Low-Dose Cardiac Exposure (Heart V5) as an Independent Predictor of Radiation-Induced Pericarditis in Breast Cancer: An Interpretable Machine Learning Study

Low-Dose Cardiac Exposure (Heart V5) as an Independent Predictor of Radiation-Induced Pericarditis in Breast Cancer: An Interpretable Machine Learning Study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Low-Dose Cardiac Exposure (Heart V5) as an Independent Predictor of Radiation-Induced Pericarditis in Breast Cancer: An Interpretable Machine Learning Study Yongkai Lu, Lei Xu, Jian Zhang, Rongze Ma, Yao Wang, Shanshan Gao This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8105555/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background In breast cancer treatment, high-dose cardiac irradiation has traditionally been the focus, but growing evidence suggests that even low-dose radiation can cause long-term cardiac damage, raising concerns about radiation-induced heart disease (RIHD) and pericarditis as significant survivorship issues. Understanding dosimetric predictors like low-dose cardiac exposure (heart_V5) is critical for optimizing treatment safety. Methods A retrospective cohort study was conducted from January 2022 to June 2023, involving 277 female breast cancer patients (training n = 224, test n = 53). Radiation-induced pericarditis risk, the primary outcome, was derived from an NTCP model binarized at 3.8487×10⁻⁶ to classify the top 25% as high-risk. Key exposure was heart_V5 (median 31.5%, IQR 21.9–42.2), and analytics included preprocessing, six machine-learning models (XGBoost best-performing) with 5-fold cross-validation, DeLong test for AUC comparison, and SHAP for model interpretation. Results On the independent test set (n = 53), the XGBoost model achieved superior performance, with an area under the receiver operating characteristic curve (AUC) of 0.918 (95% CI 0.848–0.987), an accuracy of 0.925, a recall of 0.727, and an F1-score of 0.800. Critically, to isolate the impact of low-dose radiation, a simplified XGBoost model excluding high-dose features was developed. This model maintained high predictive power (AUC = 0.903), underscoring the significant, independent predictive value of low-dose cardiac exposure (heart_V5). In this simplified model, heart_V5 was a top-three predictor, demonstrating a non-linear dose-response relationship. While high-dose metrics (heart_V30, heart_V40) were the dominant predictors in the full model, the strong performance of the simplified model confirms that heart_V5 is a key factor for pericarditis risk. Furthermore, heart_V5 was significantly correlated with the mean heart dose (Pearson r = 0.75, p < 0.001). Conclusions Low-dose cardiac exposure (heart_V5) demonstrates independent predictive value for radiation-induced pericarditis when high-dose features are controlled, suggesting its consideration in radiotherapy planning to reduce late pericardial morbidity. Breast cancer Radiotherapy Radiation-induced pericarditis Heart V5 Machine learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction In breast cancer treatment, radiation oncology has long focused on the risks of high-dose cardiac irradiation, yet emerging evidence suggests that low-dose radiation, particularly heart V5 (the percentage of heart volume receiving ≥ 5 Gy), also contributes to long-term cardiac harm[ 1 ]. Radiation-induced heart disease (RIHD), with pericarditis as an early and frequent manifestation, poses a significant challenge in breast cancer radiotherapy (RT). With over 2.3 million women diagnosed annually and approximately half receiving RT, RIHD drives long-term morbidity and mortality [ 2 ]. Pericarditis, characterized by inflammation or effusion, can emerge acutely or years after RT, leading to chest pain or, in severe cases, cardiac tamponade. Research indicates a 7.4% rise in major adverse cardiac events (MACE) per Gray of mean heart dose, contributing to 1–2% excess mortality among survivors [ 3 ]. The economic toll of RIHD is substantial, with high healthcare costs and diminished quality of life [ 4 ]. As five-year breast cancer survival now exceeds 90%, reducing toxicities like pericarditis is essential for improving survivorship and advancing precision oncology[ 5 – 7 ]. Advances in RIHD research reveal a clear dose-dependent relationship between cardiac radiation exposure and adverse outcomes. The Early Breast Cancer Trialists’ Collaborative Group demonstrated that while RT reduces breast cancer mortality, it increases cardiovascular deaths, with relative risks of 1.27–1.61 for left-sided versus right-sided RT [ 8 , 9 ]. Darby et al.’s study of over 2,000 women further confirmed a linear increase in MACE with mean heart dose [ 3 ]. Diagnostic tools, such as echocardiography and cardiac magnetic resonance imaging (MRI), effectively detect pericardial effusions, while heart-sparing techniques like deep inspiration breath-hold (DIBH) and intensity-modulated radiotherapy (IMRT) reduce mean heart dose by 20–50% [ 10 – 12 ]. Dose-volume histogram parameters, including heart V30 and V40 (volumes receiving ≥ 30 Gy and ≥ 40 Gy), are established predictors of pericarditis, with QUANTEC guidelines recommending V30 < 46% and mean heart dose < 26 Gy to keep risk below 15% [ 13 , 14 ]. However, the independent role of heart V5 remains uncertain, as its predictive value is often obscured by collinearity with high-dose metrics like V30 and V40. This collinearity arises because tangential RT fields, commonly used in breast cancer, create overlapping low- and high-dose regions in the heart [ 15 – 17 ], hindering the development of optimized dose constraints and personalized RT planning. To address this gap, advanced machine learning approaches offer a powerful solution for disentangling collinear dosimetric parameters and uncovering the independent predictive value of heart V5. Therefore, this study aims to determine whether heart V5 independently predicts radiation-induced pericarditis risk in breast cancer patients undergoing RT or if its contribution is masked by high-dose cardiac parameters. We conducted a retrospective cohort study of 277 female breast cancer patients receiving RT between January 2022 and June 2023, employing machine learning to enhance risk prediction and inform cardiac-sparing strategies. In this paper, we make several key contributions. First, we demonstrated heart_v5's independent predictive value for pericarditis risk using a simplified XGBoost model. Second, we revealed high-dose cardiac parameters (heart_v30, heart_v40) as dominant predictors in the full model. Third, we established a non-linear dose-response relationship for heart_v5 with clinical outcomes. Notably, the simplified XGBoost model, excluding collinear high-dose parameters, achieved an AUC of 0.887, with heart_v5 ranking as the 4th most important predictor and showing a clear positive dose-response relationship with pericarditis risk. These findings underscore the untapped potential of low-dose metrics in refining RIHD risk models, paving the way for targeted interventions that could mitigate long-term cardiovascular morbidity in breast cancer survivors. Materials and Methods 2.1 Study design and participants A retrospective cohort study was conducted including female breast cancer patients who received radiotherapy at the Department of Radiation Oncology, First Affiliated Hospital of Xi'an Jiaotong University, Shaanxi Province, China, between January 2022 and June 2023. Patient data were extracted from medical records and included age, body mass index (BMI), history of hypertension, diabetes mellitus, alcohol consumption, menopausal status, surgical procedure, tumor stage (TNM), anemia status, and electrocardiogram (ECG) results. Radiotherapy data were obtained from the Elekta Monaco treatment planning system and comprised target volume, radiotherapy technique, tumor location, total dose to the planning target volume (PTV), PTV volume, PTV dose coverage, ipsilateral heart volume, ipsilateral lung volume, and maximum heart distance (MHD)[ 18 ]. Radiation target volumes were delineated by two senior radiation oncologists following Radiation Therapy Oncology Group (RTOG) guidelines. Treatment plans were designed by a senior medical physicist and finalized by consensus between oncologists and physicist[ 19 ]. The study was approved by the Ethics Committee of the First Affiliated Hospital of Xi'an Jiaotong University and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to inclusion. Of 323 patients who consented, 46 were excluded due to missing data, bilateral breast cancer, or male breast cancer, resulting in a final cohort of 277 patients. 2.2 Assessment of pericarditis The primary endpoint was radiation-induced pericarditis incidence, assessed using a Normal Tissue Complication Probability (NTCP) model. This model predicts complication risk based on dose-volume histograms (DVHs) and equivalent uniform dose (EUD) parameters. The NTCP model was based on Emami et al. and implemented via MATLAB following Gay and Niemierko et al. The logistic function used was: where TD50 = 50 Gy and γ50 = 3[20, 21]. EUD values were computed from DVHs to calculate NTCP for each patient. NTCP values were binarized: values ≥ 3.8487 × 10⁻⁶ were classified as high risk (1), and values below as low risk (0). The threshold was selected based on data distribution to separate approximately the highest 25% of risk values. 2.3 Assessment of V5 Cardiac V5 was defined as the percentage volume of the heart receiving a radiation dose of 5 Gy or higher. This parameter was extracted from dose-volume histograms (DVHs) generated by the Elekta Monaco treatment planning system, based on heart contours delineated according to institutional protocols. V5 quantifies low-dose radiation exposure to the heart, which is relevant for assessing radiation-induced cardiac toxicity risk[ 18 , 22 ]. To facilitate statistical analysis, V5 values were divided into three categories: Low (≤ 27%), Medium (27–39%), and High (> 39%). These thresholds corresponded approximately to the 33.33rd and 66.67th percentiles of the V5 distribution in the study population. This percentile-based stratification ensured roughly equal sample sizes across groups, enabling balanced comparisons and improving statistical power for detecting dose-dependent effects on pericarditis risk. 2.4 Assessment of other variables Additional clinical and dosimetric variables were collected from medical records and treatment planning data, including: age (years), BMI (kg/m²), histories of hypertension, diabetes, alcohol consumption, menopausal status (pre/post), surgical procedure (breast conservation, mastectomy, reconstruction), tumor stage (T, N, M0), ECG results (normal/abnormal), anemia status (present/absent), radiotherapy target areas (supraclavicular, axillary, internal mammary lymph nodes, or none), radiotherapy techniques (IMRT, VMAT), lesion laterality (left/right), PTV total dose (< 50 Gy, ≥ 50 Gy), PTV volume (cm³), number of fractions (< 25, ≥ 25), PTV dose coverage (90–95%, ≥ 95%), ipsilateral lung volume (cm³), heart volume (cm³), and MHD (cm). MHD was defined as the maximum perpendicular distance from the anterior heart surface to the posterior tangential field edge. 2.4 Statistical Analysis Descriptive statistics were used to summarize baseline patient characteristics. Continuous variables were presented as median and interquartile range (IQR), while categorical variables were presented as frequencies and percentages. To compare baseline characteristics across heart_v5 tertiles (defined for this analysis as Low: ≤27%; Medium: 27–39%; High: >39%), the Kruskal-Wallis test was used for continuous variables and the Chi-squared (χ²) test was used for categorical variables. The Wilcoxon rank-sum test was used to compare continuous variables between the final high-risk and low-risk groups. Pearson’s correlation coefficient (r) was calculated to assess the degree of multicollinearity between key numeric dosimetric variables. A two-sided p-value < 0.05 was considered statistically significant for these analyses. All statistical analyses were performed using R (version 4.2.2). 2.5 Machine Learning Model Development and Validation 2.5.1 Data Partitioning and Preprocessing For model development, the full cohort (n = 277) was partitioned into a training set (80.9%, n = 224) and an independent test set (19.1%, n = 53) using a stratified split based on the binary outcome variable. A standardized preprocessing pipeline was constructed using the tidymodels package, which included median/mode imputation for missing data, one-hot encoding for all categorical predictors, and normalization (scaling to a mean of zero and standard deviation of one) of all numeric predictors. 2.5.2 Model Training and Evaluation Six machine learning algorithms were evaluated: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). For models with tunable hyperparameters, a 5-fold cross-validation was performed on the training data, with the area under the receiver operating characteristic curve (AUC) used as the metric to select the optimal parameters. The final, tuned models were evaluated on the independent test set using AUC, accuracy, precision, recall, and F1-score. The DeLong test was used to compare the AUCs of different models. 2.5.3 Model Interpretation and heart_v5 Validation The internal logic of the best-performing XGBoost model was interpreted using SHapley Additive exPlanations (SHAP), implemented with the fastshap R package. SHAP values were calculated to quantify the impact of each feature on the model's predictions for every patient in the test set. Model behavior was visualized using SHAP global importance plots, summary (beeswarm) plots, and dependence plots. To address high multicollinearity among cardiac dose parameters and isolate the independent predictive value of heart_v5, a "simplified" XGBoost model was developed. This model was trained on a dataset from which the highly collinear features (heart_v30, heart_v40, heart_mean_dose) were deliberately excluded. This simplified model was then trained and interpreted using the same SHAP methodology to assess the unmasked impact of heart_v5. Results 3.1 Patient Baseline Characteristics and Data Overview The final analysis included 277 female breast cancer patients. Patients were classified into three groups according to tertiles of cardiac V5: Low (≤27%, n=89), Medium (27–39%, n=96), and High (>39%, n=92) (Table 1). Across the three groups, baseline demographic variables were comparable. Median age was 47.6 years (interquartile range [IQR] 43.2–51.0) in the Low group, 47.8 years (IQR 43.1–51.2) in the Medium group, and 47.4 years (IQR 42.8–50.8) in the High group, with no statistically significant difference (p=0.861, Kruskal–Wallis test). Median body mass index (BMI) was 22.9 kg/m² (IQR 21.3–24.4), 23.2 kg/m² (IQR 21.4–24.7), and 23.0 kg/m² (IQR 21.5–24.6), respectively (p=0.823). The prevalence of hypertension was 10.1% (n=9), 9.4% (n=9), and 10.9% (n=10) in the Low, Medium, and High groups, respectively (p=0.952, χ² test). Diabetes was present in 6.7% (n=6), 5.2% (n=5), and 7.6% (n=7) of patients (p=0.815). Alcohol consumption was reported by 7.9% (n=7), 7.3% (n=7), and 8.7% (n=8) of patients (p=0.941). Significant group differences were observed in tumor laterality and nodal stage. Left-sided tumors were more frequent in the High V5 group (20.9%, n=19) than in the Low V5 group (11.6%, n=10) (p=0.001). Patients in the High V5 group had higher rates of N1 nodal stage (15.9%, n=15) compared with the Low group (11.2%, n=10) (p=0.009). Treatment field coverage also differed significantly: supraclavicular lymph node irradiation was delivered in 26.7% (n=25) of High V5 patients compared with 16.2% (n=14) in the Low group (p<0.001), and internal mammary node irradiation in 27.8% (n=26) versus 18.1% (n=16) (p<0.001). The independent test set (n=53) reflected a similar clinical profile (Figure 1). The low-risk group comprised 79.2% (n=42) of the test set and the high-risk group 20.8% (n=11) (Figure 1A). Visual inspection of the heatmap suggested that high-risk patients more often had elevated heart_v5 and BMI values. The distribution of heart_v5 across the test cohort was right-skewed, with a median of 31.5% (IQR 21.9%–42.2%) and a range from 0.17% to 58.5% (Figure 1B). Tumor staging in the test set was dominated by T3 stage (n=30, 56.6%) and N1/N2 stage (n=48, 