16.28-208.87] 95.24 [74.54–100.00] 95.23 [81.44–99.47] 95.28 [74.54–100.00] 105.28 [55.05–166.30] 103.43 [55.05–166.30] 108.12 [70.19-143.91] 89.23 [42.67-126.87] 87.05 [42.67-126.87] 92.16 [49.03-120.98] 1431.70 [752.36-2910.99] 1497.63 [777.22-2910.99] 1405.01 [752.36-2329.11] 29.56 [6.97–63.14] 23.09 [6.97–42.76] 32.62 [18.23–63.14] 21.30 [5.92–54.04] 15.84 [5.92–34.40] 25.37 [14.04–54.04] 35.89 [6.50-202.69] 49.48 [8.26-202.69] 20.71 [6.50-93.05] Comparison between the SRP (n = 52) and non-SRP (n = 68) groups identified twelve variables with significant associations to SRP development (p < 0.05). Among them, all DVH (V5-V30, MLD) and volume (ITV, PTV, TV, LPR) parameters exhibited highly significant between-group differences (p < 0.0001) and large effect sizes (absolute Cohen’s d > 1). A significant association was also observed for a history of OT (p = 0.0006). In contrast, the distribution of the two different treatment planning systems (TPSs) showed only borderline statistical significance (p = 0.0449). Notably, patient age, gender, tumor location, prescription (D, F, d), dose parameters (Dmax2, Davg2) and traditional plan quality indices (CI, HI, PTVcov) showed no significant differences between groups. These results are summarized in Table 1, and the twelve significant variables are shown in Fig. 1. List of the twelve statistically significant variables (P < 0.05) Multicollinearity analysis and predictor selection The distribution of the significant variables suggests the presence of potential multicollinearity. As shown in Fig. 2(a), patients with SRP (RP ≥ grade 2) exhibited consistently higher DVH values across all parameters (V5–V30) compared to those without SRP (RP < grade 2). For example, the median V5 was 23.1% in the non-SRP group versus 32.6% in the SRP group, and the median V20 was 9.8% and 17.8%, respectively. A similar trend was observed for MLD, which was also higher in the SRP group. The SRP group had significantly larger volumetric parameters (ITV, PTV, and TV) but a significantly lower LPR compared to the non-SRP group. This inverse relationship indicates that a larger PTV relative to IL volume is associated with a higher risk of SRP, as shown in Fig. 2(b). Figure 2(c) suggests that the proportion of patients developing SRP was lower in those planned with the Pinnacle system and higher in those planned with the Monaco system, compared to the non-SRP group. This disparity may be attributed to the different dose calculation algorithms employed by the two systems [25]. However, whether the TPS genuinely influences the probability of SRP occurrence warrants further investigation. Figure 2(d) indicates that patients with a history of OT were more susceptible to developing SRP after SBRT. Comparative analysis of parameters between patients with non-SRP (RP < 2) and SRP (RP ≥ 2). (a) Box plots of DVH parameters (V5- V30, and MLD). (b) Box plots of volumetric parameters (ITV, PTV, TV, and LPR). (c) Stacked bar chart showing the distribution of TPS. (d) Stacked bar chart illustrating the distribution of OT history Directly incorporating all univariately significant variables (p < 0.05) into a multivariable model could impair its stability and predictive accuracy due to collinearity. Therefore, a correlation analysis was performed for the 10 significant continuous variables prior to multivariate modeling to identify and exclude redundant parameters. The initial Spearman correlation matrix shown in Fig. 3(a) revealed extensive high correlations (|r| > 0.8), particularly among the IL dose-volume parameters (V5-V30 and MLD) and between target volumes (PTV, TV, ITV and LPR). Initial VIF values exceeded 20 for most of these parameters. Following the iterative VIF-based removal procedure in Fig. 3(b), the highly collinear set was distilled. Variables like V20, V25, V30, MLD, and PTV were sequentially removed due to VIF > 10. This process prioritized retaining V5 as a representative of the low-dose volume and LPR as a unique volumetric composite. Consequently, the set of variables deemed sufficiently independent for multivariate analysis consisted of: V5 (continuous), LPR (continuous), TPS (categorical) and OT (categorical). Correlation and collinearity of significant continuous parameters. (a) Matrix of Pearson correlation coefficients, visualizing interrelationships between key