OL Oncology Letters 1792-1074 1792-1082 D.A. Spandidos 10.3892/ol.2026.15909 OL-32-6-15909 Articles An interpretable machine-learning model to predict moderate-to-severe acute radiation esophagitis following regional nodal irradiation for breast cancer HeBin 1 2 * QiZhongchun 2 * ZhangHengdi 3 LiLinling 4 LiHua 2 ZouXinyun 2 TanYu 4 ZhangLing 2 KongFangyuan 2 Department of Oncology and Radiotherapy, Chengdu Qingbaijiang District People's Hospital, Chengdu, Sichuan 610300, P.R. China Department of Oncology, The General Hospital of Western Theater Command of The Chinese People's Liberation Army, Chengdu, Sichuan 610083, P.R. China Department of Ophthalmology, The General Hospital of Western Theater Command of The Chinese People's Liberation Army, Chengdu, Sichuan 610083, P.R. China School of Clinical Medicine, Chengdu Medical College, Chengdu, Sichuan 610500, P.R. China Correspondence to: Professor Ling Zhang or Professor Fangyuan Kong, Department of Oncology, The General Hospital of Western Theater Command of The Chinese People's Liberation Army, 270 Tianhui Road, Rongdu Avenue, Jinniu, Chengdu, Sichuan 610083, P.R. China, E-mail: zhangling71@163.com, E-mail: 13540212456@163.com

Contributed equally

 122026 16102026 32 6 555 27012026 16092026 Copyright: © He et al. 2026 This is an open access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.

Acute radiation esophagitis (ARE) is a clinically relevant toxicity occurring in patients with breast cancer receiving post-operative regional nodal irradiation. The present study aimed to develop and evaluate an interpretable machine-learning (ML) model for grade ≥2 ARE. A total of 324 patients were allocated by outcome-stratified random sampling (7:3) to a training set (n=228) and a held-out test set (n=96). A total of six consensus predictors were retained: Pre-existing pruritus, histological grade, body mass index (BMI), diabetes, tumor laterality and the relative esophageal volume receiving ≥30 Gy (RV_V30). In total, 10 ML algorithms were evaluated. Bootstrap internal validation (1,000 resamples) revealed that logistic regression (LR) had the lowest mean optimism among the three preferred candidates (0.027). Final model selection considered bootstrap stability and corrected performance together with held-out discrimination and calibration, model complexity and clinical interpretability. In the held-out test set, LR achieved an area under the curve value of 0.797 (95% confidence interval, 0.706–0.889), a Brier score of 0.179, a calibration slope of 1.258 and a calibration-in-the-large of −0.056. Decision curve analysis revealed positive net benefit relative to treat-all and treat-none strategies across threshold probabilities of 10–66%. SHapley Additive exPlanations identified histological grade, RV_V30 and BMI as the most influential predictors. The final model was implemented as a nomogram and web-based calculator. Overall, the present study demonstrates that this six-feature clinical-dosimetric model may support individualized risk stratification and radiotherapy planning, but prospective multicenter external validation is required before routine clinical use.

 breast cancer radiation-induced esophagitis regional lymph node radiotherapy machine-learning predictive model Funding: No funding was received. Introduction

Breast cancer continues to pose a notable global health burden, representing one of the most prevalent malignancies affecting women worldwide. According to GLOBOCAN 2024, female breast cancer accounted for an ~2.43 million new cases and 693,660 mortalities worldwide in 2024, and was the most commonly diagnosed cancer and the leading cause of cancer-associated mortality among women globally (1). For patients with lymph node metastasis or other high-risk factors, post-operative radiotherapy is important for reducing local recurrence and breast cancer-specific mortality (2,3). The increasing adoption of regional nodal irradiation (RNI) has further improved oncological outcomes (4); however, it also elevates the risk of developing toxicities to adjacent organs.

The esophagus, due to its anatomical proximity to the supraclavicular and internal mammary nodal regions, often receives incidental radiation during RNI, resulting in acute radiation esophagitis (ARE). It has been reported that symptomatic ARE occurs in 30–50% of patients with breast cancer receiving RNI (5). ARE typically manifests as odynophagia, dysphagia and retrosternal pain, which, in severe cases, may lead to malnutrition, dehydration and treatment interruptions, thereby substantially impairing the quality of life of patients. In the era of precision radiotherapy, the early identification of high-risk patients is key for improving quality of life, maintaining treatment efficacy and reducing treatment-related morbidity (6).

Previous studies have identified numerous factors associated with ARE, including esophageal dose-volume parameters, treatment techniques and patient characteristics. Mean esophageal dose (Dmean), V10, V20, and absolute volumes (AV) of an organ at risk receiving 30 or 35 Gy (AV30 and AV35) have been shown to be associated with the incidence and severity of esophagitis (7–9). Treatment techniques may also influence esophageal exposure during breast RNI (8,10), and tumor laterality has been found to be associated with the risk of developing ARE (11). However, conventional prediction approaches may not adequately capture interactions among clinical and dosimetric variables, and clinically interpretable risk tools remain limited.

Cancer is biologically heterogeneous, with recent studies identifying prognostic molecular markers, including endoplasmic reticulum stress-associated proteins, AIMP2 and transcription factors associated with disease-free survival in triple-negative breast cancer (12–14). Although these molecular studies did not evaluate radiotherapy toxicity, they reinforce the broader rationale for individualized risk assessment. In parallel, artificial intelligence (AI) is increasingly being explored in radiotherapy and immunotherapy, with emphasis on clinically validated and interpretable models (15).

Machine-learning (ML) provides a framework for analyzing complex, high-dimensional data and may capture non-linear associations that are difficult to represent with conventional approaches (16,17). This flexibility has motivated the use of ML for clinical outcome prediction (18,19). Interpretability methods, such as SHapley Additive exPlanations (SHAP) can further decompose model predictions into feature-level contributions and may improve transparency in clinical decision support (20).

Therefore, the present study aimed to develop and evaluate an interpretable ML-based model for predicting grade ≥2 ARE in patients with breast cancer receiving RNI. In the present study, univariable and multivariable analysis, least absolute shrinkage and selection operator (LASSO) regression and the Boruta algorithm for feature selection were used, followed by the comparison of 10 ML algorithms. Bootstrap resampling was used to quantify optimism and model stability. The final model was selected by considering internal validation, held-out performance, calibration, parsimony and clinical interpretability. The present study was reported in accordance with the transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD)+AI statement (21).

 Materials and methods <sec> <title>Study design

The present single-center retrospective observational study included patients with breast cancer who received post-operative RNI at The General Hospital of Western Theater Command of the Chinese People's Liberation Army (Chengdu, China) between January 2016 and June 2025. Patient data were retrieved between July and September 2025, and statistical analyses were conducted between September and December 2025. The study protocol received approval from the Institutional Review Board of The General Hospital of Western Theater Command of the Chinese People's Liberation Army (approval no. 2025EC11-ky019) and was conducted in accordance with the principles of the amended Declaration of Helsinki. Due to the retrospective design, the Institutional Review Board granted a waiver for written informed consent. All patient identifiers were anonymized during the analysis.

 Patient selection

A total of 324 patients with breast cancer who received adjuvant RNI between January 2016 and June 2025 were retrospectively enrolled. The inclusion criteria were as follows: i) Female patients with pathologically confirmed breast cancer; ii) those who underwent mastectomy or breast-conserving surgery; iii) those whose post-operative radiotherapy target volume included the supraclavicular/infraclavicular and/or internal mammary lymph node regions; iv) those whose radiotherapy technique was intensity-modulated radiotherapy (IMRT) alone or IMRT combined with electron beams; and v) those with an availability of complete clinical, laboratory and radiotherapy planning data. The key exclusion criteria included the following: i) The receipt of concurrent chemotherapy; ii) a prior history of thoracic radiotherapy; iii) pre-existing esophageal disease or swallowing dysfunction; and iv) a diagnosis of other active malignancies or notable missing key data (>10% for any variable).