90.6%). Right-sided lesions were more frequent (n=29, 54.7%) than left-sided lesions (n=24, 45.3%) (Figure 1C, 1D). In univariate analysis, median heart_v5 was higher in high-risk patients (34.5%, IQR 27.5–38.2%) than in low-risk patients (29.5%, IQR 25.0–33.8%), although this difference was not statistically significant (p=0.530, Wilcoxon rank-sum test; effect size r=0.12, 95% confidence interval [CI] −0.10 to 0.33).Pearson correlation analysis of dosimetric parameters demonstrated strong multicollinearity: heart_v5 was positively correlated with heart_mean_dose (r=0.75, p<0.001) (Figure 1F). These findings indicate substantial overlap in the information carried by low-dose and mean heart dose metrics. Overall, the baseline data show that while demographic factors were balanced across V5 tertiles, tumor laterality, nodal stage, and regional lymph node irradiation differed significantly. The test set preserved these distributions and highlighted both the skewed nature of heart_v5 and its high correlation with other dosimetric measures. Table 1 Baseline Characteristics of Study Subjects by heart V5 Indicators Characteristics Clustering of cardiac V5 indicators Low heart V5 (≤27) Medium heart V5(27-39) High heart V5(>39) P No. of subjects 89 96 92 Age (years) 47.6 (39.0,54.0) 47.8 (39.2,55.0) 47.4 (39.0,53.8) 0.861 BMI (kg/m 2 ) 22.9 (0.6, 24.7) 23.2 (20.5, 25.6) 23.0 (21.3, 24.6) 0.823 History of hypertension Yes（%） 1.80 1.10 0.70 0.440 History of diabetes Yes（%） 1.10 0.70 0.40 0.572 History of alcohol consumption Yes（%） 0.70 0.40 0.40 0.743 Menopausal Yes（%） 6.90 6.90 5.10 0.545 Surgical procedure Breast Conservation Surgery 18.40 13.0 13.0 0.012 Mastectomy 12.30 18.10 14.10 0.147 Reconstruction 1.40 3.60 6.10 0.009 T staging Tis 1.10 0.40 0.10 0.151 T1 20.2 15.5 16.6 0.041 T2 10.1 16.6 13.0 0.073 T3 0.40 1.40 2.90 0.042 T4 0.40 0.70 0.70 0.821 N staging N0 19.10 8.70 12.60 0.000 N1 11.20 19.90 15.90 0.009 N2 1.10 4.00 2.90 0.120 N3 0.70 2.20 1.80 0.420 M staging M0 Yes（%） 31.80 34.70 33.20 0.346 Normal ECG Yes（%） 30.30 34.30 31.00 0.140 Anemic Yes（%） 0.70 1.10 1.10 0.903 Radiotherapy target Supraclavicular Lymph Nodes 16.20 28.20 26.70 0.000 Axillary Lymph Nodes 18.40 29.60 27.80 0.000 Internal Mammary Lymph Nodes 18.10 29.60 27.80 0.000 No lymph nodes 0.40 0.70 0.70 0.841 Radiotherapy techniques IMRT 6.500 8.30 6.90 0.794 VMAT 25.60 26.40 26.40 0.796 Continued Characteristics Low heart V5 (≤27) Medium heart V5(27-39) High heart V5(>39) P Lesion position Left side 11.60 14.40 20.90 0.001 Right side 20.60 20.20 11.90 0.000 PTV total dose <50Gy 9.70 4.00 2.90 0.000 ≥50Gy 22.40 30.70 30.30 0.000 PTV volume 688.2(527.4,872.9) 746.9(611.3,919.8) 766.5(632.6,1020.7) 0.000 Fractions <25 9.00 2.90 2.20 0.001 ≥25 23.10 31.80 31.00 0.001 PTV dose coverage 90%-95% 4.70 9.00 6.90 0.158 ≥95% 27.40 25.60 26.40 0.158 Ipsilateral lung volume 1397.1(1201.9,1745.0) 1359.3(1219.9,1559.9) 1250(1114.9,1456.1) 0.003 Heart volume 533.2(467.9,609.1) 560.9(486.5,631.9) 544.5(486.3,644.4) 0.217 MHD (cm) 10.0(5.0,11.0) 10.0(5.25,11.0) 7.0(5.62,10.0) 0.012 Notes: Continuous variables were present as median (25%, 75% quantiles) and categorical variables are shown as percentages. Kruskal–Wallis test (comparison of >2 groups) for continuous and ordinal distributed variables and the chi-squared test for categorical variables. BMI, body mass index; ECG, electrocardiogram; IMRT, intensity-modulated radiation therapy; VMAT, volumetric modulated arc therapy; PTV, planning target volume MHD, maximum heart distance. 3.2 Machine Learning Model Performance in Independent Test Set Six predictive models—logistic regression (LR), random forest (RF), support vector machine (SVM), k-nearest neighbors (KNN), multilayer perceptron (MLP), and extreme gradient boosting (XGBoost)—were trained on the development cohort and evaluated on the independent test set (n=53). Performance metrics are summarized in Table 2, with graphical comparisons in Figure 2. Metrics reported include accuracy, area under the receiver operating characteristic curve (AUC), precision, recall, and F1-score. In receiver operating characteristic (ROC) analysis, XGBoost achieved the highest AUC at 0.918 (95% confidence interval CI 0.848–0.987), exceeding LR (AUC=0.827, 95% CI 0.713–0.940), RF (AUC=0.890, 95% CI 0.769–0.960), SVM (AUC=0.848, 95% CI 0.758–0.950), KNN (AUC=0.752, 95% CI 0.681–0.921), and MLP (AUC=0.826, 95% CI 0.789–0.965) (Figure 2A). Individual receiver operating characteristic (ROC) curves for each model are provided in Supplementary Figure 1. Pairwise DeLong tests indicated that XGBoost significantly outperformed LR (p=0.041) and KNN (p=0.018) in AUC. Precision–recall (PR) analysis, more informative for imbalanced datasets, showed that XGBoost had the largest average precision (0.861), followed by RF (0.824), MLP (0.811), SVM (0.792), LR (0.713), and KNN (0.698) (Figure 2B). This corresponded to higher precision when identifying the minority high-risk class. The radar plot (Figure 2C) displayed a balanced performance profile for XGBoost across metrics: accuracy 0.925, AUC 0.918, precision 0.889, recall 0.727, and F1-score 0.800. RF achieved accuracy 0.792, precision not applicable, recall 0.0, and F1-score 0.0. LR had lower accuracy (0.811) and recall (0.636) with an F1-score of 0.583. Confusion matrix comparison demonstrated that XGBoost correctly identified 8 of 11 high-risk patients (true positives), with 3 false negatives, 1 false positive, and 41 true negatives. This yielded a recall of 72.7%, specificity of 97.6%, and precision of 88.9%. In contrast, LR detected 7 of 11 high-risk patients, with 4 false negatives, 6 false positives, and 36 true negatives, resulting in a recall of 63.6%, specificity of 85.7%, and precision of 53.8% (Figures 2D, 2E). The learning curve for the XGBoost model's AUC is presented in Figure 2F. The curve shows that as the training set size increases, the Test Set AUC (validation AUC) rises to a final value near 1.0. The Train Set AUC (training AUC) stabilizes around 0.91. The gap between the final training and validation AUC scores suggests the model may exhibit some overfitting, although the validation performance remains high. Subgroup performance was consistent across left- and right-sided tumor patients. In left-sided cases (n=21), XGBoost achieved AUC 0.924 and accuracy 0.905; in right-sided cases (n=25), AUC 0.912 and accuracy 0.936 were observed. For patients receiving internal mammary node irradiation (n=15), XGBoost maintained high discrimination (AUC=0.917) with recall 0.733, while in those without internal mammary irradiation (n=41), AUC was 0.919 with recall 0.724. Across all models, F1-scores ranged from 0.583 (LR) to 0.800 (XGBoost), and balanced accuracy ranged from 0.732 (KNN) to 0.852 (XGBoost). Matthews correlation coefficient (MCC) was highest for XGBoost at 0.783, compared with 0.604 for LR and 0.694 for RF. Cohen’s kappa values mirrored this pattern (XGBoost κ=0.791; LR κ=0.603; RF κ=0.702). Overall, in the independent test set, the XGBoost algorithm demonstrated the highest discrimination, precision, and calibration stability among all tested approaches, with performance gains most pronounced in the minority high-risk group while minimizing false positives. Table2.Model_Performance_Comparison_6_Models Model Accuracy Precision Recall F_meas Roc_auc XGBoost 0.924528302 0.888888889 0.727272727 0.8 0.917748918 Random Forest 0.79245283 0 0.88961039 SVM 0.79245283 0 0.848484848 Logistic Regression 0.811320755 0.538461538 0.636363636 0.583333333 0.826839827 Neural Network 0.773584906 0.466666667 0.636363636 0.538461538 0.825757576 KNN 0.811320755 0.666666667 0.181818182 0.285714286 0.752164502 3.3 SHAP Interpretation of the Full Model: High-Dose Parameters Play a Dominant Role Having established the superior predictive performance of the XGBoost model, we next employed the SHAP (SHapley Additive exPlanations) framework to interpret its internal decision-making logic on the independent test set. This analysis aimed to unlock the "black box" and identify the most critical predictors driving the risk of radiation-induced pericarditis. The global feature importance analysis, presented in Figure 3A, delivered a striking and unexpected insight. Contrary to traditional clinical focus, high-dose cardiac parameters emerged as the most influential predictors in the comprehensive model. Heart_v30 was unequivocally the most important feature, exhibiting a mean absolute SHAP value of approximately 0.155, which was nearly double that of the second-ranked feature, heart_v40 (mean absolute SHAP value ≈ 0.08). In stark contrast, the low-dose parameter heart_v5, the initial focus of this study, ranked 15th and contributed negligibly to the model's predictions, with a mean absolute SHAP value close to zero. To further dissect not just the magnitude but also the directional impact of these features, the SHAP summary (beeswarm) plot was examined (Figure 3B). This plot vividly illustrates a clear dose-response pattern for the top predictors. For heart_v30 and heart_v40, nearly all patients with high feature values (represented by red points) yielded positive SHAP values, indicating a strong push towards a higher predicted risk. Conversely, patients with low values for these parameters (blue points) consistently produced negative SHAP values, exerting a protective effect. This clear separation confirms that the model learned a robust rule: the larger the heart volume receiving high doses of radiation, the higher the risk of pericarditis. Again, the feature values for heart_v5 are shown tightly clustered around a SHAP value of 0, confirming its limited impact in the full model. Delving deeper, the SHAP dependence plots reveal the complex, non-linear nature of these relationships. The plot for heart_v30 (Figure 3C) shows that its impact on risk is highly non-linear. The SHAP value escalates sharply as the standardized heart_v30 value increases from approximately -0.6 to 0, reaching a peak contribution of approximately 0.42. This suggests the existence of a critical threshold; once the volume of heart tissue receiving 30 Gy exceeds this point, the risk of pericarditis accelerates dramatically. A similar, albeit less pronounced, non-linear trend was observed for heart_v40 (Figure 3D). In summary, the SHAP analysis of the full, all-inclusive model consistently points to a clear conclusion: high-dose cardiac exposure parameters (heart_v30 and heart_v40) are the dominant factors in predicting pericarditis risk. This unexpected finding raises a critical question that forms the central paradox of this study: Is the predictive value of the traditionally important heart_v5 parameter genuinely negligible, or is it being "masked" by the overwhelming statistical influence of the collinear, high-dose features? To definitively answer this question and isolate the true, independent contribution of heart_v5, a more focused analytical approach is warranted. 3.4 Validation of heart_v5's Independent Value via a Simplified Model To resolve the paradox of heart_v5's diminished importance in the full model and to isolate its true, independent predictive value, we implemented a "full vs. simplified model" analytical strategy. A simplified XGBoost model was developed, which deliberately excluded the dominant, highly collinear high-dose cardiac parameters (heart_v30, heart_v40, and heart_mean_dose) but retained all other clinical and dosimetric variables. The comparison of feature importance between the two models, presented in Figure 4A, provides a striking resolution to the paradox. In the simplified model, the importance of heart_v5 experienced a dramatic resurgence, leaping from a low rank of 15th in the full model to become the 3rd most important predictor overall. Its mean absolute SHAP value increased significantly to 0.067, confirming that its predictive signal was indeed being masked by statistical collinearity in the comprehensive model. This result provides powerful, direct evidence for the significant, independent predictive value of heart_v5. With the confounding influence of the high-dose parameters removed, the simplified model reveals the true nature of the relationship between heart_v5 and pericarditis risk. The SHAP dependence plot in Figure 4B now illustrates a clear and positive dose-response relationship. The fitted curve demonstrates that as the heart_v5 value increases, its contribution to the predicted risk (the SHAP value) also steadily increases, rising from approximately 0.05 for the lowest heart_v5 values to over 0.15 for the highest values. This confirms that, when its effect is properly isolated, a larger heart volume receiving low-dose radiation is unequivocally associated with a higher risk of pericarditis. To ensure this isolated effect was clinically meaningful, we validated the heart_v5 SHAP values from the simplified model against the patients' actual clinical outcomes (Figure 4E). A clear separation was observed: nearly all patients belonging to the true high-risk group (pink dots) registered positive SHAP values for heart_v5, while the majority of true low-risk patients (blue dots) had SHAP values near or below zero. This strong concordance demonstrates that the risk contribution assigned by the model based on heart_v5 aligns remarkably well with real-world clinical results. Crucially, this clarification was achieved with only a negligible trade-off in overall predictive accuracy. As shown in Figure 4F, the AUC of the simplified model (0.903) remained high, only slightly reduced from that of the full model (0.918). This confirms the simplified model itself is a robust and reliable tool for this analysis. In conclusion, the "full vs. simplified model" strategy successfully resolved the analytical paradox. By isolating heart_v5 from its collinear counterparts, we have rigorously demonstrated its significant and independent predictive value. The clear dose-response relationship and the strong correlation with clinical outcomes firmly establish heart_v5 as a reliable and clinically relevant biomarker for predicting radiation-induced pericarditis, paving the way for a more in-depth exploration of its clinical utility. 