parameters. (b) Initial variance inflation factor (VIF) values for the same variable set, quantifying potential redundancy Prediction model development and validation To identify independent predictors of SRP, an initial multivariate logistic regression model was constructed using the candidate variables (V5, LPR, OT, and TPS) selected after multicollinearity reduction. The results in Table 2 confirmed that V5, LPR, and OT were statistically significant independent predictors (all p < 0.01), while TPS was not (p = 0.265). Table 2Results of the initial multivariate logistic regression analysisPredictorCoefficientStd. ErrorOdds Ratio (OR)95% CI for ORP-valueV50.1100.0351.1161.042–1.1950.002LPR-0.0570.0170.9450.915–0.976< 0.001OT1.4310.5244.1851.497–11.6940.006TPS-0.6330.5680.5310.174–1.6170.265Const-1.0631.5860.3460.015–7.7370.503 Results of the initial multivariate logistic regression analysis Both univariate and multivariate logistic regression models were constructed for the three candidate variables (V5, LPR, OT) using a robust framework incorporating standardization, SMOTE for class imbalance handling, and 5-fold stratified cross-validation. The performance of the models is summarized in Fig. 4; Table 3. As shown in Figs. 4(a) and 4(b), the multivariate model yielded the highest mean AUC (0.875) and the best balance among specificity (0.769), accuracy (0.775), and F1-score (0.756). Consequently, it was selected as the optimal predictive model. This selected multivariate model was further validated on a held-out independent test set (n = 24, 20% of the total cohort), where it exhibited robust generalizability with an AUC of 0.921, an accuracy of 0.833, a sensitivity of 0.80, and a specificity of 0.857. The comparative ROC curves of all models on the independent test set are presented in Fig. 4(c), and feature importance analysis of the multivariate model, shown in Fig. 4(d), identified LPR as the strongest predictor, with an absolute coefficient of 1.357, followed by V5 and OT. The final multivariate logistic regression model for predicting the probability of SRP is given by the following formula: Comparative analysis of univariate and multivariate Models for SRP Prediction. (a) Bar chart comparing mean AUC values (± standard deviation) of the four prediction models. (b) Radar plot displaying four key performance metrics for each model: sensitivity, specificity, accuracy, and F1-score. (c) ROC curves for all models evaluated on an independent test set (20% of total data). (d) Horizontal bar chart of feature importance based on absolute coefficient values Table 3Cross-validation performance of predictive models for SRPModelAUC(MeanSD)Sensitivity (Mean ± SD)Specificity (Mean ± SD)Accuracy (Mean ± SD)F1-score (Mean ± SD)Univariate_V50.808 ± 0.0610.773 ± 0.1720.693 ± 0.0860.725 ± 0.0370.703 ± 0.069Univariate_LPR0.866 ± 0.1020.825 ± 0.0850.659 ± 0.1740.733 ± 0.0860.733 ± 0.074Univariate_OT0.669 ± 0.1350.735 ± 0.1650.604 ± 0.1780.658 ± 0.1360.652 ± 0.127Multivariate0.875 ± 0.0700.791 ± 0.0710.769 ± 0.1600.775 ± 0.0700.756 ± 0.055 Cross-validation performance of predictive models for SRP For clinical translation, patients were stratified into three risk categories based on the model’s predicted probability: Low Risk (P < 0.3, 38.3% of the cohort), Medium Risk (0.3 ≤ P ≤ 0.7, 31.7% of the cohort) and High Risk (P > 0.7, 30.0% of the cohort). The scatter plot of V5 versus LPR in Fig. 5(a) demonstrates a clear separation of patients across the three predicted risk strata, with distinct clustering based on OT status. Patients without a history of OT, characterized by lower V5 and higher LPR, predominantly fall into the low-risk category. In contrast, those with a history of OT, exhibiting higher V5 and lower LPR, are mainly classified as high-risk. This distinct separation confirms the model’s ability to integrate V5, LPR, and OT into a coherent and effective risk stratification. Individualized SRP risk rating system. (a) Scatter plot of V5 versus LPR, with points colored by the model-predicted risk level (Low: Green, Medium: Gold, High: Red) and shaped by OT history (circle: OT = 0, square: OT = 1). (b) The calibration curve of the final multivariate model. The shaded background zones correspond to the clinical risk strata (low, medium, high), and the vertical dotted lines at 0.3 and 0.7 mark the decision thresholds