 Treatment and dosimetric parameters

All patients received radiotherapy within 8 weeks post-operatively (if adjuvant chemotherapy was not planned) or upon the completion of adjuvant chemotherapy. Patients were positioned using computed tomography (CT) simulation scans and immobilized with thermoplastic body masks. For patients who underwent mastectomy, a 5-mm-thick silicone sheet was placed on the chest wall surface during radiotherapy. The irradiation fields encompassed the whole breast/chest wall, supraclavicular/infraclavicular fossa and level II axillary lymph nodes.

For the present study, the esophagus was contoured from the inferior border of the cricoid cartilage to the inferior border of the aortic arch (Fig. S1). A representative intensity-modulated radiotherapy plan and dose distribution is presented in Fig. S2. Other organs at risk, including the spinal cord, thyroid gland, contralateral breast, lungs and heart, were also delineated. Patients received IMRT with conventional or hypofractionated regimens. As the total dose varied with fractionation and the use of a tumor-bed boost following breast-conserving surgery, the total dose was analyzed as a continuous variable. Dose-volume histograms were used to extract the Dmean, maximum dose (Dmax), total esophageal volume, relative volumes receiving 10–40 Gy (RV_V10-RV_V40) and AV receiving 10–40 Gy (AV_V10-AV_V40). Relative dose-volume metrics are expressed as the percentage of the contoured esophageal volume receiving at least the specified dose. Concurrent targeted therapy was permitted for patients with HER2-positive disease, whereas concurrent chemotherapy was not permitted. Baseline pruritus was defined as pruritus documented prior to radiotherapy initiation, either at CT simulation or at the pretreatment visit within 7 days and was recorded as present or absent. Pruritus arising during radiotherapy was not included in the prediction model to avoid post-baseline information leakage.

 Missing data

Missing data was assessed for all candidate predictors and was found to account for <5%. Little's test was used to assess whether the missingness pattern was compatible with missing completely at random (MCAR). The null hypothesis was that the data were MCAR, and a two-sided P>0.05 was interpreted as insufficient evidence to reject this hypothesis. Given the low overall missingness and the MCAR result (P>0.05), complete-case analysis was used for model development and evaluation.

 Follow-up and toxicity evaluation

ARE was graded according to the Radiation Therapy Oncology Group acute radiation morbidity scoring criteria (22). Assessments were performed during radiotherapy and at 1 week, 2 weeks and 3 months following the completion of treatment. At least two radiation oncologists reviewed the ARE status. Disagreements were adjudicated by a senior oncologist. The highest recorded ARE grade during the assessment period was used for analysis. The primary endpoint was the occurrence of grade ≥2 ARE at any assessment during this predefined period.

 Feature selection and model construction

Patients were divided into a training set (70%) and a held-out test set (30%) by outcome-stratified random sampling using a fixed seed of 52. Feature selection and model development were performed primarily in the training set. No post-enrollment imputation was required for the complete-case dataset used for model development. For algorithms requiring scaled inputs, centering and scaling parameters were estimated from the training data and then applied unchanged to the held-out data.

Candidate predictors were screened in the training set using three complementary methods: Univariable analysis followed by multivariable logistic regression (LR), LASSO regression and the Boruta algorithm (23–25). For LASSO, a fixed random seed of 43 was used with a binomial model, α=1, a maximum of 100,000 iterations and 10-fold cross-validation based on binomial deviance; λ1 standard error (λ1SE) was selected as the penalty parameter. For Boruta, a fixed random seed of 123 and maxRuns=500 were used, followed by TentativeRoughFix for tentative attributes. The final consensus feature set comprised variables retained by >2 of the three selection approaches: Pre-existing pruritus, histological grade, BMI, diabetes, tumor laterality and esophageal RV_V30.

In total, 10 supervised ML algorithms were evaluated: LR, support vector machine (SVM), gradient boosting machine (GBM), neural network (NN), random forest (RF), extreme gradient boosting (XGBoost), K-Nearest Neighbors (KNN), AdaBoost, light gradient boosting machine (LightGBM) and categorical boosting (CatBoost). For models implemented through the caret framework, a fixed set of outcome-stratified five-fold cross-validation folds (seed=520) was used for tuning and comparison. XGBoost was evaluated using its native R interface with fixed hyperparameters over the same training folds, whereas LightGBM and CatBoost were fitted using their native R interfaces. Model-specific tuning grids, fixed settings, resampling details and available random-seed information are reported in Table SI.

Internal validation was used to assess model stability and identify preferred candidate models. For the three pre-selected models, bootstrap resampling (B=1,000) was performed within the training set to quantify optimism. In each bootstrap sample, the apparent area under the curve (AUC) was compared with performance on the original training cohort, and the mean optimism was subtracted from the apparent AUC of the full training-set model to obtain the optimism-corrected AUC. Final model selection considered bootstrap optimism and corrected performance together with held-out test-set discrimination and calibration, model complexity and clinical interpretability. For threshold-dependent classification metrics, the probability threshold was determined in the training set using the maximum Youden index and then fixed for application to the held-out test set; no threshold optimization was performed in the test set. Held-out discrimination was summarized by AUC with DeLong 95% confidence intervals (CIs). Calibration was assessed using the Brier score, calibration slope and calibration-in-the-large (CITL), and clinical utility was evaluated by decision curve analysis. Pairwise AUC comparisons with LR used DeLong's test.

 Model interpretation was performed using SHAP

Training data were used as the SHAP background distribution, and global importance, beeswarm and dependence plots were generated for the final LR model. Dependence plots were prespecified for RV_V30 and BMI to illustrate the direction of the two continuous predictors with high global importance. Continuous predictor values in the dependence plots are displayed on the standardized model-input scale. For clinical implementation, a static nomogram was constructed from the final LR coefficients using the rms package (https://cran.r-project.org/package=rms), and a web-based dynamic calculator was generated using DynNom (https://cran.r-project.org/package=DynNom) and shiny (https://cran.r-project.org/package=shiny). The nomogram and web calculator were based on the same final LR model.

 Statistical analysis

Continuous variables are presented as the mean ± standard deviation when normally distributed and as median (interquartile range) otherwise; categorical variables are presented as frequencies and percentages. Binary variables were coded according to their prespecified categories, and multilevel categorical variables were represented using indicator variables as required. Between-group comparisons were performed using the independent-samples t-test for normally distributed continuous variables and the Mann-Whitney U test for non-normally distributed continuous variables. χ2 or Fisher's exact tests were used for categorical variables. All tests were two-sided, and P<0.05 was considered to indicate a statistically significant difference.

Statistical analyses were performed using R software (version 4.5.2; Posit Software, PBC). The glmnet and Boruta packages were used for LASSO and Boruta feature selection, respectively (23–25). The caret framework was used for LR, SVM, GBM, NN, RF, KNN and AdaBoost. XGBoost, LightGBM and CatBoost were fitted with their native R interfaces in the final analysis. Model-specific tuning grids, fixed settings, resampling details and available random-seed information are summarized in Table SI.

 Results <sec> <title>Baseline characteristics

A total of 324 patients with breast cancer who received post-operative RNI were included. Grade ≥2 ARE occurred in 35.8% (116/324) of the patients. The cohort was divided by outcome-stratified random sampling into a training set (n=228; 70%) and a held-out test set (n=96; 30%). The study workflow is illustrated in Fig. 1.

In accordance with TRIPOD+AI (21), baseline characteristics are reported for both the training and held-out test sets (Table I). The two subsets were generally comparable for the majority of the variables. However, internal mammary node irradiation and several low-dose esophageal dose-volume histogram (DVH) parameters, including RV_V10, RV_V15, AV_V10 and AV_V15, exhibited significant imbalances. None of these imbalanced variables was retained among the six consensus predictors.

 Feature selection and predictor identification

In total, three feature-selection approaches were applied in the training set: Univariable analysis followed by multivariable logistic regression, LASSO with 10-fold cross-validation (λ1SE=0.069) and Boruta. The corresponding reproducibility settings are described in the aforementioned section in materials and methods.

Multivariable LR was used for association screening in the training set; the resulting coefficients were interpreted as predictive associations rather than causal effects. LASSO retained five variables with non-zero coefficients: Pre-existing pruritus, tumor laterality, histological grade, BMI and diabetes. Boruta retained 10 features: Pre-existing pruritus, tumor laterality, histological grade, BMI, RV_V25, RV_V30, RV_V40, AV_V30, AV_V35 and AV_V40. The results of univariable and multivariable LR analyses for the candidate predictors are presented in Table II.