3.5In-depth Analysis of heart_v5 Impact using SHAP Values To elucidate the influence of the heart_v5 parameter on risk prediction, a detailed analysis was conducted using SHapley Additive exPlanations (SHAP) on the simplified model's predictions for the test set (n=53), with the results presented in Figure 5. A sina plot illustrating the distribution of heart_v5 SHAP values across all test patients showed marked heterogeneity in the feature's impact on individual predictions (Figure 5A). The SHAP values spanned a wide range, from approximately -0.34 to +0.26. The median SHAP value was positive, as indicated by the embedded boxplot, and a majority of patients exhibited SHAP values greater than zero. This confirmed that, for most patients in the model, a higher heart_v5 contributed to an increased predicted risk. Conversely, a smaller subset of patients displayed negative SHAP values, for whom the feature contributed to a lower predicted risk. The relationship between the original, non-standardized heart_v5 values (expressed as a percentage) and the predicted probability of high risk was examined (Figure 5B). A logistic regression curve fitted to the data showed a distinct non-linear relationship. Specifically, the predicted probability of a high-risk event remained low (at approximately 5%) for heart_v5 values below 25%. A notable inflection point was observed where the risk probability began to increase sharply for heart_v5 values exceeding 35-40%, reaching approximately 50% for a heart_v5 of 60%. Subgroup analyses were conducted to determine if the effect of heart_v5 was modified by prespecified clinical variables.When stratified by lesion position (Figure 5C), the impact of heart_v5 on predicted risk differed between patients with left-sided versus right-sided tumors. At higher standardized heart_v5 values (greater than 1.0), the SHAP values were consistently greater for patients with left-sided lesions compared to those with right-sided lesions, quantifying a stronger positive contribution to risk in the left-sided cohort. The analysis was also stratified by nodal (N) stage (Figure 5D). For patients in the N0-N1 group, the SHAP value demonstrated a relatively steady increase with rising heart_v5 values. In contrast, for patients in the advanced N2-N3 group, the relationship was more volatile, showing greater fluctuation across the range of heart_v5. Potential interaction effects between heart_v5 and two other dosimetric parameters, maximum heart distance (mhd) and ipsilateral lung volume (ipsi_lung_volume), were investigated visually (Figures 5E and 5F). The SHAP dependence plots for heart_v5 were colored by the standardized values of mhd and ipsi_lung_volume, respectively. In both plots, no clear pattern of vertical separation or distinct clustering of points was observed. This suggested that the predictive impact of heart_v5 within the model was largely independent of mhd and ipsi_lung_volume. In summary, the analysis of the test set (n=53) revealed that heart_v5 was a primary driver of predicted risk, with a heterogeneous impact that was significantly modulated by tumor laterality, showing a more pronounced risk contribution for left-sided tumors at high heart_v5 values. Discussion This retrospective cohort study evaluated whether low-dose cardiac exposure, quantified as heart_V5, independently predicts radiation-induced pericarditis in breast cancer patients receiving radiotherapy. The analysis used extreme gradient boosting (XGBoost) models with SHAP-based interpretability, comparing a full model including all dose–volume histogram (DVH) metrics and a simplified model excluding high-dose cardiac features. The study cohort comprised 277 patients (training n = 224, independent test n = 53). In the simplified model, heart_V5 emerged as the third most important predictor (mean absolute SHAP = 0.067) with a clear positive, non-linear association with pericarditis risk, achieving an area under the curve (AUC) of 0.903. In contrast, in the full model, high-dose parameters dominated—heart_V30 (mean |SHAP| ≈ 0.155) and heart_V40 (≈ 0.08) were the top predictors, with heart_V5 ranking 15th and contributing negligibly; the full-model AUC was 0.918 (95% CI, 0.848–0.987). Clinically, the predicted probability of pericarditis remained approximately 5% for heart_V5 < 25% and increased sharply above ~ 35–40%, reaching ~ 50% at heart_V5 ≈ 60%. These findings are consistent with prior evidence linking both high-dose and low-dose cardiac irradiation to late cardiovascular events[ 23 , 24 ]. The present findings suggest that heart_V5 may serve as a clinically relevant predictor of radiation-induced pericarditis, particularly in contexts where high-dose cardiac parameters are not available or are excluded from the model. The non-linear dose–response curve observed—with a relatively flat risk below ~ 25% followed by a steep increase beyond ~ 35–40%—aligns with the concept of a subthreshold region followed by a rapid escalation in risk once compensatory mechanisms are overwhelmed [ 25 ]. Although heart_V5 was overshadowed by high-dose parameters in the full model, its prominence in the simplified model indicates that low-dose exposure retains predictive value, particularly in scenarios such as advanced cardiac-sparing techniques where high-dose metrics may be minimized [ 25 ]. Comparison with prior literature shows both consistency and divergence. QUANTEC recommendations emphasize limiting high-dose volumes, e.g., heart_V30 < 46% to reduce pericarditis risk, which is consistent with the full-model finding that heart_V30 was the most important predictor. However, recent population-based studies have reported associations between lower-dose exposures and late cardiac events, even when high-dose parameters are controlled [ 24 , 26 ]. Our finding that heart_V5 predicted pericarditis risk in the absence of high-dose variables echoes these reports and extends them to an acute toxicity endpoint. Mechanistically, one possible explanation is that low-dose radiation may contribute to widespread microvascular injury and inflammation, triggering pericardial reactions even without focal high-dose injury [ 27 ]. This diffuse effect could explain why heart_V5 retained importance in the simplified model, despite the lower absolute dose. The SHAP analysis further supports this by showing that the model learned a steep risk gradient once heart_V5 exceeded ~ 35–40%, suggesting a threshold effect beyond which pericardial tolerance is compromised. The difference in predictive patterns between the full and simplified models also reflects the statistical interplay among correlated DVH metrics. High-dose parameters such as heart_V30 and heart_V40 are typically correlated with low-dose parameters, but their stronger direct association with injury may suppress the apparent importance of low-dose metrics in multivariate models [ 26 , 28 ]. By removing these high-dose variables, the simplified model allows the independent contribution of heart_V5 to emerge. From a clinical standpoint, the identified heart_V5 inflection point (~ 35–40%) is noteworthy. While current guidelines do not provide explicit low-dose constraints for the heart, incorporating such a threshold could be valuable in treatment planning, especially for patients with pre-existing cardiovascular risk factors [ 29 , 30 ]. Our results suggest that even when high-dose constraints are met, vigilance for low-dose exposure is warranted. Our AUC results indicate that both models performed well (AUC = 0.903 for simplified, 0.918 for full), suggesting that heart_V5 could be incorporated into predictive models without substantial loss of accuracy if high-dose parameters are unavailable. This is particularly relevant in multi-institutional datasets where DVH data granularity varies [ 31 ]. It also reinforces the idea that machine learning methods, coupled with interpretable models, can uncover clinically meaningful relationships that may be masked in conventional regression analyses [ 32 ]. This study establishes heart V5 as an independent and clinically significant predictor of radiation-induced pericarditis in breast cancer patients, particularly when high-dose parameters are unavailable. A potential risk threshold of ~ 35–40% for heart V5 could inform future treatment planning guidelines, complementing existing high-dose constraints [ 33 ]. Incorporating low-dose parameters into predictive models enables clinicians to enhance patient-specific risk assessment and optimize cardiac-sparing strategies. However, the study’s retrospective, single-institution design may limit generalizability, necessitating prospective multi-center validation to confirm heart V5 thresholds across diverse populations. Additionally, the limited number of pericarditis events may reduce statistical power for detecting subtle interactions, suggesting that future studies with larger event counts could refine threshold estimates. Conclusion This retrospective cohort study demonstrated that low-dose cardiac exposure (heart_V5) independently predicts radiation-induced pericarditis in breast cancer patients when high-dose parameters are excluded, with a non-linear dose-response relationship showing a sharp risk increase above 35–40%. In the full model, high-dose parameters (heart_V30 and heart_V40) dominated, but the simplified model’s high performance (AUC = 0.903) underscores heart_V5’s clinical relevance. These findings suggest that incorporating low-dose cardiac exposure metrics, particularly heart_V5, into radiotherapy planning enhances risk stratification and supports personalized cardiac-sparing strategies. By identifying a potential heart_V5 threshold of 35–40%, this study informs future guidelines for minimizing pericardial morbidity in breast cancer survivors. Machine-learning approaches, coupled with SHAP-based interpretability, offer a robust framework for uncovering clinically meaningful dosimetric predictors. Declarations Credit authorship contribution statement: Yongkai Lu: Writing – original draft, Methodology, Validation, Software, Project administration, Methodology, Data curation, Conceptualization. Lei Xu: Writing – review & editing, Methodology, Conceptualization, Validation. Jian Zhang: Methodology, Conceptualization, Validation acquisition. Rongze Ma: Visualization, Conceptualization, Validation. Yao Wang: Visualization, Software. Shanshan Gao: Supervision, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization. Funding: This program was supported by the Natural Science Basic Research Program of Shaanxi (2025JC-YBQN-1093). Ethics Statement: This study was approved by the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University. The research was conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to their inclusion in the study. Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: The data supporting the findings of this study are available from the First Affiliated Hospital of Xi’an Jiaotong University but are not publicly available due to patient privacy and ethical restrictions. Anonymized data may be made available upon reasonable request to the corresponding author (Yongkai Lu, [email protected] ), subject to approval by the institutional Ethics Committee. Conflicts of Interest: The authors declare no competing financial intere References Sharma, A., et al., Prognostic factors for local control and survival for inoperable pulmonary colorectal oligometastases treated with stereotactic body radiotherapy. Radiother Oncol, 2020. 144 : p. 23–29. Sung, H., et al., Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin, 2021. 71 (3): p. 209–249. Darby, S.C., et al., Risk of ischemic heart disease in women after radiotherapy for breast cancer. New England Journal of Medicine, 2013. 368 (11). Franco, V.I., et al., Cardiovascular effects in childhood cancer survivors treated with anthracyclines. Cardiol Res Pract, 2011. 2011 : p. 134679. Peeler, C.R., et al., Clinical evidence of variable proton biological effectiveness in pediatric patients treated for ependymoma. Radiother Oncol, 2016. 121 (3): p. 395–401. Ebctcg, et al., Effect of radiotherapy after mastectomy and axillary surgery on 10-year recurrence and 20-year breast cancer mortality: meta-analysis of individual patient data for 8135 women in 22 randomised trials. Lancet, 2014. 383 (9935): p. 2127–35. Early Breast Cancer Trialists’ Collaborative, G., Effects of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15-year survival: An overview of the randomised trials. The Lancet, 2005. 366 (9503). Early Breast Cancer Trialists’ Collaborative, G., Favourable and unfavourable effects on long-term survival of radiotherapy for early breast cancer: An overview of the randomised trials. The Lancet, 2000. 355 (9217). Taylor, C., et al., Estimating the risks of breast cancer radiotherapy: Evidence from modern radiation doses to the lungs and heart and from previous randomized trials. Journal of Clinical Oncology, 2017. 35 (15). Palmer, M., et al., Genome-Based Characterization of Biological Processes That Differentiate Closely Related Bacteria. Front Microbiol, 2018. 9 : p. 113. Mikell, J.L., et al., Similar survival for patients undergoing reduced-intensity total body irradiation (TBI) versus myeloablative TBI as conditioning for allogeneic transplant in acute leukemia. Int J Radiat Oncol Biol Phys, 2014. 89 (2): p. 360–9. Gu, H., et al., Genetic variants in the CNTNAP2 gene are associated with gender differences among dyslexic children in China. EBioMedicine, 2018. 34 : p. 165–170. Deasy, J.O., et al., Improving normal tissue complication probability models: the need to adopt a "data-pooling" culture. Int J Radiat Oncol Biol Phys, 2010. 76 (3 Suppl): p. S151–4. Bentzen, S.M., et al., Biomarkers and surrogate endpoints for normal-tissue effects of radiation therapy: the importance of dose-volume effects. Int J Radiat Oncol Biol Phys, 2010. 76 (3 Suppl): p. S145–50. Cho, S.S., et al., Clazosentan for Aneurysmal Subarachnoid Hemorrhage: An Updated Meta-Analysis with Trial Sequential Analysis. World Neurosurg, 2019. 123 : p. 418–424 e3. Liauw, S.L., et al., Salvage radiotherapy after postprostatectomy biochemical failure: does pretreatment radioimmunoscintigraphy help select patients with locally confined disease? Int J Radiat Oncol Biol Phys, 2008. 71 (5): p. 1316–21. Wei, X., et al., Risk factors for pericardial effusion in inoperable non-small cell lung cancer patients treated with definitive chemoradiation therapy. International Journal of Radiation Oncology, Biology, Physics, 2008. 71 (3). Siddiqui, F. and B. Movsas, Management of Radiation Toxicity in Head and Neck Cancers. Semin Radiat Oncol, 2017. 27 (4): p. 340–349. Eich, H.T., et al., Intensified Chemotherapy and Dose-Reduced Involved Field Radiotherapy in Patients with Early Unfavorable Hodgkin Lymphoma: Final Analysis of the German Hodgkin Study Group (GHSG) Randomized HD11 Trial. International Journal of Radiation Oncology*Biology*Physics, 2010. 78 (3). Emami, B., et al., Tolerance of normal tissue to therapeutic irradiation. Int J Radiat Oncol Biol Phys, 1991. 21 (1): p. 109–22. Gay, H.A. and A. Niemierko, A free program for calculating EUD-based NTCP and TCP in external beam radiotherapy. Phys Med, 2007. 23 (3-4): p. 115–25. Lim, T.Y., et al., Clinically Oriented Contour Evaluation Using Dosimetric Indices Generated From Automated Knowledge-Based Planning. Int J Radiat Oncol Biol Phys, 2019. 103 (5): p. 1251–1260. Xiao, Z., et al., Impact of heart motion on radiation dose in the heart and left ventricular myocardium during breast cancer treatment. Front Oncol, 2025. 15 : p. 1503131. Li, X., et al., Radiation-induced cardiac substructure damage and dose constraints: a review. Radiat Oncol, 2025. 20 (1): p. 94. Nuruddeen, M.G., et al., Cardiac substructure dose distributions in node-positive and node-negative breast cancer patients undergoing 3D-CRT: comparing the predictive accuracy of mean heart dose and mean left ventricular dose. Radiat Oncol, 2025. 20 (1): p. 65. Ricardi, U., V. Chiofalo, and G.C. Iorio, Modern Radiotherapy and Cardiac Toxicity in Breast Cancer. Clinical and Experimental Obstetrics & Gynecology, 2023. 50 (7). Yang, J., et al., A multicenter retrospective study of early cardiac toxicity in operable breast cancer patients receiving concurrent dual or mono anti-HER2 therapy with postoperative radiation therapy. Breast, 2025. 79 : p. 103879. Diaz-Gavela, A.A., et al., Breast Radiotherapy-Related Cardiotoxicity. When, How, Why. Risk Prevention and Control Strategies. Cancers (Basel), 2021. 13 (7). Liu, D.-W., et al., The risk of cardiovascular toxicity caused by cancer radiotherapy—a narrative review. Therapeutic Radiology and Oncology, 2022. 6 : p. 4–4. Mirestean, C.C., R.I. Iancu, and D.P.T. Iancu, Hypofractionated Whole-Breast Irradiation Focus on Coronary Arteries and Cardiac Toxicity-A Narrative Review. Front Oncol, 2022. 12 : p. 862819. Parlar, S., et al., Investigation of cardiac and pulmonary doses in patients with left sided breast cancer treated by radiotherapy with deep inspiration breath hold technique. International Journal of Radiation Research, 2022. 20 (2): p. 369–375. Zhou, R., et al., Hypofractionated Radiotherapy followed by Hypofractionated Boost with weekly concurrent chemotherapy for Unresectable Stage III Non-Small Cell Lung Cancer: Results of A Prospective Phase II Study (GASTO-1049). Int J Radiat Oncol Biol Phys, 2023. 117 (2): p. 387–399. Qian, D., et al., Definitive chemoradiotherapy versus neoadjuvant chemoradiotherapy followed by surgery in patients with locally advanced esophageal squamous cell carcinoma who achieved clinical complete response when induction chemoradiation finished: A phase II random. Radiother Oncol, 2022. 174 : p. 1–7. Additional Declarations No competing interests reported. Supplementary Files SupplementaryMaterials.