Figure 5(b) presents the calibration curve of the final multivariate model, which assesses the agreement between predicted probabilities and observed SRP incidence. The plotted points lie close to the ideal 45-degree diagonal line across the entire probability range, indicating excellent calibration. This visual impression is supported by a low Brier score of 0.13 and a calibration slope of 1.06, both of which demonstrate that the model’s probability estimates are accurate. The shaded background zones correspond to the clinical risk strata (low, medium, high), and the vertical dotted lines at 0.3 and 0.7 mark the decision thresholds. The sample sizes (n) shown for each decile further confirm that the calibration assessment is robust across all risk levels. Discussion This study demonstrates that SRP after lung SBRT can be predicted by three fundamentally distinct factors: a low-dose parameter (V5), a novel volumetric ratio (LPR), and a clinical sensitizing event (OT). Beyond developing a predictive model with excellent performance (cross-validated AUC 0.875; test set AUC 0.921), the findings challenge prevailing dosimetric dogmas, introduce a conceptually new risk dimension, and reveal an important interaction between physical treatment parameters and biological susceptibility. Re-evaluating the dosimetric determinants of SRP in the SBRT Although all dosimetric parameters of IL from V5 through V30 demonstrated strong associations in univariate analysis (p < 0.0001) with large effect sizes, their near-perfect collinearity (Spearman |r| > 0.8; initial VIF > 20) rendered them statistically interchangeable. The iterative VIF-based elimination process sequentially removed V10, V20, V25, V30, and MLD, ultimately retaining only V5. This statistical selection indicates that, among this correlated family of dose-volume metrics, V5 represents the most informative surrogate for the low-dose bath phenomenon. Furthermore, the incremental predictive information contained in higher dose thresholds becomes largely redundant once V5 is known. This finding directly challenges the traditional reliance on V20 and MLD as cardinal predictors in CFRT [11, 12, 26]. The distinct radiobiology of hypofractionation offers an explanation: under this regimen, the volumetric spread of low-dose radiation may carry greater biological weight than the magnitude of high-dose regions [27, 28]. This interpretation is consistent with the hypothesis that SBRT triggers qualitatively different mechanisms of lung injury compared to conventional fractionation [29, 30]. The pronounced vascular endothelial damage and pro-inflammatory cytokine cascade triggered by large single fractions may exhibit a lower dose threshold for initiation than previously appreciated [31–33]. These insights carry practical implications for plan optimization and the prioritization of clinical constraints, suggesting that in the SBRT setting, limiting the volume of IL exposed to very low doses (e.g., V5) may be more important than controlling traditional intermediate‑dose parameters. LPR: Beyond absolute volumes to functional reserve The identification of the LPR as the single strongest predictor introduces a conceptually distinct risk dimension that transcends conventional volumetric thinking. While absolute target volumes (ITV, PTV, TV) were highly significant in univariate analysis (p < 0.0001; Cohen’s d > 1), their elimination during multicollinearity reduction in favor of LPR is instructive: absolute tumor burden appears to matter less than tumor burden relative to the available lung reserve. Patients with a low LPR—a large PTV occupying a small IL—face a double jeopardy: not only is a greater proportion of functioning tissue exposed to radiation, but the remaining unirradiated lung may possess insufficient reserve to compensate for the inflammatory and fibrotic sequelae. This ratio-based perspective aligns with the fundamental principles of pulmonary physiology, wherein functional capacity is determined by the interplay between demand and reserve. The marked separation in LPR between SRP and non-SRP groups (median 23.5% vs. 41.6% in Fig. 2(b)) underscores its discriminatory power. Importantly, LPR captures information orthogonal to V5: two patients with identical V5 values may experience vastly different risks depending on whether that 5 Gy dose is distributed