After integrating the three selection approaches and accounting for collinearity among related DVH and inflammatory variables, six consensus predictors were retained: Pre-existing pruritus, histological grade, BMI, diabetes, tumor laterality and esophageal RV_V30. In the present study, RV_V30 denotes the percentage of the contoured esophageal volume receiving ≥30 Gy; it emerged from the multi-strategy feature-selection process among the candidate DVH metrics. The selection process is illustrated in Fig. 2, and univariable discrimination and correlation analyses are provided in Fig. S3.

 Model development and performance evaluation

Using the six consensus predictors, 10 ML models were developed. The models implemented in caret were evaluated using the same outcome-stratified five-fold cross-validation folds within the training set. Based on internal performance, stability considerations and parsimony, CatBoost, LR and SVM were carried forward for 1,000-resample bootstrap internal validation.

The bootstrap internal-validation results are presented in Table III. The optimism-corrected AUC values were 0.804 for CatBoost, 0.744 for SVM and 0.715 for logistic regression. The corresponding mean optimism values were 0.113, 0.089 and 0.027, respectively. Thus, LR exhibited substantially lower bootstrap optimism than the two more complex pre-selected models, although its optimism-corrected AUC was not the highest.

Threshold-dependent held-out test-set metrics are summarized in Table IV. Using the threshold determined in the training set (0.396), LR achieved an accuracy of 0.729, sensitivity of 0.735, a specificity of 0.726, a precision of 0.595 and a F1 score of 0.658. Performance profiles varied across the other algorithms, and no single model was uniformly superior across all threshold-dependent metrics.

The discrimination and calibration results for all 10 models are reported in Table V. LR achieved the highest held-out AUC value of 0.797 (95% CI, 0.706–0.889), followed by SVM (0.791; 95% CI, 0.699–0.883) and RF (0.786; 95% CI, 0.693–0.878). CatBoost achieved a held-out AUC value of 0.762 (95% CI, 0.659–0.865), as shown in Fig. 3A. In the final 10-model comparison, CatBoost exhibited a training-cohort AUC of 0.923 (95% CI, 0.887–0.959; Fig. S4A). The held-out set was inspected during candidate-model evaluation and served as the early-stopping monitor for LightGBM. Accordingly, these estimates are internal test-set results and should not be interpreted as independent validation performance. For LR, the Brier score was 0.179, the calibration slope was 1.258 and CITL was −0.056. These values indicate generally acceptable calibration with a modest deviation from the ideal slope of 1. Decision curve analysis revealed that LR provided greater net benefit than both treat-all and treat-none strategies across threshold probabilities of 10–66% (Fig. 3B). Considering bootstrap stability, held-out discrimination and calibration, parsimony and clinical interpretability, LR was selected as the final model. The test-set receiver operating characteristic, decision-curve and calibration results are illustrated in Fig. 3, and the corresponding training-cohort plots are provided in Fig. S4.

 Model interpretation

SHAP analysis was used to interpret the final LR model (Fig. 4A-D). Histological grade exhibited the highest global importance, followed by RV_V30 and BMI; pre-existing pruritus, tumor laterality and diabetes also contributed to model predictions (Fig. 4A). The beeswarm plot exhibited positive SHAP contributions for a higher histological grade, higher RV_V30 and a higher BMI (Fig. 4B). Left-sided laterality also tended to contribute toward a higher predicted risk, whereas the encoded higher value for laterality (right) tended to contribute negatively. Dependence plots for RV_V30 and BMI further illustrated the direction and distribution of these continuous feature effects (Fig. 4C and D).

 Visual nomogram and clinical deployment

To support clinical interpretation, the final LR model was presented as a static nomogram (Fig. 5A). The nomogram incorporates the six consensus predictors: Pre-existing pruritus, histological grade, BMI, diabetes, tumor laterality and esophageal RV_V30. Each predictor contributes points according to its regression coefficient, and the summed score is mapped to the predicted probability of grade ≥2 ARE.

An open-access web-based calculator was also developed for point-of-care risk estimation (Fig. 5B). The calculator uses the same final LR model as the static nomogram. A numerical consistency check confirmed concordant predicted probabilities between the nomogram implementation and the final LR model (correlation=1.000; maximum absolute difference, 1.16×10−8). The web application is accessible at: https://hebin7290.shinyapps.io/Radiation-esophagitis/. As an illustrative example, a hypothetical patient with pre-existing pruritus, grade 3 histology, BMI 25 kg/m2, no diabetes, a left-sided tumor and RV_V30 of 20% would have a predicted probability of grade ≥2 ARE of ~0.62.

 Discussion

The present study developed an interpretable clinical-dosimetric model for grade ≥2 ARE following RNI in patients with breast cancer. A total of six predictors captured baseline symptoms, tumor characteristics, metabolic status, anatomy and esophageal dose exposure. Bootstrap internal validation revealed that the more flexible pre-selected models had greater optimism, whereas LR had the lowest mean optimism. Final model selection therefore considered stability together with held-out discrimination and calibration, model complexity and interpretability rather than relying on a single performance metric. The held-out AUC value of LR was 0.797, with a Brier score of 0.179 and modest calibration deviation. These findings support LR as a clinically transparent candidate for further external validation. The selected variables should be interpreted as predictive markers rather than causal determinants of ARE.

Histological grade exhibited the strongest global contribution in the SHAP analysis and was associated with the risk of developing ARE in the predictive model. This association should not be interpreted as evidence that tumor grade directly alters esophageal radiosensitivity. High-grade breast cancers can exhibit hypoxia-related and inflammatory phenotypes (26), and hypoxia and inflammatory signaling can interact through pathways, including hypoxia-inducible factor and NF-κB (27). Systemic inflammatory indices have also been evaluated in radiation-esophagitis prediction in other thoracic malignancies (28–30). These observations provide a possible biological context; however, any association between tumor grade and host normal-tissue susceptibility remains hypothesis-generating and requires dedicated biological and prospective validation.

RV_V30 was retained as the principal dosimetric predictor, supporting the relevance of intermediate-dose esophageal exposure. Prospective evidence in breast RNI has also identified clinically relevant esophageal dose-volume constraints. For example, Wang et al (11) reported RV_V30 and AV_V30 constraints in conventionally fractionated RNI, whereas Ji et al (9) reported associations for AV30 and AV35 in a smaller cohort. In the present study, RV_V30 was selected from a broader candidate set that included Dmean, Dmax and multiple relative and absolute dose-volume metrics. As RV_V30 is modifiable during planning, reducing intermediate-dose esophageal exposure while maintaining target coverage is a plausible optimization strategy. Since no universal RV_V30 cut-off was derived in the present study, this finding should be interpreted as supporting dose minimization rather than a fixed planning constraint.

BMI and diabetes were both retained in the final prediction model. Their contributions may reflect aspects of metabolic status, inflammation, vascular health or treatment geometry rather than a single causal pathway. Obesity is associated with chronic inflammation and oxidative stress (31), and diabetes can impair microvascular function (32). A higher BMI can also affect radiotherapy geometry and planning complexity (33). These mechanisms provide plausible context for the observed predictive associations but were not directly measured in the present study.

Pre-existing pruritus was retained as a predictive marker. By contrast to a formal patient-reported outcome measure, pruritus was retrospectively documented as a binary present/absent variable before radiotherapy. Previous research has linked normal-tissue reactions to inter-individual susceptibility (34), while taxane-related skin toxicity and neuroimmune mechanisms may contribute to persistent pruritus (35–37). A recent breast-radiotherapy study also identified pruritus as a contributor to esophagitis prediction using SHAP (38). Although these observations support further investigation, they do not establish pruritus as a mechanistic marker of radiosensitivity.

A possible explanation is that residual treatment-related symptoms may reflect a host state that also modifies tolerance to subsequent irradiation. Taxane-related cutaneous toxicity and radiation-induced mucosal injury involve inflammatory and epithelial-injury pathways (35,39–41); however, the data of the present study do not establish a shared mechanism or a causal association between pre-existing pruritus and esophageal radiosensitivity. This interpretation is therefore hypothesis-generating and requires prospective validation. The binary ascertainment used in the present retrospective cohort also did not capture pruritus severity, duration or patient-reported burden. Future studies using standardized grading and patient-reported assessment rather than a simple present/absent variable are warranted.