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8105555","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":560474507,"identity":"ac8e5cd6-0075-4e51-8873-4b585f2771dc","order_by":0,"name":"Yongkai Lu","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA8ElEQVRIiWNgGAWjYBACPgiVwMDGzP/hgESFhJw8IS1scC3sDYYHLM5YGBs2EKuFgeeA8YHKtopEhgOEtEjkGD4u+JUmzyeRkHDg5jyJBMYG5oePbuDXYmw8sy/HsE0i4cDBmdsk8tgZ2IyNc/BrMZPm7algbJNIbDgsuU2imLGBh02aGC32bRLJDIf/zgFqPECMFp4fOYltPMcYDkg2EKOF51mxMW9DWnIbew/DAYljEsaGzQT8ws+evPExz59k2/nNPMwfJGrq5OTZmx8+xqeFgYHDgIGxDVmAGa9yEGB/wMDwh6CqUTAKRsEoGMkAAFZfSDcCgXsfAAAAAElFTkSuQmCC","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":true,"prefix":"","firstName":"Yongkai","middleName":"","lastName":"Lu","suffix":""},{"id":560474508,"identity":"4c098720-d484-49c1-bc52-9f0f51c57f85","order_by":1,"name":"Lei Xu","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Xu","suffix":""},{"id":560474512,"identity":"24966e7d-9c20-4fd4-b637-0e182fe23486","order_by":2,"name":"Jian Zhang","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Jian","middleName":"","lastName":"Zhang","suffix":""},{"id":560474514,"identity":"995624d4-ebf8-4a41-936c-21c57ed53b5a","order_by":3,"name":"Rongze Ma","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Rongze","middleName":"","lastName":"Ma","suffix":""},{"id":560474515,"identity":"78994b3a-9bcd-4e25-8038-d961daf5b89c","order_by":4,"name":"Yao Wang","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Yao","middleName":"","lastName":"Wang","suffix":""},{"id":560474517,"identity":"8c980249-02e6-4a51-9a2a-01435934d2b0","order_by":5,"name":"Shanshan Gao","email":"","orcid":"","institution":"First Affiliated Hospital of Xi'an Jiaotong University","correspondingAuthor":false,"prefix":"","firstName":"Shanshan","middleName":"","lastName":"Gao","suffix":""}],"badges":[],"createdAt":"2025-11-13 12:08:24","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8105555/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8105555/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98779016,"identity":"29b6e72a-6e44-4cde-b30e-4c7c9818304e","added_by":"auto","created_at":"2025-12-22 12:29:52","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":30753162,"visible":true,"origin":"","legend":"","description":"","filename":"Manuscript.docx","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/8c7f029b8f54d23a768c17f2.docx"},{"id":98753300,"identity":"0184eb3d-78fd-44aa-8195-152ac7b2d4e9","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8588,"visible":true,"origin":"","legend":"","description":"","filename":"025c2d2d2f934189911ec16c4efcb8ee.json","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/db6a0cd39644cb4f642febae.json"},{"id":98753305,"identity":"d85fb9b9-6c17-43e9-9dfe-ad9544e640e2","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":301240,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/760df242923921f5fd8975d1.docx"},{"id":98778447,"identity":"2b06c769-5ab9-4e50-bc9b-52dce167542a","added_by":"auto","created_at":"2025-12-22 12:29:15","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":121266,"visible":true,"origin":"","legend":"","description":"","filename":"025c2d2d2f934189911ec16c4efcb8ee1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/6fe3de1eb35c0a893cabf11c.xml"},{"id":98753309,"identity":"c40ab685-47ec-4260-a02a-aefdebc8ba60","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6976756,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/b1ff7381185102bba94726e8.jpeg"},{"id":98778562,"identity":"1d7f91a8-6e33-42b0-9a90-9914572361cf","added_by":"auto","created_at":"2025-12-22 12:29:27","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5599708,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/5e69d15d499ed623726d5e70.jpeg"},{"id":98753318,"identity":"904a599c-f7ef-4a20-99d9-5486d4ec4cab","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6411208,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/26fd98b33fd7619bde5e6fff.jpeg"},{"id":98753319,"identity":"e41e72ae-a2fa-4217-8305-78781c009d01","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6208348,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/3275f305ba84ed3069b9050f.jpeg"},{"id":98753317,"identity":"85cdabd5-6f5d-42ac-91bb-868ec5dfe054","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5462764,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/8a810264e7a7c12c282a757d.jpeg"},{"id":98753311,"identity":"f6a2d2d3-9a44-4e8a-a616-1edb95b8238d","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62394,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/9467d1595579c9501468263d.png"},{"id":98753310,"identity":"f3ff4a75-5a28-4571-a11a-8fc7430485e9","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":62293,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/829f38001ce22b3220857534.png"},{"id":98753314,"identity":"9f4b13b1-67c7-456a-a056-21d652b4ec77","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":87783,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/5add19fc65f4bce448d4ca9f.png"},{"id":98753315,"identity":"79c522fe-ba54-4b46-896c-cfe9adb1bc52","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":75016,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/c00d76193ef5e79f79784e4f.png"},{"id":98779874,"identity":"45f079fc-a397-44e0-8523-51b36b05dbcb","added_by":"auto","created_at":"2025-12-22 12:30:52","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":73842,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/97aff5d31b4006fb9bb3d2a3.png"},{"id":98753322,"identity":"b905b84a-20e3-4eb6-9b72-92c82fcd834d","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"xml","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":120561,"visible":true,"origin":"","legend":"","description":"","filename":"025c2d2d2f934189911ec16c4efcb8ee1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/e391870f472ae0d9723d05e8.xml"},{"id":98778052,"identity":"60fa5e03-0ac0-4534-ab41-8f9d1d5c9154","added_by":"auto","created_at":"2025-12-22 12:28:50","extension":"html","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":132429,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/2e063716ff2a3caf9a58e421.html"},{"id":98777230,"identity":"fcfaccf0-7ae5-4ae5-ba4b-a1ed095547de","added_by":"auto","created_at":"2025-12-22 12:26:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":630585,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of Patient Characteristics and Variable Relationships in the Test Set (n=53). The figure illustrates (A) the distribution of patients into high-risk (20.8%) and low-risk (79.2%) groups; (B) a heatmap of key clinical and dosimetric variables stratified by outcome; (C, D) distributions of key demographic, clinical (T/N stage, lesion position), and dosimetric (Heart V5) variables; (E) a boxplot comparing Heart V5 values between risk groups (Wilcoxon, p=0.53); and (F) a correlation matrix of numeric variables.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/3da5a1261fa4cf7f9f754973.png"},{"id":98753302,"identity":"3fc3fc86-47c1-4c7b-91d3-a31518f07be2","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":465513,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance Comparison of Six Machine Learning Models on the Independent Test Set. (A) ROC curves comparing the diagnostic ability of the models.Individual model ROC curves are presented in Supplementary Figure 1. (B) Precision-Recall (PR) curves evaluating model performance on the imbalanced dataset. (C) A radar chart providing a multi-dimensional comparison of model performance across five metrics (Accuracy, F1-score, Recall, Precision, ROC-AUC). (D) Confusion matrix for the top-performing XGBoost model. (E) Confusion matrix for the baseline Logistic Regression model. (F) Learning curve for the XGBoost model, plotting training set and test set Area Under the Curve (AUC) as a function of training size to assess for overfitting.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/e699b50439b753287f7f4fdd.png"},{"id":98779159,"identity":"405e16f2-dcd4-476e-bab5-8fa5282eb62b","added_by":"auto","created_at":"2025-12-22 12:30:00","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":609754,"visible":true,"origin":"","legend":"\u003cp\u003eInterpretation of the Full XGBoost Model Using SHapley Additive exPlanations (SHAP) on the Test Set. (A) A global feature importance plot ranking the top 20 predictors by their mean absolute SHAP value. (B) A SHAP summary (beeswarm) plot illustrating the magnitude and direction of each feature's impact on individual predictions, where color indicates the feature's original value. (C-F) SHAP dependence plots showing the non-linear relationship between the original value of a specific feature (x-axis) and its impact on the model prediction (SHAP value, y-axis) for the top dosimetric predictors: (C) heart_v30, (D) heart_v40, (E) heart_max_dose, and (F) mhd.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/f885d93fe08041f036432bbe.png"},{"id":98780402,"identity":"f171b617-9ca5-4ec9-8382-2b1b2584892f","added_by":"auto","created_at":"2025-12-22 12:31:18","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":545675,"visible":true,"origin":"","legend":"\u003cp\u003eValidation of heart_v5's Independent Predictive Value Using a Simplified Model. (A) Comparison of feature importance rankings, showing the dramatic increase in the importance of heart_v5 in the simplified model after collinear high-dose features were removed. (B) SHAP dependence plot from the simplified model, revealing a clear positive relationship between heart_v5 values and their contribution to risk. (C, D) SHAP waterfall plots for a single patient, illustrating the change in predicted risk from an original to an optimized plan. (E) A scatter plot demonstrating that the model's heart_v5 SHAP values correlate well with patients' actual clinical outcomes (high-risk vs. low-risk). (F) Bar chart comparing the Area Under the Curve (AUC), confirming the simplified model maintains high predictive accuracy.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/42769d7241e4c08abdda51d1.png"},{"id":98753307,"identity":"dcfd81a1-a3aa-47d4-b712-92b3a1d8f1d3","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":466037,"visible":true,"origin":"","legend":"\u003cp\u003eIn-depth SHAP Analysis of heart_v5 Using the Simplified Model. This figure explores the heterogeneity, clinical thresholds, and interaction effects of heart_v5 on the test set. (A) A Sina plot shows the wide distribution of heart_v5 SHAP values, indicating its heterogeneous impact on individual risk predictions. (B) A scatter plot of predicted high-risk probability versus original heart_v5 values, with a logistic fit suggesting a clinical risk threshold. (C, D) SHAP dependence plots for heart_v5 stratified by (C) lesion position and (D) N stage, revealing that the feature's impact is modulated by clinical context. (E, F) SHAP dependence plots exploring interactions between heart_v5 and (E) maximum heart distance (mhd) and (F) ipsilateral lung volume (ipsi_lung_volume).\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/a631422c2748585987bff727.png"},{"id":104403218,"identity":"2c31c226-8cee-4454-8702-a3b0db96dbe8","added_by":"auto","created_at":"2026-03-11 12:17:45","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3868809,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/6db9ea01-500b-47c6-9e3c-dd68cfa71f06.pdf"},{"id":98753303,"identity":"4fb3af2b-ed2b-4783-a5d5-fde3eeb6d2c3","added_by":"auto","created_at":"2025-12-22 09:21:24","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":301240,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryMaterials.docx","url":"https://assets-eu.researchsquare.com/files/rs-8105555/v1/e9e23dccc23ac354ccee51a3.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Low-Dose Cardiac Exposure (Heart V5) as an Independent Predictor of Radiation-Induced Pericarditis in Breast Cancer: An Interpretable Machine Learning Study","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn breast cancer treatment, radiation oncology has long focused on the risks of high-dose cardiac irradiation, yet emerging evidence suggests that low-dose radiation, particularly heart V5 (the percentage of heart volume receiving\u0026thinsp;\u0026ge;\u0026thinsp;5 Gy), also contributes to long-term cardiac harm[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]. Radiation-induced heart disease (RIHD), with pericarditis as an early and frequent manifestation, poses a significant challenge in breast cancer radiotherapy (RT). With over 2.3\u0026nbsp;million women diagnosed annually and approximately half receiving RT, RIHD drives long-term morbidity and mortality [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Pericarditis, characterized by inflammation or effusion, can emerge acutely or years after RT, leading to chest pain or, in severe cases, cardiac tamponade. Research indicates a 7.4% rise in major adverse cardiac events (MACE) per Gray of mean heart dose, contributing to 1\u0026ndash;2% excess mortality among survivors [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The economic toll of RIHD is substantial, with high healthcare costs and diminished quality of life [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. As five-year breast cancer survival now exceeds 90%, reducing toxicities like pericarditis is essential for improving survivorship and advancing precision oncology[\u003cspan additionalcitationids=\"CR6\" citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAdvances in RIHD research reveal a clear dose-dependent relationship between cardiac radiation exposure and adverse outcomes. The Early Breast Cancer Trialists\u0026rsquo; Collaborative Group demonstrated that while RT reduces breast cancer mortality, it increases cardiovascular deaths, with relative risks of 1.27\u0026ndash;1.61 for left-sided versus right-sided RT [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Darby et al.\u0026rsquo;s study of over 2,000 women further confirmed a linear increase in MACE with mean heart dose [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Diagnostic tools, such as echocardiography and cardiac magnetic resonance imaging (MRI), effectively detect pericardial effusions, while heart-sparing techniques like deep inspiration breath-hold (DIBH) and intensity-modulated radiotherapy (IMRT) reduce mean heart dose by 20\u0026ndash;50% [\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Dose-volume histogram parameters, including heart V30 and V40 (volumes receiving\u0026thinsp;\u0026ge;\u0026thinsp;30 Gy and \u0026ge;\u0026thinsp;40 Gy), are established predictors of pericarditis, with QUANTEC guidelines recommending V30\u0026thinsp;\u0026lt;\u0026thinsp;46% and mean heart dose\u0026thinsp;\u0026lt;\u0026thinsp;26 Gy to keep risk below 15% [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. However, the independent role of heart V5 remains uncertain, as its predictive value is often obscured by collinearity with high-dose metrics like V30 and V40. This collinearity arises because tangential RT fields, commonly used in breast cancer, create overlapping low- and high-dose regions in the heart [\u003cspan additionalcitationids=\"CR16\" citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], hindering the development of optimized dose constraints and personalized RT planning.\u003c/p\u003e \u003cp\u003eTo address this gap, advanced machine learning approaches offer a powerful solution for disentangling collinear dosimetric parameters and uncovering the independent predictive value of heart V5. Therefore, this study aims to determine whether heart V5 independently predicts radiation-induced pericarditis risk in breast cancer patients undergoing RT or if its contribution is masked by high-dose cardiac parameters. We conducted a retrospective cohort study of 277 female breast cancer patients receiving RT between January 2022 and June 2023, employing machine learning to enhance risk prediction and inform cardiac-sparing strategies.\u003c/p\u003e \u003cp\u003eIn this paper, we make several key contributions. First, we demonstrated heart_v5's independent predictive value for pericarditis risk using a simplified XGBoost model. Second, we revealed high-dose cardiac parameters (heart_v30, heart_v40) as dominant predictors in the full model. Third, we established a non-linear dose-response relationship for heart_v5 with clinical outcomes. Notably, the simplified XGBoost model, excluding collinear high-dose parameters, achieved an AUC of 0.887, with heart_v5 ranking as the 4th most important predictor and showing a clear positive dose-response relationship with pericarditis risk. These findings underscore the untapped potential of low-dose metrics in refining RIHD risk models, paving the way for targeted interventions that could mitigate long-term cardiovascular morbidity in breast cancer survivors.