across a large or small lung volume. This orthogonality, confirmed by multicollinearity diagnostics, explains why the combination of V5 and LPR yields synergistic predictive value. OT: The clinical dimension of pulmonary susceptibility The independent contribution of OT introduces the critical dimension of biological susceptibility into a model otherwise dominated by physical parameters. Remarkably, this effect persists despite the variable’s binary and composite nature: even without granular data on specific agents or timing, a simple “yes/no” history of OT nearly triples the odds of SRP. This finding likely reflects the cumulative subclinical injury inflicted by OT—chemotherapy-induced microvascular damage, immunotherapy-mediated immune priming, or targeted therapy effects on alveolar repair mechanisms. The presence of OT effectively lowers the threshold at which a given V5 or LPR translates into clinical pneumonitis. This interaction is visually evident in Fig. 5(a), where some patients with OT and intermediate or low dosimetric parameters also develop SRP. Integrating the three dimensions: a visual framework for clinical decision-making The multivariate model reveals that SRP risk is not governed by a single dominant factor but emerges from the interplay of three distinct dimensions: the low-dose volume (V5), the functional lung reserve (LPR), and the sensitising clinical condition (OT). To translate this multidimensional relationship into a practical tool, two-dimensional risk maps stratified by OT status were generated and are presented in Fig. 6. These maps plot V5 against LPR, with a contour line marking the 30% predicted SRP probability threshold for low risk. By locating a patient’s LPR and planned V5 on the appropriate map, clinicians can instantly determine whether the combination lies within the low-risk region and, if not, the extent of deviation from safety. Risk maps for SRP stratified by OT. Relationship between V5 and LPR in patients without (a) and with (b) OT. The solid line indicates the 0.3 predicted probability threshold for SRP (low-risk cutoff). The shaded area represents the low-risk region (P < 0.3). Points A–F on the boundary show maximum V5 values for low risk at representative LPR levels, with values provided in the table As shown in Fig. 6(a), for patients without a history of OT, the low-risk region lies below the solid line defined by Eq. (2), which corresponds to a predicted SRP probability of exactly 30%: The positive slope of the line reflects the protective effect of larger functional lung reserve: as LPR increases, higher V5 values are tolerated without exceeding the risk threshold. This continuous trade-off between low-dose volume and functional reserve is exemplified by three points on the boundary: A (LPR = 30, V5 = 23.9%), B (LPR = 60, V5 = 35.8%), and C (LPR = 80, V5 = 43.7%). For a patient with limited reserve (LPR = 30), V5 must remain below 23.9% to stay low-risk, while for one with abundant reserve (LPR = 80), V5 can be as high as 43.7%—highlighting how anatomical factors modulate permissible dosimetric constraints. Figure 6(b) shows that a history of OT shifts the risk boundary downward to the line defined by: Indicating the substantial risk increment conferred by the sensitising condition. Points D (LPR = 30, V5 = 15.4%), E (LPR = 60, V5 = 26.5%), and F (LPR = 80, V5 = 34.4%) on this line illustrate the tightened trade-off. Even with abundant lung reserve (point F), V5 must remain below 34.4% to keep the predicted SRP probability below 0.3; for average or limited LPR (points D and E), the permissible V5 becomes exceedingly small. Thus, OT governs the risk equation, leaving minimal flexibility for dosimetric compromise. This visual framework synthesizes the three dimensions into a practical tool. The clinician can select the appropriate map and plots the point (LPR, planned V5). If it falls within the shaded low-risk area, the plan is acceptable; if above the boundary, plan modifications (e.g., reducing V5 via improved conformality or beam geometry) should be considered. When modification is not feasible, the patient should be counselled regarding the elevated risk. The maps also underscore that LPR is a fixed anatomical determinant of intrinsic vulnerability: patients with low LPR require stricter V5 constraints, particularly in the presence of OT. Conversely, for