Tumor laterality was retained as a model predictor. In the current coding, right was the higher numeric category and had a negative model coefficient relative to left, consistent with a greater contribution from left-sided laterality in the SHAP analysis. This association may partly reflect treatment geometry, particularly when internal mammary nodal irradiation places beam paths closer to the esophagus. Techniques, such as deep-inspiration breath-hold can modify organ-at-risk geometry and dose exposure (42). Laterality should therefore be regarded as an anatomical prediction marker rather than a causal determinant of ARE.

Published breast cancer RNI studies have largely emphasized dosimetric predictors, with heterogeneous sample sizes, fractionation schedules and validation strategies. Calibration and decision-analytic measures have not been reported consistently, which can make apparent model performance difficult to compare across cohorts.

Yaney et al (8) analyzed 531 conventionally fractionated RNI cases and proposed practical dose cut-points, including mean esophageal dose, V10 and V20. Wang et al (43) prospectively evaluated esophageal constraints in hypofractionated RNI and reported moderate discrimination for parsimonious models. Ji et al (9) examined a smaller hypofractionated cohort and reported high apparent discrimination for AV30 and AV35. These studies illustrate both the clinical relevance of esophageal dosimetry and the sensitivity of reported performance to cohort size and validation design.

The results of the present study add a mixed clinical-dosimetric model that incorporates RV_V30, BMI, diabetes, laterality and a baseline symptom. The comparison of multiple algorithms also illustrates the potential for optimism in flexible learners when the sample size is modest. CatBoost, for example, had a high apparent AUC, but substantially greater bootstrap optimism than LR. A similar observation has been reported in radiotherapy-toxicity modeling outside the breast-RNI setting: In an externally validated head-and-neck dysphagia study, penalized LR using dose-volume metrics performed as well as support vector and RF models (external-validation AUC of 0.82), demonstrating that additional model complexity does not necessarily improve transportable performance (44). This does not imply that ensemble methods are intrinsically inferior; rather, higher-capacity models may be more vulnerable to optimism and calibration instability in limited datasets. Bootstrap methods are useful for quantifying this optimism, although internal validation does not replace external evaluation (45,46). When held-out discrimination is similar, calibration, stability, parsimony and interpretability are therefore key complementary selection criteria.

Internal validation was used to identify preferred candidates and quantify model optimism. Final model selection then considered bootstrap-corrected performance together with held-out test-set discrimination and calibration, model complexity and clinical interpretability. LR was selected as it provided a favorable balance across these criteria rather than because it was uniformly superior on every metric. The held-out test results should therefore be interpreted as internal held-out evaluation, not as external validation.

The model is intended to support pre-radiotherapy risk stratification rather than dictate treatment. Decision curve analysis revealed positive net benefit for LR over treat-all and treat-none strategies across threshold probabilities of 10–66%. In practice, in the event that the predicted risk of a patient exceeds a threshold judged clinically meaningful by the treating team, the plan could be reviewed for opportunities to reduce RV_V30, while maintaining target coverage, and the patient could receive closer symptom monitoring and earlier supportive care. The static nomogram and web-based calculator provide transparent tools for this risk estimate (Fig. 5).

The present study has several limitations which should be mentioned. First, its retrospective design may have introduced selection bias through eligibility criteria, complete-case inclusion and the availability of clinical and dosimetric records. Second, the single-center cohort and modest sample size may limit generalizability. Institutional differences in esophageal contouring, RNI technique, fractionation, treatment-planning systems, accelerators and supportive care may alter predictor distributions and model calibration. Third, calibration estimates from the held-out test set (n=96) remain imprecise. The calibration slope of 1.258 and CITL of −0.056 indicate modest deviation from ideal calibration. Fourth, pre-existing pruritus was recorded only as present or absent and did not capture severity, duration or standardized patient-reported burden. Fifth, the held-out set was inspected during candidate-model evaluation and served as the early-stopping monitor for LightGBM. It therefore provided internal test-set evaluation rather than independent or external validation. Sixth, the present study did not formally compare the final clinical-dosimetric model with a clinical-only baseline model. The incremental value of RV_V30 therefore remains to be quantified. Prospective multicenter studies are thus required to validate the model across different patient populations, RNI techniques, fractionation schedules, planning systems and accelerators. Such studies should also evaluate recalibration and compare clinical-only and clinical-dosimetric models. This comparison should assess discrimination, calibration and decision-curve performance to quantify the incremental value of RV_V30.

In conclusion, an interpretable six-feature clinical-dosimetric model was developed and internally evaluated for grade ≥2 ARE after breast cancer RNI. The final LR model incorporates pre-existing pruritus, histological grade, BMI, diabetes, tumor laterality and esophageal RV_V30. It provided a favorable balance of discrimination, calibration, stability, parsimony and interpretability in the present cohort and was implemented as a nomogram and web-based calculator. Prospective multicenter external validation and, where necessary, recalibration are required before routine clinical implementation.

 Supplementary Material Supporting Data Supporting Data Acknowledgements

The authors wish to thank Dr Yong Diao (Department of Oncology, The General Hospital of Western Theater Command of the Chinese People's Liberation Army, Chengdu, China) for sharing clinical insights regarding patient management and toxicity assessment in radiotherapy. Their expertise contributed to the clinical perspective of the present study.

 Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

 Authors' contributions

BH and ZQ contributed equally to this work. BH and LZ conceived the study and designed the analytical framework. BH drafted the manuscript and coordinated the project. ZQ, LL and HL contributed to patient data collection, clinical variable assessment and model validation. XZ and YT performed the formal data analysis, developed the machine-learning models and software tools, and created the visualizations. HZ contributed to the acquisition and interpretation of clinical and radiotherapy data and critically revised the manuscript. FK contributed to study conception and design, interpretation of the results, critical revision of the manuscript and supervision of the study. LZ supervised the study and provided critical methodological and clinical guidance. BH and ZQ confirm the authenticity of all the raw data. All authors read and approved the final version of the manuscript.

 Ethics approval and consent to participate

The present study was approved by the Ethics Committee of The General Hospital of Western Theater Command (approval no. 2025EC11-ky019) and was performed in accordance with the ethical standards of the Declaration of Helsinki. Due to the retrospective nature of the study, the requirement for written informed consent was waived by the Ethics Committee of The General Hospital of Western Theater Command.

 Patient consent for publication

Not applicable.

 Competing interests

The authors declare that they have no competing interests.

 Use of artificial intelligence tools

During the preparation of this work, artificial intelligence tools were used to improve the readability and language of the manuscript or to generate images, and subsequently, the authors revised and edited the content produced by the artificial intelligence tools as necessary, taking full responsibility for the ultimate content of the present manuscript.