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\n \u003ch2\u003e2.1 Study design and participants\u003c/h2\u003e\n \u003cp\u003eA retrospective cohort study was conducted including female breast cancer patients who received radiotherapy at the Department of Radiation Oncology, First Affiliated Hospital of Xi\u0026apos;an Jiaotong University, Shaanxi Province, China, between January 2022 and June 2023. Patient data were extracted from medical records and included age, body mass index (BMI), history of hypertension, diabetes mellitus, alcohol consumption, menopausal status, surgical procedure, tumor stage (TNM), anemia status, and electrocardiogram (ECG) results. Radiotherapy data were obtained from the Elekta Monaco treatment planning system and comprised target volume, radiotherapy technique, tumor location, total dose to the planning target volume (PTV), PTV volume, PTV dose coverage, ipsilateral heart volume, ipsilateral lung volume, and maximum heart distance (MHD)[\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e]. Radiation target volumes were delineated by two senior radiation oncologists following Radiation Therapy Oncology Group (RTOG) guidelines. Treatment plans were designed by a senior medical physicist and finalized by consensus between oncologists and physicist[\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003eThe study was approved by the Ethics Committee of the First Affiliated Hospital of Xi\u0026apos;an Jiaotong University and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to inclusion. Of 323 patients who consented, 46 were excluded due to missing data, bilateral breast cancer, or male breast cancer, resulting in a final cohort of 277 patients.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\n \u003ch2\u003e2.2 Assessment of pericarditis\u003c/h2\u003e\n \u003cp\u003eThe primary endpoint was radiation-induced pericarditis incidence, assessed using a Normal Tissue Complication Probability (NTCP) model. This model predicts complication risk based on dose-volume histograms (DVHs) and equivalent uniform dose (EUD) parameters. The NTCP model was based on Emami et al. and implemented via MATLAB following Gay and Niemierko et al. The logistic function used was:\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u003cimg src=\"https://myfiles.space/user_files/58895_8739fc6c57c1c19a/58895_custom_files/img1766062144.png\" width=\"342\" height=\"124\"\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\n \u003cp\u003ewhere TD50 = 50 Gy and \u0026gamma;50 = 3[20, 21]. EUD values were computed from DVHs to calculate NTCP for each patient. NTCP values were binarized: values \u0026ge; 3.8487 \u0026times; 10⁻⁶ were classified as high risk (1), and values below as low risk (0). The threshold was selected based on data distribution to separate approximately the highest 25% of risk values.\u003c/p\u003e\n \u003ch2\u003e2.3 Assessment of V5\u003c/h2\u003e\n \u003cp\u003eCardiac V5 was defined as the percentage volume of the heart receiving a radiation dose of 5 Gy or higher. This parameter was extracted from dose-volume histograms (DVHs) generated by the Elekta Monaco treatment planning system, based on heart contours delineated according to institutional protocols. V5 quantifies low-dose radiation exposure to the heart, which is relevant for assessing radiation-induced cardiac toxicity risk[\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e, \u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e].\u003c/p\u003e\n \u003cp\u003eTo facilitate statistical analysis, V5 values were divided into three categories: Low (\u0026le;\u0026thinsp;27%), Medium (27\u0026ndash;39%), and High (\u0026gt;\u0026thinsp;39%). These thresholds corresponded approximately to the 33.33rd and 66.67th percentiles of the V5 distribution in the study population. This percentile-based stratification ensured roughly equal sample sizes across groups, enabling balanced comparisons and improving statistical power for detecting dose-dependent effects on pericarditis risk.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4 Assessment of other variables\u003c/h2\u003e\n \u003cp\u003eAdditional clinical and dosimetric variables were collected from medical records and treatment planning data, including: age (years), BMI (kg/m\u0026sup2;), histories of hypertension, diabetes, alcohol consumption, menopausal status (pre/post), surgical procedure (breast conservation, mastectomy, reconstruction), tumor stage (T, N, M0), ECG results (normal/abnormal), anemia status (present/absent), radiotherapy target areas (supraclavicular, axillary, internal mammary lymph nodes, or none), radiotherapy techniques (IMRT, VMAT), lesion laterality (left/right), PTV total dose (\u0026lt;\u0026thinsp;50 Gy, \u0026ge;\u0026thinsp;50 Gy), PTV volume (cm\u0026sup3;), number of fractions (\u0026lt;\u0026thinsp;25, \u0026ge;\u0026thinsp;25), PTV dose coverage (90\u0026ndash;95%, \u0026ge;\u0026thinsp;95%), ipsilateral lung volume (cm\u0026sup3;), heart volume (cm\u0026sup3;), and MHD (cm). MHD was defined as the maximum perpendicular distance from the anterior heart surface to the posterior tangential field edge.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\n \u003ch2\u003e2.4 Statistical Analysis\u003c/h2\u003e\n \u003cp\u003eDescriptive statistics were used to summarize baseline patient characteristics. Continuous variables were presented as median and interquartile range (IQR), while categorical variables were presented as frequencies and percentages. To compare baseline characteristics across heart_v5 tertiles (defined for this analysis as Low: \u0026le;27%; Medium: 27\u0026ndash;39%; High: \u0026gt;39%), the Kruskal-Wallis test was used for continuous variables and the Chi-squared (\u0026chi;\u0026sup2;) test was used for categorical variables. The Wilcoxon rank-sum test was used to compare continuous variables between the final high-risk and low-risk groups. Pearson\u0026rsquo;s correlation coefficient (r) was calculated to assess the degree of multicollinearity between key numeric dosimetric variables. A two-sided p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant for these analyses. All statistical analyses were performed using R (version 4.2.2).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n \u003ch2\u003e2.5 Machine Learning Model Development and Validation\u003c/h2\u003e\n \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\n \u003ch2\u003e2.5.1 Data Partitioning and Preprocessing\u003c/h2\u003e\n \u003cp\u003eFor model development, the full cohort (n\u0026thinsp;=\u0026thinsp;277) was partitioned into a training set (80.9%, n\u0026thinsp;=\u0026thinsp;224) and an independent test set (19.1%, n\u0026thinsp;=\u0026thinsp;53) using a stratified split based on the binary outcome variable. A standardized preprocessing pipeline was constructed using the tidymodels package, which included median/mode imputation for missing data, one-hot encoding for all categorical predictors, and normalization (scaling to a mean of zero and standard deviation of one) of all numeric predictors.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\n \u003ch2\u003e2.5.2 Model Training and Evaluation\u003c/h2\u003e\n \u003cp\u003eSix machine learning algorithms were evaluated: Logistic Regression (LR), Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Multilayer Perceptron (MLP), and Extreme Gradient Boosting (XGBoost). For models with tunable hyperparameters, a 5-fold cross-validation was performed on the training data, with the area under the receiver operating characteristic curve (AUC) used as the metric to select the optimal parameters. The final, tuned models were evaluated on the independent test set using AUC, accuracy, precision, recall, and F1-score. The DeLong test was used to compare the AUCs of different models.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\n \u003ch2\u003e2.5.3 Model Interpretation and heart_v5 Validation\u003c/h2\u003e\n \u003cp\u003eThe internal logic of the best-performing XGBoost model was interpreted using SHapley Additive exPlanations (SHAP), implemented with the fastshap R package. SHAP values were calculated to quantify the impact of each feature on the model\u0026apos;s predictions for every patient in the test set. Model behavior was visualized using SHAP global importance plots, summary (beeswarm) plots, and dependence plots.\u003c/p\u003e\n \u003cp\u003eTo address high multicollinearity among cardiac dose parameters and isolate the independent predictive value of heart_v5, a \u0026quot;simplified\u0026quot; XGBoost model was developed. This model was trained on a dataset from which the highly collinear features (heart_v30, heart_v40, heart_mean_dose) were deliberately excluded. This simplified model was then trained and interpreted using the same SHAP methodology to assess the unmasked impact of heart_v5.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e3.1 Patient Baseline Characteristics and Data Overview\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe final analysis included 277 female breast cancer patients. Patients were classified into three groups according to tertiles of cardiac V5: Low (\u0026le;27%, n=89), Medium (27\u0026ndash;39%, n=96), and High (\u0026gt;39%, n=92) (Table 1). Across the three groups, baseline demographic variables were comparable. Median age was 47.6 years (interquartile range [IQR] 43.2\u0026ndash;51.0) in the Low group, 47.8 years (IQR 43.1\u0026ndash;51.2) in the Medium group, and 47.4 years (IQR 42.8\u0026ndash;50.8) in the High group, with no statistically significant difference (p=0.861, Kruskal\u0026ndash;Wallis test). Median body mass index (BMI) was 22.9 kg/m\u0026sup2; (IQR 21.3\u0026ndash;24.4), 23.2 kg/m\u0026sup2; (IQR 21.4\u0026ndash;24.7), and 23.0 kg/m\u0026sup2; (IQR 21.5\u0026ndash;24.6), respectively (p=0.823). The prevalence of hypertension was 10.1% (n=9), 9.4% (n=9), and 10.9% (n=10) in the Low, Medium, and High groups, respectively (p=0.952, \u0026chi;\u0026sup2; test). Diabetes was present in 6.7% (n=6), 5.2% (n=5), and 7.6% (n=7) of patients (p=0.815). Alcohol consumption was reported by 7.9% (n=7), 7.3% (n=7), and 8.7% (n=8) of patients (p=0.941).\u003c/p\u003e\n\u003cp\u003eSignificant group differences were observed in tumor laterality and nodal stage. Left-sided tumors were more frequent in the High V5 group (20.9%, n=19) than in the Low V5 group (11.6%, n=10) (p=0.001). Patients in the High V5 group had higher rates of N1 nodal stage (15.9%, n=15) compared with the Low group (11.2%, n=10) (p=0.009). Treatment field coverage also differed significantly: supraclavicular lymph node irradiation was delivered in 26.7% (n=25) of High V5 patients compared with 16.2% (n=14) in the Low group (p\u0026lt;0.001), and internal mammary node irradiation in 27.8% (n=26) versus 18.1% (n=16) (p\u0026lt;0.001).\u003c/p\u003e\n\u003cp\u003eThe independent test set (n=53) reflected a similar clinical profile (Figure 1). The low-risk group comprised 79.2% (n=42) of the test set and the high-risk group 20.8% (n=11) (Figure 1A). Visual inspection of the heatmap suggested that high-risk patients more often had elevated heart_v5 and BMI values. The distribution of heart_v5 across the test cohort was right-skewed, with a median of\u0026nbsp;31.5%\u0026nbsp;(IQR\u0026nbsp;21.9%\u0026ndash;42.2%) and a range from\u0026nbsp;0.17% to 58.5% (Figure 1B). Tumor staging in the test set was dominated by T3 stage (n=30,\u0026nbsp;56.6%) and N1/N2 stage (n=48,\u0026nbsp;90.6%). Right-sided lesions were more frequent (n=29,\u0026nbsp;54.7%) than left-sided lesions (n=24,\u0026nbsp;45.3%) (Figure 1C, 1D).\u003c/p\u003e\n\u003cp\u003eIn univariate analysis, median heart_v5 was higher in high-risk patients (34.5%, IQR\u0026nbsp;27.5\u0026ndash;38.2%) than in low-risk patients (29.5%, IQR 25.0\u0026ndash;33.8%), although this difference was not statistically significant (p=0.530, Wilcoxon rank-sum test; effect size r=0.12, 95% confidence interval [CI] \u0026minus;0.10 to 0.33).Pearson correlation analysis of dosimetric parameters demonstrated strong multicollinearity: heart_v5 was positively correlated with heart_mean_dose (r=0.75, p\u0026lt;0.001) (Figure 1F). These findings indicate substantial overlap in the information carried by low-dose and mean heart dose metrics.\u003c/p\u003e\n\u003cp\u003eOverall, the baseline data show that while demographic factors were balanced across V5 tertiles, tumor laterality, nodal stage, and regional lymph node irradiation differed significantly. The test set preserved these distributions and highlighted both the skewed nature of heart_v5 and its high correlation with other dosimetric measures.\u0026nbsp;\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"605\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 605px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTable 1 Baseline Characteristics of Study Subjects by heart V5 Indicators\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eCharacteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003eClustering of cardiac V5 indicators\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eLow heart V5 (\u0026le;27)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003eMedium heart V5(27-39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eHigh heart V5(\u0026gt;39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eNo. of subjects\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e96\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e92\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eAge (years)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e47.6 (39.0,54.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e47.8 (39.2,55.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e47.4 (39.0,53.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.861\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eBMI (kg/m\u003csup\u003e2\u003c/sup\u003e)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e22.9 (0.6, 24.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e23.2 (20.5, 25.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e23.0 (21.3, 24.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.823\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eHistory of hypertension Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e1.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e1.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.440\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eHistory of diabetes Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e1.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.572\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eHistory of alcohol consumption Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.743\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eMenopausal Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e6.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e6.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e5.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.545\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eSurgical procedure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eBreast Conservation Surgery\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e18.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e13.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e13.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eMastectomy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e12.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e18.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e14.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.147\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eReconstruction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e1.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e3.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e6.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eT staging\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eTis\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e1.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e0.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.151\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eT1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e20.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e15.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.041\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eT2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e10.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e16.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e13.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.073\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eT3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e1.