patients with high LPR and no OT, moderately higher V5 may be acceptable. Limitations and future directions Despite its integrative framework, this study has several limitations. First, while the model demonstrated robust internal performance (cross-validated AUC 0.875; test set AUC 0.921), external validation in independent cohorts is necessary to confirm its generalizability, particularly as the 43.3% SRP incidence observed here may reflect institution-specific factors. Second, the binary OT variable does not distinguish therapy types (chemotherapy vs. immunotherapy vs. targeted agents) or timing, despite emerging evidence that checkpoint inhibitors may potentiate risk through distinct mechanisms. Third, pre-existing lung disease and smoking status—potential modifiers of pulmonary reserve—were unavailable. Finally, the model predicts SRP occurrence but not severity, timing, or treatment response. These limitations point to clear next steps. Prospective studies should evaluate whether integrating this visual framework into treatment planning reduces SRP incidence. Radiomics may capture parenchymal features beyond LPR and V5, while circulating biomarkers could further refine the biological susceptibility dimension—potentially transforming binary OT into a continuous measure. Ultimately, the goal is prevention: this three-dimensional framework can guide personalized strategies—such as refining treatment plans, implementing prophylaxis, or scheduling surveillance—thereby evolving from a visual aid into a validated clinical tool. Conclusion In this study, SRP after lung SBRT was determined by three distinct factors: a low-dose parameter (V5), a volumetric ratio reflecting functional lung reserve (LPR), and other anticancer therapy administered prior to SBRT (OT). Through rigorous multicollinearity analysis, V5 was identified as the optimal surrogate among all correlated dose-volume metrics for the low-dose bath phenomenon, while LPR quantifies the relationship between tumor burden and available lung reserve and emerged as the dominant predictor in the final model. The independent contribution of OT highlights the biological dimension of pulmonary susceptibility, modifying the threshold at which dosimetric parameters translate into clinical toxicity. The resulting three-factor model demonstrated excellent discriminative performance (cross-validated AUC 0.875; test set AUC 0.921) and calibration (Brier score 0.13), confirming its robustness within this cohort. Translation of the model into risk maps stratified by OT provides a clinical tool for individualized plan evaluation, enabling assessment of the trade-off between low-dose spread and lung functional reserve. These findings challenge conventional emphasis on high-dose metrics in the CFRT context and establish a parsimonious framework integrating physical, anatomical, and biological determinants of lung injury. Nevertheless, external validation in prospective, multi-institutional cohorts remains warranted before widespread clinical implementation. Author contributions X. Yang, Q.C. Wang, C.L. Ruan, and L.Z. Wang conducted the literature search and study design. X. Yang and H.Y. Gong collected the data. X. Yang and Q.C. Wang, processed statistical information, and contributed to data interpretation. X. Yang drafted the manuscript, which was subsequently revised by D. Huang and H.Y. Gong. Funding No funding to declare. Data availability All relevant data and materials have been included in the article. Further inquiries can be directed to the corresponding authors. Declarations Ethics approval and consent to participate This retrospective cohort study was conducted in accordance with the Declaration of Helsinki and approved by the Clinical Research Ethics Committee of Renmin Hospital of Wuhan University (approval number: 2024K-K037), with a waiver for informed consent due to the retrospective nature. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xiong Yang and Quncai Wang contributed equally to this work. Contributor Information Xiong Yang, Email: rm003627@whu.edu.cn. Hongyun Gong, Email: dr_gonghongyun@163.com. David Huang, Email: huangd2013@outlook.com. References Associated Data Data Availability Statement All relevant data and materials have been included in the article. Further inquiries can be directed to the corresponding authors.