 References SungHFilhoAMLaversanneMFerlayJSiegelRLSoerjomataramIJemalABrayFGlobal cancer statistics 2024: GLOBOCAN estimates of incidence and mortality worldwide for 34 cancers in 186 countriesCA Cancer J Clin76e70090202642417444 EBCTCG (Early Breast Cancer Trialists' Collaborative Group)McGalePTaylorCCorreaCCutterDDuaneFEwertzMGrayRMannuGPetoREffect 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 trialsLancet38321272135201410.1016/S0140-6736(14)60488-824656685 PoortmansPMWeltensCFortpiedCKirkoveCPeignaux-CasasnovasKBudachVvan der LeijFVonkEWeidnerNRiveraSInternal mammary and medial supraclavicular lymph node chain irradiation in stage I–III breast cancer (EORTC 22922/10925): 15-year results of a randomised, phase 3 trialLancet Oncol2116021610202010.1016/S1470-2045(20)30472-133152277 PoortmansPMColletteSKirkoveCVan LimbergenEBudachVStruikmansHColletteLFourquetAMaingonPValliMInternal mammary and medial supraclavicular irradiation in breast cancerN Engl J Med373317327201510.1056/NEJMoa141536926200978 WestKSchneiderMWrightCBeldham-CollinsRCoburnNTiverKGebskiVStuartKERadiation-induced oesophagitis in breast cancer: Factors influencing onset and severity for patients receiving supraclavicular nodal irradiationJ Med Imaging Radiat Oncol64113119202010.1111/1754-9485.1294331486274 BakerSFairchildARadiation-induced esophagitis in lung cancerLung Cancer (Auckl)7119127201628210168 WangDQZhangNDongLHZhongYHWuHFZhongQZJinJHouXRJingHTangYDose-volume predictors for radiation esophagitis in patients with breast cancer undergoing hypofractionated regional nodal radiation therapyInt J Radiat Oncol Biol Phys117186197202310.1016/j.ijrobp.2023.03.06037001764 YaneyAAyanASPanXJhawarSHealyEBeyerSLindseyKKuhnKTedrickKWhiteJRBazanJGDosimetric parameters associated with radiation-induced esophagitis in breast cancer patients undergoing regional nodal irradiationRadiother Oncol155167173202110.1016/j.radonc.2020.10.04233157173 JiMCLiZJLiKWangYXYangBLvLLSuYZhangZWHuoZCQiQDosimetric risk factors for radiation esophagitis in patients with breast cancer following regional nodal radiationWorld J Clin Cases1229953003202410.12998/wjcc.v12.i17.299538898857 BazanJGHealyEBeyerSKuhnKDiCostanzoDSmithTLJhawarSWhiteJRClinical effectiveness of an adaptive treatment planning algorithm for intensity modulated radiation therapy versus 3D conformal radiation therapy for node-positive breast cancer patients undergoing regional nodal irradiation/postmastectomy radiation therapyInt J Radiat Oncol Biol Phys10811591171202010.1016/j.ijrobp.2020.07.02732711036 WangDQShenDXHanNLuYFYangWFTieJHouXRWangXHZhongYHYuXLDosimetric and clinical predictors for radiation esophagitis in patients with breast cancer undergoing conventional fractionated regional nodal irradiation: A prospective studyInt J Cancer15810551066202610.1002/ijc.7016140966125 ShreyaSPandeySGhoshDKPuneet PrasadSBGrossetCFJainBPAnalysis of endoplasmic reticulum stress-associated proteins as prognostic markers in breast cancerCurr Genomics26621632202510.2174/011389202937415825103006484542164555 QiuJZhouTWangDHongWQianDMengXLiuXPan-cancer analysis identifies AIMP2 as a potential biomarker for breast cancerCurr Genomics24307329202310.2174/011389202925594123101414205038235352 WangHHaoRLiuWZhangYMaSLuYHuJQiYIdentification of transcription factors associated with the disease-free survival of triple-negative breast cancer through weighted gene co-expression network analysisCytoJournal2171202410.25259/Cytojournal_127_202439917004 ShiQXuanMLvCZhangZGengXHuangDHuXRecent applications of artificial intelligence in cancer radiotherapy and immunotherapy: Current status and future directionsFront Immunol171776472202610.3389/fimmu.2026.177647242459675 ShamoutFZhuTCliftonDAMachine learning for clinical outcome predictionIEEE Rev Biomed Eng14116126202110.1109/RBME.2020.300781632746368 GreenerJGKandathilSMMoffatLJonesDTA guide to machine learning for biologistsNat Rev Mol Cell Biol234055202210.1038/s41580-021-00407-034518686 KheneZEBigotPDoumercNOuzaidIBoissierRNouhaudFXAlbigesLBernhardJCIngelsABorchielliniDApplication of machine learning models to predict recurrence after surgical resection of nonmetastatic renal cell carcinomaEur Urol Oncol6323330202310.1016/j.euo.2022.07.00735987730 YuYDLeeKSKimJMRyuJHLeeJGLeeKWKimBWKimDSKorean Organ Transplantation Registry Study GroupArtificial intelligence for predicting survival following deceased donor liver transplantation: Retrospective multi-center studyInt J Surg105106838202210.1016/j.ijsu.2022.10683836028137 AmannJVetterDBlombergSNChristensenHCCoffeeMGerkeSGilbertTKHagendorffTHolmSLivneMTo explain or not to explain?-Artificial intelligence explainability in clinical decision support systemsPLoS Digit Health1e0000016202210.1371/journal.pdig.000001636812545 CollinsGSMoonsKGMDhimanPRileyRDBeamALVan CalsterBGhassemiMLiuXReitsmaJBvan SmedenMTRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methodsBMJ385e078378202410.1136/bmj-2023-07837838626948 CoxJDStetzJPajakTFToxicity criteria of the radiation therapy oncology group (RTOG) and the European organization for research and treatment of cancer (EORTC)Int J Radiat Oncol Biol Phys3113411346199510.1016/0360-3016(95)00060-C7713792 KursaMBRudnickiWRFeature selection with the Boruta packageJ Stat Softw36113201010.18637/jss.v036.i11 EngebretsenSBohlinJStatistical predictions with glmnetClin Epigenetics11123201910.1186/s13148-019-0730-131443682 FriedmanJHastieTTibshiraniRRegularization paths for generalized linear models via coordinate descentJ Stat Softw33122201010.18637/jss.v033.i0120808728 BosRZhongHHanrahanCFMommersECSemenzaGLPinedoHMAbeloffMDSimonsJWvan DiestPJvan der WallELevels of hypoxia-inducible factor-1 alpha during breast carcinogenesisJ Natl Cancer Inst93309314200110.1093/jnci/93.4.30911181778 D'IgnazioLBatieMRochaSHypoxia and inflammation in cancer, focus on HIF and NF-κBBiomedicines521201710.3390/biomedicines502002128536364 ZengXYeLLuoMZengDChenYPrognostic value of pretreatment systemic immune-inflammation index in Chinese esophageal squamous cell carcinoma patients receiving radical radiotherapy: A meta-analysisMedicine (Baltimore)102e34117202310.1097/MD.000000000003411737352061 YuYZhengHLiuLLiHZhengQWangZWuYLiJPredicting severe radiation esophagitis in patients with locally advanced esophageal squamous cell carcinoma receiving definitive chemoradiotherapy: Construction and validation of a model based in the clinical and dosimetric parameters as well as inflammatory indexesFront Oncol11687035202110.3389/fonc.2021.68703534249736 QiuJKeDLinHYuYZhengQLiHZhengHLiuLLiJUsing inflammatory indexes and clinical parameters to predict radiation esophagitis in patients with small-cell lung cancer undergoing chemoradiotherapyFront Oncol12898653202210.3389/fonc.2022.89865336483030 IyengarNMGucalpADannenbergAJHudisCAObesity and cancer mechanisms: Tumor microenvironment and inflammationJ Clin Oncol3442704276201610.1200/JCO.2016.67.428327903155 MehnatiPBaradaranBVahidianFNadiriazamSFunctional response difference between diabetic/normal cancerous patients to inflammatory cytokines and oxidative stresses after radiotherapyRep Pract Oncol Radiother25730737202010.1016/j.rpor.2020.06.00832684862 WintersEPooleCChallenges and impact of patient obesity in radiation therapy practiceRadiography (Lond)26e158e163202010.1016/j.radi.2020.01.00532052747 BarnettGCWestCMLDunningAMElliottRMColesCEPharoahPDPBurnetNGNormal tissue reactions to radiotherapy: Towards tailoring treatment dose by genotypeNat Rev Cancer9134142200910.1038/nrc258719148183 SibaudVLebœufNRRocheHBelumVRGladieffLDeslandresMMontastrucMEcheAVigariosEDalencFLacoutureMEDermatological adverse events with taxane chemotherapyEur J Dermatol26427443201610.1684/ejd.2016.283327550571 YosipovitchGBernhardJDClinical practice. Chronic pruritusN Engl J Med36816251634201310.1056/NEJMcp120881423614588 AgelopoulosKPereiraMPWiegmannHStänderSCutaneous neuroimmune crosstalk in pruritusTrends Mol Med28452462202210.1016/j.molmed.2022.03.00535422379 ZhangHWWangYRLiJHuangWXuBPangHWJiangCLPredicting radiation esophagitis in patients undergoing synchronous boost radiotherapy post-breast-conserving surgeryDose Response2315593258251335802202510.1177/1559325825133580240297669 SonisSTThe pathobiology of mucositisNat Rev Cancer4277284200410.1038/nrc131815057287 BentzenSMPreventing or reducing late side effects of radiation therapy: Radiobiology meets molecular pathologyNat Rev Cancer6702713200610.1038/nrc195016929324 KochAReinhardtPElicinOAebersoldDMSchanneDHPredictive biomarkers of radiotherapy-related dermatitis, xerostomia, mucositis and dysphagia in head and neck cancer: A systematic reviewRadiother Oncol203110689202510.1016/j.radonc.2024.11068939706342 PandeliCSmythLMLDavidSSeeAWDose reduction to organs at risk with deep-inspiration breath-hold during right breast radiotherapy: A treatment planning studyRadiat Oncol14223201910.1186/s13014-019-1430-x31822293 WangDQYanLWuHFZhongYHHuangXBJinJZhongQZZhaoLNWangXHJingHProspective cohort study to validate esophageal dose constraints and predictive models for esophagitis in patients with breast cancer undergoing hypofractionated regional nodal radiotherapyRadiother Oncol212111120202510.1016/j.radonc.2025.11112040921329 DeanJAWongKHGayHWelshLCJonesABSchickUOhJHApteANewboldKBhideSIncorporating spatial dose metrics in machine learning-based normal tissue complication probability (NTCP) models of severe acute dysphagia resulting from head and neck radiotherapyClin Transl Radiat Oncol82739201829399642 SteyerbergEWHarrellFEJrPrediction models need appropriate internal, internal-external, and external validationJ Clin Epidemiol69245247201610.1016/j.jclinepi.2015.04.00525981519 EfronBTibshiraniRImprovements on cross-validation: The 632+ bootstrap methodJ Am Stat Assoc92548560199710.2307/2965703