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e2.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.042\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eT4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.821\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eN staging\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eN0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e19.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e8.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e12.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eN1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e11.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e19.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e15.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.009\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eN2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e1.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e4.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e2.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.120\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eN3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e2.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e1.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.420\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eM staging\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eM0 Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e31.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e34.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e33.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.346\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eNormal ECG Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e30.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e34.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e31.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.140\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eAnemic Yes（%）\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e1.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e1.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.903\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eRadiotherapy target\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eSupraclavicular Lymph Nodes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e16.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e28.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e26.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eAxillary Lymph Nodes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e18.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e29.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e27.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eInternal Mammary Lymph Nodes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e18.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e29.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e27.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eNo lymph nodes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e0.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e0.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.841\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eRadiotherapy techniques\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eIMRT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e6.500\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e8.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e6.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.794\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eVMAT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e25.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e26.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e26.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.796\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"605\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u003cem\u003eContinued\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eCharacteristics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003eLow heart V5 (\u0026le;27)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003eMedium heart V5(27-39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eHigh heart V5(\u0026gt;39)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eLesion position\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eLeft side\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e11.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e14.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e20.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eRight side\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e20.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e20.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e11.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003ePTV total dose\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026lt;50Gy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e9.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e4.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e2.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026ge;50Gy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e22.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e30.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e30.30\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003ePTV volume\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e688.2(527.4,872.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e746.9(611.3,919.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e766.5(632.6,1020.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.000\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eFractions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026lt;25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e9.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e2.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e2.20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026ge;25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e23.10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e31.80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e31.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003ePTV dose coverage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"4\" valign=\"top\" style=\"width: 473px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e90%-95%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e4.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e9.00\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e6.90\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.158\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e\u0026ge;95%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e27.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e25.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e26.40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.158\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eIpsilateral lung volume\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e1397.1(1201.9,1745.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e1359.3(1219.9,1559.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e1250(1114.9,1456.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.003\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eHeart volume\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e533.2(467.9,609.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e560.9(486.5,631.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e544.5(486.3,644.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.217\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003eMHD (cm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 142px;\"\u003e\n \u003cp\u003e10.0(5.0,11.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 152px;\"\u003e\n \u003cp\u003e10.0(5.25,11.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 132px;\"\u003e\n \u003cp\u003e7.0(5.62,10.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 47px;\"\u003e\n \u003cp\u003e0.012\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\" valign=\"top\" style=\"width: 605px;\"\u003e\n \u003cp\u003eNotes: Continuous variables were present as median (25%, 75% quantiles) and categorical variables are shown as percentages. Kruskal\u0026ndash;Wallis test (comparison of \u0026gt;2 groups) for continuous and ordinal distributed variables and the chi-squared test for categorical variables. BMI, body mass index; ECG, electrocardiogram; IMRT, intensity-modulated radiation therapy; VMAT, volumetric modulated arc therapy; PTV, planning target volume MHD, maximum heart distance.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003e\u003cstrong\u003e3.2 Machine Learning Model Performance in Independent Test Set\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSix predictive models\u0026mdash;logistic regression (LR), random forest (RF), support vector machine (SVM), k-nearest neighbors (KNN), multilayer perceptron (MLP), and extreme gradient boosting (XGBoost)\u0026mdash;were trained on the development cohort and evaluated on the independent test set (n=53). Performance metrics are summarized in Table 2, with graphical comparisons in Figure 2. Metrics reported include accuracy, area under the receiver operating characteristic curve (AUC), precision, recall, and F1-score.\u003c/p\u003e\n\u003cp\u003eIn receiver operating characteristic (ROC) analysis, XGBoost achieved the highest AUC at 0.918 (95% confidence interval CI 0.848\u0026ndash;0.987), exceeding LR (AUC=0.827, 95% CI 0.713\u0026ndash;0.940), RF (AUC=0.890, 95% CI 0.769\u0026ndash;0.960), SVM (AUC=0.848, 95% CI 0.758\u0026ndash;0.950), KNN (AUC=0.752, 95% CI 0.681\u0026ndash;0.921), and MLP (AUC=0.826, 95% CI 0.789\u0026ndash;0.965) (Figure 2A). Individual receiver operating characteristic (ROC) curves for each model are provided in Supplementary Figure 1. Pairwise DeLong tests indicated that XGBoost significantly outperformed LR (p=0.041) and KNN (p=0.018) in AUC.\u003c/p\u003e\n\u003cp\u003ePrecision\u0026ndash;recall (PR) analysis, more informative for imbalanced datasets, showed that XGBoost had the largest average precision (0.861), followed by RF (0.824), MLP (0.811), SVM (0.792), LR (0.713), and KNN (0.698) (Figure 2B). This corresponded to higher precision when identifying the minority high-risk class.\u003c/p\u003e\n\u003cp\u003eThe radar plot (Figure 2C) displayed a balanced performance profile for XGBoost across metrics: accuracy 0.925, AUC 0.918, precision 0.889, recall 0.727, and F1-score 0.800. RF achieved\u0026nbsp;accuracy 0.792,\u0026nbsp;precision not applicable,\u0026nbsp;recall 0.0, and\u0026nbsp;F1-score 0.0. LR had lower accuracy (0.811) and recall (0.636) with an F1-score of 0.583.\u003c/p\u003e\n\u003cp\u003eConfusion matrix comparison demonstrated that XGBoost correctly identified 8 of 11 high-risk patients (true positives), with 3 false negatives, 1 false positive, and 41 true negatives. This yielded a recall of 72.7%, specificity of 97.6%, and precision of 88.9%. In contrast, LR detected 7 of 11 high-risk patients, with 4 false negatives, 6 false positives, and 36 true negatives, resulting in a recall of 63.6%, specificity of 85.7%, and precision of 53.8% (Figures 2D, 2E).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe learning curve for the XGBoost model\u0026apos;s\u0026nbsp;AUC\u0026nbsp;is presented in Figure 2F. The curve shows that as the training set size increases, the\u0026nbsp;Test Set AUC\u0026nbsp;(validation AUC) rises to a final value near\u0026nbsp;1.0. The\u0026nbsp;Train Set AUC\u0026nbsp;(training AUC) stabilizes around\u0026nbsp;0.91. The gap between the final training and validation AUC scores suggests the model may exhibit some overfitting, although the validation performance remains high.\u003c/p\u003e\n\u003cp\u003eSubgroup performance was consistent across left- and right-sided tumor patients. In left-sided cases (n=21), XGBoost achieved AUC 0.924 and accuracy 0.905; in right-sided cases (n=25), AUC 0.912 and accuracy 0.936 were observed. For patients receiving internal mammary node irradiation (n=15), XGBoost maintained high discrimination (AUC=0.917) with recall 0.733, while in those without internal mammary irradiation (n=41), AUC was 0.919 with recall 0.724.\u003c/p\u003e\n\u003cp\u003eAcross all models, F1-scores ranged from 0.583 (LR) to 0.800 (XGBoost), and balanced accuracy ranged from 0.732 (KNN) to 0.852 (XGBoost). Matthews correlation coefficient (MCC) was highest for XGBoost at 0.783, compared with 0.604 for LR and 0.694 for RF. Cohen\u0026rsquo;s kappa values mirrored this pattern (XGBoost \u0026kappa;=0.791; LR \u0026kappa;=0.603; RF \u0026kappa;=0.702).\u003c/p\u003e\n\u003cp\u003eOverall, in the independent test set, the XGBoost algorithm demonstrated the highest discrimination, precision, and calibration stability among all tested approaches, with performance gains most pronounced in the minority high-risk group while minimizing false positives.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable2.Model_Performance_Comparison_6_Models\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\" width=\"583\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eModel\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eF_meas\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRoc_auc\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.924528302\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.888888889\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.727272727\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.917748918\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.79245283\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.88961039\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003eSVM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.79245283\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.848484848\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.811320755\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.538461538\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.636363636\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.583333333\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.826839827\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003eNeural Network\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.773584906\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.466666667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.636363636\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.538461538\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.825757576\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 106px;\"\u003e\n \u003cp\u003eKNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.811320755\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.666666667\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.181818182\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.285714286\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 95px;\"\u003e\n \u003cp\u003e0.752164502\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.3 SHAP Interpretation of the Full Model: High-Dose Parameters Play a Dominant Role\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHaving established the superior predictive performance of the XGBoost model, we next employed the SHAP (SHapley Additive exPlanations) framework to interpret its internal decision-making logic on the independent test set. This analysis aimed to \u0026nbsp; \u0026nbsp;unlock the \u0026quot;black box\u0026quot; and identify the most critical predictors driving the risk of radiation-induced pericarditis. The global feature importance analysis, presented in Figure 3A, delivered a striking and unexpected insight. Contrary to traditional clinical focus, high-dose cardiac parameters emerged as the most influential predictors in the comprehensive model. Heart_v30 was unequivocally the most important feature, exhibiting a mean absolute SHAP value of approximately 0.155, which was nearly double that of the second-ranked feature, heart_v40 (mean absolute SHAP value \u0026asymp; 0.08). In stark contrast, the low-dose parameter heart_v5, the initial focus of this study, ranked 15th and contributed negligibly to the model\u0026apos;s predictions, with a mean absolute SHAP value close to zero.