Study flowchart for development and evaluation of the prediction model for grade ≥2 acute radiation esophagitis. Eligible patients were divided by outcome-stratified 7:3 random sampling into a training set (n=228) and a held-out test set (n=96). Feature selection, five-fold CV and bootstrap internal validation were performed in the model-development workflow. Held-out performance was assessed using discrimination, calibration, threshold-dependent classification metrics and DCA. RNI, regional nodal irradiation; CV, cross-validation; DCA; decision curve analysis; RT, radiotherapy; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; AUC, area under the curve; LR, logistic regression; SVM, support vector machine; GBM, gradient boosting machine; NN, neural network; RF, random forest; XGBoost, extreme gradient boosting; KNN, K-Nearest Neighbors; LightGBM, light gradient boosting machine; CatBoost, categorical boosting; SHAP, SHapley Additive exPlanations.

Study flowchart for development and evaluation of the prediction model for grade ≥2 acute radiation esophagitis. Eligible patients were divided by outcome-stratified 7:3...

Feature selection and identification of the six consensus predictors. (A) Variable-importance ranking from the Boruta algorithm. (B) A 10-fold cross-validation curve used to select the LASSO penalty parameter. (C) LASSO coefficient profiles across the penalty path. (D) Intersection of variables identified by multivariable logistic regression, LASSO and Boruta. LASSO, least absolute shrinkage and selection operator.

Feature selection and identification of the six consensus predictors. (A) Variable-importance ranking from the Boruta algorithm. (B) A 10-fold cross-validation curve used to...

Performance of the prediction models in the held-out test set. (A) ROC curves for the 10 machine-learning algorithms; LR achieved an AUC of 0.797 (95% CI, 0.706–0.889). CatBoost achieved an AUC of 0.762 (95% CI, 0.659–0.865). (B) DCA calculated from the established predicted risks. LR showed greater net benefit than treat-all and treat-none strategies across threshold probabilities of 10–66%. (C) Calibration curves for the 10 models. For LR, the Brier score was 0.179, calibration slope was 1.258, and calibration-in-the-large was −0.056. ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; LR, logistic regression; DCA, decision curve analysis; LR, logistic regression; SVM, support vector machine; GBM, gradient boosting machine; XGBoost, extreme gradient boosting; KNN, K-Nearest Neighbors; LightGBM, light gradient boosting machine; CatBoost, categorical boosting; CI, confidence interval.

Performance of the prediction models in the held-out test set. (A) ROC curves for the 10 machine-learning algorithms; LR achieved an AUC of 0.797 (95% CI, 0.706–0.889). CatBoost...

SHAP interpretation of the final Logistic Regression model. (A) Global feature-importance bar plot ranked by mean absolute SHAP value. (B) Beeswarm plot showing the direction and magnitude of feature contributions; color represents the encoded feature value from low to high. (C) SHAP dependence plot for RV_V30. (D) SHAP dependence plot for BMI. For tumor laterality, the higher encoded value corresponds to right and the lower value to left. In panels C and D, RV_V30 and BMI are displayed on the standardized model-input scale. SHAP, SHapley Additive exPlanations; BMI, body mass index; RV_V30, relative esophageal volume receiving ≥30 Gy.

SHAP interpretation of the final Logistic Regression model. (A) Global feature-importance bar plot ranked by mean absolute SHAP value. (B) Beeswarm plot showing the direction and...

Visualization and clinical deployment of the final Logistic Regression model. (A) Static nomogram for predicting grade ≥2 acute radiation esophagitis from pre-existing pruritus, histological grade, BMI, diabetes, tumor laterality and esophageal RV_V30. (B) Web-based calculator using the same final Logistic Regression model for individualized risk estimation. BMI, body mass index; RV_V30, relative esophageal volume receiving ≥30 Gy.

Visualization and clinical deployment of the final Logistic Regression model. (A) Static nomogram for predicting grade ≥2 acute radiation esophagitis from pre-existing...

Baseline characteristics of the training and held-out test sets.