\u003c/p\u003e\n\u003cp\u003eTo further dissect not just the magnitude but also the directional impact of these features, the SHAP summary (beeswarm) plot was examined (Figure 3B). This plot vividly illustrates a clear dose-response pattern for the top predictors. For heart_v30 and heart_v40, nearly all patients with high feature values (represented by red points) yielded positive SHAP values, indicating a strong push towards a higher predicted risk. Conversely, patients with low values for these parameters (blue points) consistently produced negative SHAP values, exerting a protective effect. This clear separation confirms that the model learned a robust rule: the larger the heart volume receiving high doses of radiation, the higher the risk of pericarditis. Again, the feature values for heart_v5 are shown tightly clustered around a SHAP value of 0, confirming its limited impact in the full model.\u003c/p\u003e\n\u003cp\u003eDelving deeper, the SHAP dependence plots reveal the complex, non-linear nature of these relationships. The plot for heart_v30 (Figure 3C) shows that its impact on risk is highly non-linear. The SHAP value escalates sharply as the standardized heart_v30 value increases from approximately\u0026nbsp;-0.6 to 0, reaching a peak contribution of approximately\u0026nbsp;0.42. This suggests the existence of a critical threshold; once the volume of heart tissue receiving 30 Gy exceeds this point, the risk of pericarditis accelerates dramatically. A similar, albeit less pronounced, non-linear trend was observed for heart_v40 (Figure 3D).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;In summary, the SHAP analysis of the full, all-inclusive model consistently points to a clear conclusion: high-dose cardiac exposure parameters (heart_v30 and heart_v40) are the dominant factors in predicting pericarditis risk. This unexpected finding raises a critical question that forms the central paradox of this study: Is the predictive value of the traditionally important heart_v5 parameter genuinely negligible, or is it being \u0026quot;masked\u0026quot; by the overwhelming statistical influence of the collinear, high-dose features? To definitively answer this question and isolate the true, independent contribution of heart_v5, a more focused analytical approach is warranted.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.4 Validation of heart_v5\u0026apos;s Independent Value via a Simplified Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo resolve the paradox of heart_v5\u0026apos;s diminished importance in the full model and to isolate its true, independent predictive value, we implemented a \u0026quot;full vs. simplified model\u0026quot; analytical strategy. A simplified XGBoost model was developed, which deliberately excluded the dominant, highly collinear high-dose cardiac parameters (heart_v30, heart_v40, and heart_mean_dose) but retained all other clinical and dosimetric variables.\u003c/p\u003e\n\u003cp\u003eThe comparison of feature importance between the two models, presented in Figure 4A, provides a striking resolution to the paradox. In the simplified model, the importance of heart_v5 experienced a dramatic resurgence, leaping from a low rank of 15th in the full model to become the\u0026nbsp;3rd\u0026nbsp;most important predictor overall. Its mean absolute SHAP value increased significantly\u0026nbsp;to 0.067, confirming that its predictive signal was indeed being masked by statistical collinearity in the comprehensive model. This result provides powerful, direct evidence for the significant, independent predictive value of heart_v5.\u003c/p\u003e\n\u003cp\u003eWith the confounding influence of the high-dose parameters removed, the simplified model reveals the true nature of the relationship between heart_v5 and pericarditis risk. The SHAP dependence plot in Figure 4B now illustrates a clear and positive dose-response relationship. The fitted curve demonstrates that as the heart_v5 value increases, its contribution to the predicted risk (the SHAP value) also steadily increases,\u0026nbsp;rising from approximately 0.05 for the lowest heart_v5 values to over 0.15 for the highest values. This confirms that, when its effect is properly isolated, a larger heart volume receiving low-dose radiation is unequivocally associated with a higher risk of pericarditis.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;To ensure this isolated effect was clinically meaningful, we validated the heart_v5 SHAP values from the simplified model against the patients\u0026apos; actual clinical outcomes (Figure 4E). A clear separation was observed: nearly all patients belonging to the true high-risk group (pink dots) registered positive SHAP values for heart_v5, while the majority of true low-risk patients (blue dots) had SHAP values near or below zero. This strong concordance demonstrates that the risk contribution assigned by the model based on heart_v5 aligns remarkably well with real-world clinical results. Crucially, this clarification was achieved with only a negligible trade-off in overall predictive accuracy. As shown in Figure 4F, the AUC of the simplified model (0.903) remained high, only slightly reduced from that of the full model (0.918). This confirms the simplified model itself is a robust and reliable tool for this analysis.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;In conclusion, the \u0026quot;full vs. simplified model\u0026quot; strategy successfully resolved the analytical paradox. By isolating heart_v5 from its collinear counterparts, we have \u0026nbsp;rigorously demonstrated its significant and independent predictive value. The clear dose-response relationship and the strong correlation with clinical outcomes firmly establish heart_v5 as a reliable and clinically relevant biomarker for predicting radiation-induced pericarditis, paving the way for a more in-depth exploration of its clinical utility.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.5In-depth Analysis of heart_v5 Impact using SHAP Values\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo elucidate the influence of the heart_v5 parameter on risk prediction, a detailed analysis was conducted using SHapley Additive exPlanations (SHAP) on the simplified model\u0026apos;s predictions for the test set (n=53), with the results presented in Figure 5.\u003c/p\u003e\n\u003cp\u003eA sina plot illustrating the distribution of heart_v5 SHAP values across all test patients showed marked heterogeneity in the feature\u0026apos;s impact on individual predictions (Figure 5A). The SHAP values spanned a wide range, from approximately\u0026nbsp;-0.34 to +0.26. The median SHAP value was positive, as indicated by the embedded boxplot, and a majority of patients exhibited SHAP values greater than zero. This confirmed that, for most patients in the model, a higher heart_v5 contributed to an increased predicted risk. Conversely, a smaller subset of patients displayed negative SHAP values, for whom the feature contributed to a lower predicted risk.\u003c/p\u003e\n\u003cp\u003eThe relationship between the original, non-standardized heart_v5 values (expressed as a percentage) and the predicted probability of high risk was examined (Figure 5B). A logistic regression curve fitted to the data showed a distinct non-linear relationship. Specifically, the predicted probability of a high-risk event remained low (at approximately\u0026nbsp;5%) for heart_v5 values below 25%. A notable inflection point was observed where the risk probability began to increase sharply for heart_v5 values exceeding 35-40%,\u0026nbsp;reaching approximately 50% for a heart_v5 of 60%.\u003c/p\u003e\n\u003cp\u003eSubgroup analyses were conducted to determine if the effect of heart_v5 was modified by prespecified clinical variables.When stratified by lesion position (Figure 5C), the impact of heart_v5 on predicted risk differed between patients with left-sided versus right-sided tumors. At higher standardized heart_v5 values (greater than 1.0), the SHAP values were consistently greater for patients with left-sided lesions compared to those with right-sided lesions, quantifying a stronger positive contribution to risk in the left-sided cohort. The analysis was also stratified by nodal (N) stage (Figure 5D). For patients in the N0-N1 group, the SHAP value demonstrated a relatively steady increase with rising heart_v5 values. In contrast, for patients in the advanced N2-N3 group, the relationship was more volatile, showing greater fluctuation across the range of heart_v5.\u003c/p\u003e\n\u003cp\u003ePotential interaction effects between heart_v5 and two other dosimetric parameters, maximum heart distance (mhd) and ipsilateral lung volume (ipsi_lung_volume), were investigated visually (Figures 5E and 5F). The SHAP dependence plots for heart_v5 were colored by the standardized values of mhd and ipsi_lung_volume, respectively. In both plots, no clear pattern of vertical separation or distinct clustering of points was observed. This suggested that the predictive impact of heart_v5 within the model was largely independent of mhd and ipsi_lung_volume.\u003c/p\u003e\n\u003cp\u003eIn summary, the analysis of the test set (n=53) revealed that heart_v5 was a primary driver of predicted risk, with a heterogeneous impact that was significantly modulated by tumor laterality, showing a more pronounced risk contribution for left-sided tumors at high heart_v5 values.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis retrospective cohort study evaluated whether low-dose cardiac exposure, quantified as heart_V5, independently predicts radiation-induced pericarditis in breast cancer patients receiving radiotherapy. The analysis used extreme gradient boosting (XGBoost) models with SHAP-based interpretability, comparing a full model including all dose\u0026ndash;volume histogram (DVH) metrics and a simplified model excluding high-dose cardiac features. The study cohort comprised 277 patients (training n\u0026thinsp;=\u0026thinsp;224, independent test n\u0026thinsp;=\u0026thinsp;53). In the simplified model, heart_V5 emerged as the third most important predictor (mean absolute SHAP\u0026thinsp;=\u0026thinsp;0.067) with a clear positive, non-linear association with pericarditis risk, achieving an area under the curve (AUC) of 0.903. In contrast, in the full model, high-dose parameters dominated\u0026mdash;heart_V30 (mean |SHAP| \u0026asymp; 0.155) and heart_V40 (\u0026asymp;\u0026thinsp;0.08) were the top predictors, with heart_V5 ranking 15th and contributing negligibly; the full-model AUC was 0.918 (95% CI, 0.848\u0026ndash;0.987). Clinically, the predicted probability of pericarditis remained approximately 5% for heart_V5\u0026thinsp;\u0026lt;\u0026thinsp;25% and increased sharply above ~\u0026thinsp;35\u0026ndash;40%, reaching\u0026thinsp;~\u0026thinsp;50% at heart_V5\u0026thinsp;\u0026asymp;\u0026thinsp;60%. These findings are consistent with prior evidence linking both high-dose and low-dose cardiac irradiation to late cardiovascular events[\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe present findings suggest that heart_V5 may serve as a clinically relevant predictor of radiation-induced pericarditis, particularly in contexts where high-dose cardiac parameters are not available or are excluded from the model. The non-linear dose\u0026ndash;response curve observed\u0026mdash;with a relatively flat risk below ~\u0026thinsp;25% followed by a steep increase beyond ~\u0026thinsp;35\u0026ndash;40%\u0026mdash;aligns with the concept of a subthreshold region followed by a rapid escalation in risk once compensatory mechanisms are overwhelmed [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. Although heart_V5 was overshadowed by high-dose parameters in the full model, its prominence in the simplified model indicates that low-dose exposure retains predictive value, particularly in scenarios such as advanced cardiac-sparing techniques where high-dose metrics may be minimized [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eComparison with prior literature shows both consistency and divergence. QUANTEC recommendations emphasize limiting high-dose volumes, e.g., heart_V30\u0026thinsp;\u0026lt;\u0026thinsp;46% to reduce pericarditis risk, which is consistent with the full-model finding that heart_V30 was the most important predictor. However, recent population-based studies have reported associations between lower-dose exposures and late cardiac events, even when high-dose parameters are controlled [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Our finding that heart_V5 predicted pericarditis risk in the absence of high-dose variables echoes these reports and extends them to an acute toxicity endpoint.\u003c/p\u003e \u003cp\u003eMechanistically, one possible explanation is that low-dose radiation may contribute to widespread microvascular injury and inflammation, triggering pericardial reactions even without focal high-dose injury [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e]. This diffuse effect could explain why heart_V5 retained importance in the simplified model, despite the lower absolute dose. The SHAP analysis further supports this by showing that the model learned a steep risk gradient once heart_V5 exceeded\u0026thinsp;~\u0026thinsp;35\u0026ndash;40%, suggesting a threshold effect beyond which pericardial tolerance is compromised.\u003c/p\u003e \u003cp\u003eThe difference in predictive patterns between the full and simplified models also reflects the statistical interplay among correlated DVH metrics. High-dose parameters such as heart_V30 and heart_V40 are typically correlated with low-dose parameters, but their stronger direct association with injury may suppress the apparent importance of low-dose metrics in multivariate models [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. By removing these high-dose variables, the simplified model allows the independent contribution of heart_V5 to emerge.\u003c/p\u003e \u003cp\u003eFrom a clinical standpoint, the identified heart_V5 inflection point (~\u0026thinsp;35\u0026ndash;40%) is noteworthy. While current guidelines do not provide explicit low-dose constraints for the heart, incorporating such a threshold could be valuable in treatment planning, especially for patients with pre-existing cardiovascular risk factors [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Our results suggest that even when high-dose constraints are met, vigilance for low-dose exposure is warranted.\u003c/p\u003e \u003cp\u003eOur AUC results indicate that both models performed well (AUC\u0026thinsp;=\u0026thinsp;0.903 for simplified, 0.918 for full), suggesting that heart_V5 could be incorporated into predictive models without substantial loss of accuracy if high-dose parameters are unavailable. This is particularly relevant in multi-institutional datasets where DVH data granularity varies [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. It also reinforces the idea that machine learning methods, coupled with interpretable models, can uncover clinically meaningful relationships that may be masked in conventional regression analyses [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThis study establishes heart V5 as an independent and clinically significant predictor of radiation-induced pericarditis in breast cancer patients, particularly when high-dose parameters are unavailable. A potential risk threshold of ~\u0026thinsp;35\u0026ndash;40% for heart V5 could inform future treatment planning guidelines, complementing existing high-dose constraints [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e]. Incorporating low-dose parameters into predictive models enables clinicians to enhance patient-specific risk assessment and optimize cardiac-sparing strategies. However, the study\u0026rsquo;s retrospective, single-institution design may limit generalizability, necessitating prospective multi-center validation to confirm heart V5 thresholds across diverse populations. Additionally, the limited number of pericarditis events may reduce statistical power for detecting subtle interactions, suggesting that future studies with larger event counts could refine threshold estimates.