Characteristics Training (n=228) Held-out test set (n=96) P-value
Hepatic dysfunction 0.617
  No 198 (86.84) 86 (89.58)
  Yes 30 (13.16) 10 (10.42)
Pruritus 0.200
  No 154 (67.54) 57 (59.38)
  Yes 74 (32.46) 39 (40.62)
Hypertension >0.999
  No 191 (83.77) 81 (84.38)
  Yes 37 (16.23) 15 (15.62)
Diabetes >0.999
  No 210 (92.11) 89 (92.71)
  Yes 18 (7.89) 7 (7.29)
Family history 0.919
  No 223 (97.81) 93 (96.88)
  Yes 5 (2.19) 3 (3.12)
Tumor laterality 0.504
  Left 132 (57.89) 51 (53.12)
  Right 96 (42.11) 45 (46.88)
T stage 0.232
  T1 49 (21.49) 17 (17.71)
  T2 147 (64.47) 61 (63.54)
  T3 22 (9.65) 16 (16.67)
  T4 10 (4.39) 2 (2.08)
N stage 0.416
  N0 21 (9.21) 5 (5.21)
  N1 97 (42.54) 49 (51.04)
  N2 56 (24.56) 20 (20.83)
  N3 54 (23.68) 22 (22.92)
Lymphatic vessel infiltration 0.241
  No 96 (42.11) 33 (34.38)
  Yes 132 (57.89) 63 (65.62)
Histological grade 0.294
  Grade 1–2 151 (66.23) 70 (72.92)
  Grade 3 77 (33.77) 26 (27.08)
Molecular type 0.502
  Luminal A 26 (11.40) 14 (14.58)
  Luminal B 102 (44.74) 48 (50.00)
  HER-2 positive 55 (24.12) 17 (17.71)
  Triple negative 45 (19.74) 17 (17.71)
Surgical modality 0.776
  Modified radical mastectomy 210 (92.11) 90 (93.75)
  Breast-conserving surgery 18 (7.89) 6 (6.25)
Axillary lymph node dissection 0.827
  No 10 (4.39) 3 (3.12)
  Yes 218 (95.61) 93 (96.88)
Targeted therapy 0.598
  No 153 (67.11) 68 (70.83)
  Yes 75 (32.89) 28 (29.17)
Endocrine therapy 0.689
  No 83 (36.40) 32 (33.33)
  Yes 145 (63.60) 64 (66.67)
Chemotherapy timing 0.430
  Adjuvant chemotherapy 136 (59.65) 52 (54.17)
  Neoadjuvant therapy 92 (40.35) 44 (45.83)
Tumor bed boost 0.881
  No 211 (92.54) 90 (93.75)
  Yes 17 (7.46) 6 (6.25)
Internal mammary lymph nodes irradiation 0.009
  No 223 (97.81) 87 (90.62)
  Yes 5 (2.19) 9 (9.38)
Radiotherapy technique 0.359
  Hybrid IMRT and electron 81 (35.53) 40 (41.67)
  IMRT 147 (64.47) 56 (58.33)
Age, years 54.21 (9.81) 53.34 (9.41) 0.461
Body mass index, kg/m2 24.16 (3.48) 24.50 (3.33) 0.411
Lymph nodes, n removed 13.09 (7.31) 12.89 (6.69) 0.812
Positive lymph nodes, n 4.57 (6.64) 4.39 (4.98) 0.802
Neutrophil, ×109/l 3.20 (1.51) 2.97 (1.42) 0.207
Lymphocyte, ×109/l 1.28 (0.49) 1.27 (0.47) 0.830
Monocyte, ×109/l 0.38 (0.22) 0.36 (0.15) 0.465
Platelet, ×109/l 192.66 186.44 (63.12) 0.412
(62.01) (62.01)
Albumin, g/l 44.19 (3.66) 43.61 (3.43) 0.180
Neutrophil-to-lymphocyte ratio 2.74 (1.42) 2.62 (1.65) 0.516
Platelet-to-lymphocyte ratio 168.50 (78.72) 168.92 (88.75) 0.966
Lymphocyte-to-monocyte ratio 4.31 (6.12) 4.66 (5.16) 0.624
Systemic inflammation response index 1.09 (1.01) 0.99 (0.93) 0.441
Systemic immune-inflammation index 526.13 (302.12) 500.42 (360.49) 0.510
Total dose, Gy 50.34 (3.06) 50.24 (2.61) 0.761
Radiotherapy days, days 24.61 (2.77) 24.61 (2.30) 0.999
Esophageal volume, cm3 12.50 (6.39) 12.43 (6.30) 0.931
Maximum esophageal dose, Gy 49.87 (6.80) 49.98 (6.94) 0.892
Mean esophageal dose, Gy 16.22 (7.40) 17.40 (7.46) 0.190
RV_V10, % 46.79 (21.74) 53.33 (24.52) 0.018
RV_V15, % 39.42 (19.97) 44.76 (22.18) 0.034
RV_V20, % 33.89 (18.77) 37.63 (20.29) 0.111
RV_V25, % 29.03 (17.38) 31.36 (18.89) 0.284
RV_V30, % 24.14 (16.36) 25.82 (17.87) 0.413
RV_V35, % 18.93 (15.17) 19.77 (15.59) 0.651
RV_V40, % 13.95 (13.35) 14.36 (13.33) 0.798
AV_V10, cm3 5.70 (3.94) 6.74 (5.12) 0.049
AV_V15, cm3 4.69 (2.97) 5.51 (3.79) 0.038
AV_V20, cm3 3.96 (2.51) 4.46 (2.76) 0.109
AV_V25, cm3 3.33 (2.19) 3.59 (2.18) 0.335
AV_V30, cm3 2.73 (1.96) 2.88 (1.93) 0.530
AV_V35, cm3 2.10 (1.69) 2.18 (1.69) 0.708
AV_V40, cm3 1.55 (1.46) 1.56 (1.40) 0.948

Data are presented as n (%) for categorical variables and mean (SD) for continuous variables. P-values compare the training and held-out test sets. IMRT, intensity-modulated radiotherapy; RV, relative volume (%); AV, absolute volume (cm3); SD, standard deviation.

Univariable and multivariable logistic regression analyses of candidate predictors in the training set.

Variable No (n=146) Yes (n=82) OR (univariable) OR (multivariable)
Hepatic dysfunction
  No 127 (87.0%) 71 (86.6%)
  Yes 19 (13.0%) 11 (13.4%) 1.04 (0.47–2.30; P=0.931)
Pruritus
  No 107 (73.3%) 47 (57.3%)
  Yes 39 (26.7%) 35 (42.7%) 2.04 (1.15–3.62; P=0.014) 2.02 (1.03–3.98; P=0.042)
Hypertension
  No 123 (84.2%) 68 (82.9%)
  Yes 23 (15.8%) 14 (17.1%) 1.10 (0.53–2.28; P=0.795)
Diabetes
  No 140 (95.9) 70 (85.4%)
  Yes 6 (4.1 %) 12 (14.6%) 4.00 (1.44–11.11; P=0.008) 3.80 (1.14–12.61; P=0.029)
Family history
  No 142 (97.3%) 81 (98.8%)
  Yes 4 (2.7%) 1 (1.2%) 0.44 (0.05–3.99; P=0.464)
Tumor laterality
  Left 73 (50.0%) 59 (72.0%)
  Right 73 (50.0%) 23 (28.0%) 0.39 (0.22–0.70; P=0.002) 0.64 (0.31–1.32; P=0.228)
T stage
  T1 33 (22.6%) 16 (19.5%)
  T2 94 (64.4%) 53 (64.6%) 1.16 (0.59–2.31; P=0.666)
  T3 14 (9.6%) 8 (9.8%) 1.18 (0.41–3.38; P=0.760)
  T4 5 (3.4%) 5 (6.1%) 2.06 (0.52–8.16; P=0.302)
N stage
  N0 15 (10.3%) 6 (7.3%)
  N1 65 (44.5%) 32 (39.0%) 1.23 (0.44–3.47; P=0.695)
  N2 35 (24%) 21 (25.6%) 1.50 (0.50–4.46; P=0.466)
  N3 31 (21.2%) 23 (28.0%) 1.85 (0.62–5.51; P=0.267)
Lymphatic vessel infiltration
  No 66 (45.2%) 30 (36.6%)
  Yes 80 (54.8%) 52 (63.4%) 1.43 (0.82–2.49; P=0.207)
Histological grade
  Grade 1–2 111 (76.0%) 40 (48.8%)
  Grade 3 35 (24.0%) 42 (51.2%) 3.33 (1.87–5.92; P<0.001) 3.49 (1.79–6.83; P<0.001)
Molecular type
  Luminal A 19 (13.0%) 7 (8.5%)
  Luminal B 60 (41.1%) 42 (51.2%) 1.90 (0.73–4.92; P=0.186)
  HER-2 positive 39 (26.7%) 16 (19.5%) 1.11 (0.39–3.16; P=0.840)
  Triple negative 28 (19.2%) 17 (20.7%) 1.65 (0.57–4.74; P=0.354)
Surgical modality
  Modified radical mastectomy 135 (92.5%) 75 (91.5%)
  Breast-conserving surgery 11 (7.5%) 7 (8.5%) 1.15 (0.43–3.08; P=0.788)
Axillary lymph node dissection
  No 6 (4.1 %) 4 (4.9%)
  Yes 140 (95.9%) 78 (95.1%) 0.84 (0.23–3.05; P=0.786)
Targeted therapy
  No 93 (63.7%) 60 (73.2%)
  Yes 53 (36.3%) 22 (26.8%) 0.64 (0.36–1.16; P=0.145)
Endocrine therapy
  No 56 (38.4%) 27 (32.9%)
  Yes 90 (61.6%) 55 (67.1%) 1.27 (0.72–2.24; P=0.414)
Chemotherapy timing
  Adjuvant chemotherapy 83 (56.8%) 53 (64.6%)
  Neoadjuvant therapy 63 (43.2)% 29 (35.4%) 0.72 (0.41–1.26; P=0.251)
Tumor bed boost
  No 136 (93.2%) 75 (91.5%)
  Yes 10 (6.8%) 7 (8.5%) 1.27 (0.46–3.47; P=0.642)
Internal mammary lymph node irradiation
  No 143 (97.9%) 80 (97.6)
  Yes 3 (2.1%) 2 (2.4%) 1.19 (0.20–7.28; P=0.849)
Radiotherapy technique
  Hybrid IMRT and electron 49 (33.6%) 32 (39%)
  IMRT 97 (66.4%) 50 (61%) 0.79 (0.45–1.38; P=0.409)
Age, years 53.6±9.9 55.2±9.6 1.02 (0.99–1.05; P=0.241)
BMI, kg/m2 23.7±3.5 24.9±3.3 1.10 (1.02–1.19; P=0.017) 1.12 (1.02–1.23; P=0.014)
Lymph nodes removed, n 13.2±8.1 12.9±5.7 0.99 (0.96–1.03; P=0.769)
Positive lymph nodes, n 4.4±7.2 4.8±5.5 1.01 (0.97–1.05; P=0.649)
Neutrophil, ×109/l 3.1±1.3 3.3±1.9 1.10 (0.92–1.31; P=0.307)
Lymphocyte, ×109/l 1.3±0.5 1.3±0.5 1.11 (0.64–1.92; P=0.720)
Monocyte, ×109/l 0.4±0.2 0.4±0.2 0.52 (0.13–2.10; P=0.362)
Platelet, ×109/l 195.3±62.4 187.9±61.4 1.00 (0.99–1.00; P=0.387)
Albumin, g/l 44.3±3.9 44.0±3.3 0.97 (0.90–1.05; P=0.506)
Neutrophil-to-lymphocyte ratio 2.7±1.4 2.7±1.4 0.99 (0.82–1.20; P=0.939)
Platelet-to-lymphocyte ratio 173.1±81.5 160.3±73.2 1.00 (0.99–1.00; P=0.242)
Lymphocyte-to-monocyte ratio 3.9±1.9 5.1±9.9 1.05 (0.95–1.17; P=0.352)
Systemic inflammation response index 1.1±0.9 1.1±1.2 1.01 (0.78–1.32; P=0.929)
Systemic immune-inflammation index 532.8±304.6 514.3±299.1 1.00 (1.00–1.00; P=0.658)
Total dose, Gy 50.3±3.0 50.4±3.1 1.02 (0.93–1.11; P=0.712)
Radiotherapy days, days 24.5±2.9 24.9±2.5 1.05 (0.95–1.17; P=0.330)
Esophageal volume, cm3 12.8±6.8 12.0±5.5 0.98 (0.94–1.03; P=0.401)
Maximum esophageal dose, Gy 49.0±7.1 51.4±5.9 1.06 (1.01–1.11; P=0.014) 1.07 (0.99–1.15; P=0.074)
Mean esophageal dose, Gy 15.3±7.2 17.9±7.5 1.05 (1.01–1.09; P=0.013) 0.91 (0.75–1.11; P=0.367)
RV_V10, % 45.4±21.1 49.2±22.7 1.01 (1.00–1.02; P=0.209)
RV_V15, % 37.3±18.9 43.2±21.4 1.01 (1.00–1.03; P=0.034) 1.01 (0.92–1.11; P=0.812)
RV_V20, % 31.5±17.9 38.2±19.6 1.02 (1.00–1.03; P=0.011) 1.04 (0.89–1.22; P=0.584)
RV_V25, % 26.6±16.9 33.4±17.5 1.02 (1.01–1.04; P=0.005) 0.92 (0.75–1.13; P=0.423)
RV_V30, % 21.7±16.1 28.5±16.0 1.03 (1.01–1.04; P=0.003) 1.36 (1.06–1.74; P=0.016)
RV_V35, % 16.8±14.7 22.7±15.4 1.03 (1.01–1.04; P=0.006) 0.73 (0.55–0.96; P=0.024)
RV_V40, % 12.2±12.5 17.1±14.2 1.03 (1.01–1.05; P=0.008) 1.09 (0.91–1.29; P=0.343)
AV_V10, cm3 5.7±4.3 5.6±3.3 0.99 (0.93–1.07; P=0.859)
AV_V15, cm3 4.6±3.1 4.9±2.7 1.04 (0.95–1.13; P=0.444)
AV_V20, cm3 3.8±2.6 4.3±2.3 1.08 (0.97–1.20; P=0.172)
AV_V25, cm3 3.1±2.3 3.7±1.9 1.12 (0.99–1.27; P=0.077)
AV_V30, cm3 2.5±2.1 3.1±1.7 1.17 (1.02–1.35; P=0.028) 0.29 (0.07–1.19; P=0.086)
AV_V35, cm3 1.9±1.7 2.5±1.7 1.23 (1.05–1.44; P=0.013) 3.93 (0.29–52.43; P=0.300)
AV_V40, cm3 1.4±1.4 1.9±1.5 1.27 (1.06–1.53; P=0.012) 0.90 (0.15–5.59; P=0.914)

Outcome: Grade ≥2 acute radiation esophagitis. Data are presented as n (%) for categorical variables and mean ± SD for continuous variables. Blank cells indicate variables that were not retained in the final multivariable model. OR, odds ratio; CI, confidence interval; IMRT, intensity-modulated radiotherapy; RV, relative volume; AV, absolute volume; SD, standard deviation.

Bootstrap internal validation (B=1,000) for the three pre-selected models in the training cohort.

Model Apparent AUC Mean optimism (SD) Optimism-corrected AUC (95% CI) 95% percentile interval for bootstrap-refit AUC on original training set
Categorical boosting 0.917 0.113 (0.018) 0.804 (0.764–0.838) 0.822–0.886
Support vector machine 0.833 0.089 (0.025) 0.744 (0.698–0.795) 0.745–0.815
Logistic regression 0.741 0.027 (0.034) 0.715 (0.650–0.778) 0.708–0.741

Optimism-corrected AUC was computed as the apparent AUC minus the mean optimism across bootstrap samples. The 95% CI was derived from the percentile distribution of corrected AUC values across bootstrap resamples (B=1,000). Bootstrap refit AUC was obtained by fitting the model in each bootstrap sample and evaluating it on the original training data (‘test-on-original’); the 95% CI was computed using percentile intervals. AUC, area under the receiver operating characteristic curve; CI, confidence interval; SD, standard deviation.

Threshold-dependent classification metrics for the 10 machine-learning models in the held-out test set.

Model Threshold Accuracy Sensitivity Specificity Precision F1 score
Logistic regression 0.396 0.729 0.735 0.726 0.595 0.658
Support vector machine 0.331 0.729 0.794 0.694 0.587 0.675
Gradient boosting machine 0.346 0.688 0.676 0.694 0.548 0.605
Neural network 0.386 0.677 0.735 0.645 0.532 0.617
Random forest 0.301 0.708 0.794 0.661 0.562 0.659
Extreme gradient boosting 0.370 0.719 0.794 0.677 0.574 0.667
K-nearest neighbors 0.441 0.646 0.559 0.694 0.500 0.528
AdaBoost 0.437 0.719 0.765 0.694 0.578 0.658
Light gradient boosting machine 0.366 0.719 0.824 0.661 0.571 0.675
Categorical boosting 0.361 0.719 0.853 0.645 0.569 0.682

Thresholds were determined in the training set using the maximum Youden index and were then fixed for application to the held-out test set; no threshold optimization was performed in the held-out test set.

Discrimination and calibration performance of the 10 machine-learning models in the held-out test set.

Model AUC (95% CI, DeLong) Brier score Calibration slope CITL
Logistic regression 0.797 (0.706–0.889) 0.179 1.258 −0.056
Support vector machine 0.791 (0.699–0.883) 0.188 1.530 −0.114
Random forest 0.786 (0.693–0.878) 0.181 0.936 0.153
Extreme gradient boosting 0.771 (0.668–0.875) 0.184 1.044 −0.090
Light gradient boosting machine 0.764 (0.660–0.869) 0.199 2.773 −0.068
Categorical boosting 0.762 (0.659–0.865) 0.189 0.894 −0.184
Gradient boosting machine 0.755 (0.656–0.854) 0.189 0.898 −0.021
AdaBoost 0.744 (0.636–0.853) 0.210 1.676 −0.285
K-nearest neighbors 0.711 (0.601–0.822) 0.211 0.165 0.086
Neural network 0.679 (0.563–0.794) 0.232 0.216 −0.313

AUC 95% CIs were calculated using DeLong's method. Lower Brier scores indicate better overall prediction accuracy. The ideal calibration slope is 1 and the ideal CITL is 0; a slope >1 indicates under-dispersed predicted risks, whereas a slope <1 indicates predictions that are too extreme. AUC, area under the receiver operating characteristic curve; CI, confidence interval; CITL, calibration-in-the-large.