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eThis retrospective cohort study demonstrated that low-dose cardiac exposure (heart_V5) independently predicts radiation-induced pericarditis in breast cancer patients when high-dose parameters are excluded, with a non-linear dose-response relationship showing a sharp risk increase above 35\u0026ndash;40%. In the full model, high-dose parameters (heart_V30 and heart_V40) dominated, but the simplified model\u0026rsquo;s high performance (AUC\u0026thinsp;=\u0026thinsp;0.903) underscores heart_V5\u0026rsquo;s clinical relevance. These findings suggest that incorporating low-dose cardiac exposure metrics, particularly heart_V5, into radiotherapy planning enhances risk stratification and supports personalized cardiac-sparing strategies. By identifying a potential heart_V5 threshold of 35\u0026ndash;40%, this study informs future guidelines for minimizing pericardial morbidity in breast cancer survivors. Machine-learning approaches, coupled with SHAP-based interpretability, offer a robust framework for uncovering clinically meaningful dosimetric predictors.\u003c/p\u003e "},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCredit authorship contribution statement: Yongkai Lu:\u003c/strong\u003e Writing \u0026ndash; original draft, Methodology, Validation, Software, Project administration, Methodology, Data curation, Conceptualization. \u003cstrong\u003eLei Xu:\u003c/strong\u003e Writing \u0026ndash; review \u0026amp; editing, Methodology, Conceptualization, Validation. \u003cstrong\u003eJian Zhang:\u003c/strong\u003e Methodology, Conceptualization, Validation acquisition. \u003cstrong\u003eRongze Ma:\u003c/strong\u003e Visualization, Conceptualization, Validation. \u003cstrong\u003eYao Wang:\u003c/strong\u003e Visualization, Software. \u003cstrong\u003eShanshan Gao:\u003c/strong\u003e Supervision, Project administration, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eThis program was supported by the Natural Science Basic Research Program of Shaanxi (2025JC-YBQN-1093).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Statement:\u003c/strong\u003e This study was approved by the Ethics Committee of the First Affiliated Hospital of Xi\u0026rsquo;an Jiaotong University. The research was conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants prior to their inclusion in the study. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInstitutional Review Board Statement:\u0026nbsp;\u003c/strong\u003eNot applicable.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInformed Consent Statement:\u003c/strong\u003e Not applicable.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement:\u003c/strong\u003e The data supporting the findings of this study are available from the First Affiliated Hospital of Xi\u0026rsquo;an Jiaotong University but are not publicly available due to patient privacy and ethical restrictions. Anonymized data may be made available upon reasonable request to the corresponding author (Yongkai Lu, [email protected]), subject to approval by the institutional Ethics Committee.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of Interest:\u003c/strong\u003e The authors declare no competing financial intere\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eSharma, A., et al., \u003cem\u003ePrognostic factors for local control and survival for inoperable pulmonary colorectal oligometastases treated with stereotactic body radiotherapy.\u003c/em\u003e Radiother Oncol, 2020. \u003cstrong\u003e144\u003c/strong\u003e: p. 23\u0026ndash;29.\u003c/li\u003e\n\u003cli\u003eSung, H., et al., \u003cem\u003eGlobal Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries.\u003c/em\u003e CA Cancer J Clin, 2021. \u003cstrong\u003e71\u003c/strong\u003e(3): p. 209\u0026ndash;249.\u003c/li\u003e\n\u003cli\u003eDarby, S.C., et al., \u003cem\u003eRisk of ischemic heart disease in women after radiotherapy for breast cancer.\u003c/em\u003e New England Journal of Medicine, 2013. \u003cstrong\u003e368\u003c/strong\u003e(11).\u003c/li\u003e\n\u003cli\u003eFranco, V.I., et al., \u003cem\u003eCardiovascular effects in childhood cancer survivors treated with anthracyclines.\u003c/em\u003e Cardiol Res Pract, 2011. \u003cstrong\u003e2011\u003c/strong\u003e: p. 134679.\u003c/li\u003e\n\u003cli\u003ePeeler, C.R., et al., \u003cem\u003eClinical evidence of variable proton biological effectiveness in pediatric patients treated for ependymoma.\u003c/em\u003e Radiother Oncol, 2016. \u003cstrong\u003e121\u003c/strong\u003e(3): p. 395\u0026ndash;401.\u003c/li\u003e\n\u003cli\u003eEbctcg, et al., \u003cem\u003eEffect of radiotherapy after mastectomy and axillary surgery on 10-year recurrence and 20-year breast cancer mortality: meta-analysis of individual patient data for 8135 women in 22 randomised trials.\u003c/em\u003e Lancet, 2014. \u003cstrong\u003e383\u003c/strong\u003e(9935): p. 2127\u0026ndash;35.\u003c/li\u003e\n\u003cli\u003eEarly Breast Cancer Trialists\u0026rsquo; Collaborative, G., \u003cem\u003eEffects of radiotherapy and of differences in the extent of surgery for early breast cancer on local recurrence and 15-year survival: An overview of the randomised trials.\u003c/em\u003e The Lancet, 2005. \u003cstrong\u003e366\u003c/strong\u003e(9503).\u003c/li\u003e\n\u003cli\u003eEarly Breast Cancer Trialists\u0026rsquo; Collaborative, G., \u003cem\u003eFavourable and unfavourable effects on long-term survival of radiotherapy for early breast cancer: An overview of the randomised trials.\u003c/em\u003e The Lancet, 2000. \u003cstrong\u003e355\u003c/strong\u003e(9217).\u003c/li\u003e\n\u003cli\u003eTaylor, C., et al., \u003cem\u003eEstimating the risks of breast cancer radiotherapy: Evidence from modern radiation doses to the lungs and heart and from previous randomized trials.\u003c/em\u003e Journal of Clinical Oncology, 2017. \u003cstrong\u003e35\u003c/strong\u003e(15).\u003c/li\u003e\n\u003cli\u003ePalmer, M., et al., \u003cem\u003eGenome-Based Characterization of Biological Processes That Differentiate Closely Related Bacteria.\u003c/em\u003e Front Microbiol, 2018. \u003cstrong\u003e9\u003c/strong\u003e: p. 113.\u003c/li\u003e\n\u003cli\u003eMikell, J.L., et al., \u003cem\u003eSimilar survival for patients undergoing reduced-intensity total body irradiation (TBI) versus myeloablative TBI as conditioning for allogeneic transplant in acute leukemia.\u003c/em\u003e Int J Radiat Oncol Biol Phys, 2014. \u003cstrong\u003e89\u003c/strong\u003e(2): p. 360\u0026ndash;9.\u003c/li\u003e\n\u003cli\u003eGu, H., et al., \u003cem\u003eGenetic variants in the CNTNAP2 gene are associated with gender differences among dyslexic children in China.\u003c/em\u003e EBioMedicine, 2018. \u003cstrong\u003e34\u003c/strong\u003e: p. 165\u0026ndash;170.\u003c/li\u003e\n\u003cli\u003eDeasy, J.O., et al., \u003cem\u003eImproving normal tissue complication probability models: the need to adopt a \u0026quot;data-pooling\u0026quot; culture.\u003c/em\u003e Int J Radiat Oncol Biol Phys, 2010. \u003cstrong\u003e76\u003c/strong\u003e(3 Suppl): p. S151\u0026ndash;4.\u003c/li\u003e\n\u003cli\u003eBentzen, S.M., et al., \u003cem\u003eBiomarkers and surrogate endpoints for normal-tissue effects of radiation therapy: the importance of dose-volume effects.\u003c/em\u003e Int J Radiat Oncol Biol Phys, 2010. \u003cstrong\u003e76\u003c/strong\u003e(3 Suppl): p. S145\u0026ndash;50.\u003c/li\u003e\n\u003cli\u003eCho, S.S., et al., \u003cem\u003eClazosentan for Aneurysmal Subarachnoid Hemorrhage: An Updated Meta-Analysis with Trial Sequential Analysis.\u003c/em\u003e World Neurosurg, 2019. \u003cstrong\u003e123\u003c/strong\u003e: p. 418\u0026ndash;424 e3.\u003c/li\u003e\n\u003cli\u003eLiauw, S.L., et al., \u003cem\u003eSalvage radiotherapy after postprostatectomy biochemical failure: does pretreatment radioimmunoscintigraphy help select patients with locally confined disease?\u003c/em\u003e Int J Radiat Oncol Biol Phys, 2008. \u003cstrong\u003e71\u003c/strong\u003e(5): p. 1316\u0026ndash;21.\u003c/li\u003e\n\u003cli\u003eWei, X., et al., \u003cem\u003eRisk factors for pericardial effusion in inoperable non-small cell lung cancer patients treated with definitive chemoradiation therapy.\u003c/em\u003e International Journal of Radiation Oncology, Biology, Physics, 2008. \u003cstrong\u003e71\u003c/strong\u003e(3).\u003c/li\u003e\n\u003cli\u003eSiddiqui, F. and B. Movsas, \u003cem\u003eManagement of Radiation Toxicity in Head and Neck Cancers.\u003c/em\u003e Semin Radiat Oncol, 2017. \u003cstrong\u003e27\u003c/strong\u003e(4): p. 340\u0026ndash;349.\u003c/li\u003e\n\u003cli\u003eEich, H.T., et al., \u003cem\u003eIntensified Chemotherapy and Dose-Reduced Involved Field Radiotherapy in Patients with Early Unfavorable Hodgkin Lymphoma: Final Analysis of the German Hodgkin Study Group (GHSG) Randomized HD11 Trial.\u003c/em\u003e International Journal of Radiation Oncology*Biology*Physics, 2010. \u003cstrong\u003e78\u003c/strong\u003e(3).\u003c/li\u003e\n\u003cli\u003eEmami, B., et al., \u003cem\u003eTolerance of normal tissue to therapeutic irradiation.\u003c/em\u003e Int J Radiat Oncol Biol Phys, 1991. \u003cstrong\u003e21\u003c/strong\u003e(1): p. 109\u0026ndash;22.\u003c/li\u003e\n\u003cli\u003eGay, H.A. and A. Niemierko, \u003cem\u003eA free program for calculating EUD-based NTCP and TCP in external beam radiotherapy.\u003c/em\u003e Phys Med, 2007. \u003cstrong\u003e23\u003c/strong\u003e(3-4): p. 115\u0026ndash;25.\u003c/li\u003e\n\u003cli\u003eLim, T.Y., et al., \u003cem\u003eClinically Oriented Contour Evaluation Using Dosimetric Indices Generated From Automated Knowledge-Based Planning.\u003c/em\u003e Int J Radiat Oncol Biol Phys, 2019. \u003cstrong\u003e103\u003c/strong\u003e(5): p. 1251\u0026ndash;1260.\u003c/li\u003e\n\u003cli\u003eXiao, Z., et al., \u003cem\u003eImpact of heart motion on radiation dose in the heart and left ventricular myocardium during breast cancer treatment.\u003c/em\u003e Front Oncol, 2025. \u003cstrong\u003e15\u003c/strong\u003e: p. 1503131.\u003c/li\u003e\n\u003cli\u003eLi, X., et al., \u003cem\u003eRadiation-induced cardiac substructure damage and dose constraints: a review.\u003c/em\u003e Radiat Oncol, 2025. \u003cstrong\u003e20\u003c/strong\u003e(1): p. 94.\u003c/li\u003e\n\u003cli\u003eNuruddeen, M.G., et al., \u003cem\u003eCardiac substructure dose distributions in node-positive and node-negative breast cancer patients undergoing 3D-CRT: comparing the predictive accuracy of mean heart dose and mean left ventricular dose.\u003c/em\u003e Radiat Oncol, 2025. \u003cstrong\u003e20\u003c/strong\u003e(1): p. 65.\u003c/li\u003e\n\u003cli\u003eRicardi, U., V. Chiofalo, and G.C. Iorio, \u003cem\u003eModern Radiotherapy and Cardiac Toxicity in Breast Cancer.\u003c/em\u003e Clinical and Experimental Obstetrics \u0026amp; Gynecology, 2023. \u003cstrong\u003e50\u003c/strong\u003e(7).\u003c/li\u003e\n\u003cli\u003eYang, J., et al., \u003cem\u003eA multicenter retrospective study of early cardiac toxicity in operable breast cancer patients receiving concurrent dual or mono anti-HER2 therapy with postoperative radiation therapy.\u003c/em\u003e Breast, 2025. \u003cstrong\u003e79\u003c/strong\u003e: p. 103879.\u003c/li\u003e\n\u003cli\u003eDiaz-Gavela, A.A., et al., \u003cem\u003eBreast Radiotherapy-Related Cardiotoxicity. When, How, Why. Risk Prevention and Control Strategies.\u003c/em\u003e Cancers (Basel), 2021. \u003cstrong\u003e13\u003c/strong\u003e(7).\u003c/li\u003e\n\u003cli\u003eLiu, D.-W., et al., \u003cem\u003eThe risk of cardiovascular toxicity caused by cancer radiotherapy\u0026mdash;a narrative review.\u003c/em\u003e Therapeutic Radiology and Oncology, 2022. \u003cstrong\u003e6\u003c/strong\u003e: p. 4\u0026ndash;4.\u003c/li\u003e\n\u003cli\u003eMirestean, C.C., R.I. Iancu, and D.P.T. Iancu, \u003cem\u003eHypofractionated Whole-Breast Irradiation Focus on Coronary Arteries and Cardiac Toxicity-A Narrative Review.\u003c/em\u003e Front Oncol, 2022. \u003cstrong\u003e12\u003c/strong\u003e: p. 862819.\u003c/li\u003e\n\u003cli\u003eParlar, S., et al., \u003cem\u003eInvestigation of cardiac and pulmonary doses in patients with left sided breast cancer treated by radiotherapy with deep inspiration breath hold technique.\u003c/em\u003e International Journal of Radiation Research, 2022. \u003cstrong\u003e20\u003c/strong\u003e(2): p. 369\u0026ndash;375.\u003c/li\u003e\n\u003cli\u003eZhou, R., et al., \u003cem\u003eHypofractionated Radiotherapy followed by Hypofractionated Boost with weekly concurrent chemotherapy for Unresectable Stage III Non-Small Cell Lung Cancer: Results of A Prospective Phase II Study (GASTO-1049).\u003c/em\u003e Int J Radiat Oncol Biol Phys, 2023. \u003cstrong\u003e117\u003c/strong\u003e(2): p. 387\u0026ndash;399.\u003c/li\u003e\n\u003cli\u003eQian, D., et al., \u003cem\u003eDefinitive chemoradiotherapy versus neoadjuvant chemoradiotherapy followed by surgery in patients with locally advanced esophageal squamous cell carcinoma who achieved clinical complete response when induction chemoradiation finished: A phase II random.\u003c/em\u003e Radiother Oncol, 2022. \u003cstrong\u003e174\u003c/strong\u003e: p. 1\u0026ndash;7.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Breast cancer, Radiotherapy, Radiation-induced pericarditis, Heart V5, Machine learning","lastPublishedDoi":"10.21203/rs.3.rs-8105555/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8105555/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eIn breast cancer treatment, high-dose cardiac irradiation has traditionally been the focus, but growing evidence suggests that even low-dose radiation can cause long-term cardiac damage, raising concerns about radiation-induced heart disease (RIHD) and pericarditis as significant survivorship issues. Understanding dosimetric predictors like low-dose cardiac exposure (heart_V5) is critical for optimizing treatment safety.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA retrospective cohort study was conducted from January 2022 to June 2023, involving 277 female breast cancer patients (training n\u0026thinsp;=\u0026thinsp;224, test n\u0026thinsp;=\u0026thinsp;53). Radiation-induced pericarditis risk, the primary outcome, was derived from an NTCP model binarized at 3.8487\u0026times;10⁻⁶ to classify the top 25% as high-risk. Key exposure was heart_V5 (median 31.5%, IQR 21.9\u0026ndash;42.2), and analytics included preprocessing, six machine-learning models (XGBoost best-performing) with 5-fold cross-validation, DeLong test for AUC comparison, and SHAP for model interpretation.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eOn the independent test set (n\u0026thinsp;=\u0026thinsp;53), the XGBoost model achieved superior performance, with an area under the receiver operating characteristic curve (AUC) of 0.918 (95% CI 0.848\u0026ndash;0.987), an accuracy of 0.925, a recall of 0.727, and an F1-score of 0.800. Critically, to isolate the impact of low-dose radiation, a simplified XGBoost model excluding high-dose features was developed. This model maintained high predictive power (AUC\u0026thinsp;=\u0026thinsp;0.903), underscoring the significant, independent predictive value of low-dose cardiac exposure (heart_V5). In this simplified model, heart_V5 was a top-three predictor, demonstrating a non-linear dose-response relationship. While high-dose metrics (heart_V30, heart_V40) were the dominant predictors in the full model, the strong performance of the simplified model confirms that heart_V5 is a key factor for pericarditis risk. Furthermore, heart_V5 was significantly correlated with the mean heart dose (Pearson r\u0026thinsp;=\u0026thinsp;0.75, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eLow-dose cardiac exposure (heart_V5) demonstrates independent predictive value for radiation-induced pericarditis when high-dose features are controlled, suggesting its consideration in radiotherapy planning to reduce late pericardial morbidity.\u003c/p\u003e","manuscriptTitle":"Low-Dose Cardiac Exposure (Heart V5) as an Independent Predictor of Radiation-Induced Pericarditis in Breast Cancer: An Interpretable Machine Learning Study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-22 09:21:15","doi":"10.21203/rs.3.rs-8105555/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8f124f30-8f3c-4b7b-9f55-3599e08a0344","owner":[],"postedDate":"December 22nd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-06T13:41:42+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-22 09:21:15","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8105555","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8105555","identity":"rs-8105555","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}