A simple-to-use clinical nomogram to predict radiation esophagitis among esophageal cancer patients receiving radiotherapy: a retrospective cohort study

Abstract

Background:

Moderate-to-severe acute radiation esophagitis (MSARE) is a treatment-limiting toxicity of thoracic radiotherapy (RT) in esophageal cancer (EC). Early identification of high-risk patients is essential to individualize supportive care and preserve treatment continuity.

Purpose:

To develop and internally validate a parsimonious clinical nomogram for predicting MSARE, interpret it with SHapley Additive exPlanations (SHAP), and benchmark it against four machine learning (ML) algorithms.

Methods:

We retrospectively reviewed 151 EC patients who completed thoracic RT between January 2022 and December 2024. Candidate predictors were screened with LASSO regression and entered into a multivariable logistic model, visualized as a nomogram. Discrimination, calibration, clinical utility, and learning behavior were assessed by the area under the receiver operating characteristic curve (AUC), calibration plots, decision curve analysis (DCA), and a learning curve. Logistic regression was benchmarked against XGBoost, LightGBM, AdaBoost, and K-Nearest Neighbors (KNN) under stratified 5-fold cross-validation.

Results:

MSARE occurred in 81 of 151 patients (53.6%). Three independent predictors were retained: upper esophageal tumor (OR 7.32), hypertension (OR 2.35), and serum albumin (OR 0.85; all P < 0.05). The nomogram achieved a 5-fold cross-validated training AUC of 0.832 and validation AUC of 0.816 (95% CI: 0.676–0.949), with good calibration (P = 0.423) and positive net benefit on DCA. SHAP ranked albumin as the most influential predictor. Logistic regression yielded the highest validation AUC; DeLong tests showed no significant differences among models.

Conclusions:

An interpretable three-variable nomogram demonstrates promising internal discrimination for MSARE, yields comparable performance to four ML algorithms, and supports individualized supportive-care planning by clinicians.

Highlights

• A parsimonious nomogram derived from three routinely measured clinical variables — upper esophageal tumor location, history of hypertension and serum albumin — enables rapid bedside estimation of MSARE risk during esophageal cancer radiotherapy, without the need for additional imaging or dosimetric data.

• With a 5-fold cross-validated AUC of 0.816, the model showed no statistically significant AUC difference compared with XGBoost, LightGBM, AdaBoost and KNN (all DeLong P > 0.05) and combines interpretability with discrimination, providing clinicians with an actionable tool for individualized supportive-care intensification.

Introduction

Esophageal cancer (EC) accounts for an estimated 511,000 new cases and 445,000 cancer-related deaths annually, ranking 11th in incidence and 7th in mortality among all malignancies worldwide (1). Multimodality treatment incorporating definitive or perioperative radiotherapy (RT), with or without concurrent chemotherapy, has substantially improved local control and survival outcomes in both resectable and unresectable disease (2). However, owing to the anatomical proximity of the esophagus to the planned target volume, acute radiation esophagitis (ARE) remains one of the most frequent and dose-limiting toxicities of thoracic RT, with reported incidences for grade ≥ 2 disease ranging from approximately 20% to more than 50% across published EC cohorts.

According to the Radiation Therapy Oncology Group (RTOG) and European Organisation for Research and Treatment of Cancer (EORTC) toxicity criteria, RE grade ≥ 2—characterized by the requirement for continuous narcotic analgesia, marked dysphagia or worse symptoms—is defined as moderate-to-severe ARE (MSARE, 3). MSARE has been consistently linked with weight loss, malnutrition, unplanned treatment interruption, hospitalization, and, in severe cases, esophageal perforation or tracheoesophageal fistula (4–6). Treatment interruption in particular, has been associated with deteriorated local control and overall survival in radical chemoradiotherapy regimens, providing a strong rationale for early identification of at-risk patients and timely activation of supportive measures, including nutritional optimization, prophylactic analgesia, and early dietary modification.

Most existing prediction tools for radiation-induced esophagitis have been developed in cohorts of lung cancer or mixed thoracic malignancy and rely on dosimetric or imaging-derived parameters that are not available at the time of initial clinical decision-making (7–9). Comparatively few models have been derived specifically for EC, and fewer have been built solely from variables that can be obtained without additional examinations. In parallel, the recent expansion of machine learning (ML) into clinical prediction has raised the question of whether conventional logistic regression remains adequate. Systematic reviews indicate that, in low-dimensional clinical datasets, ML algorithms rarely outperform well-specified logistic models in terms of discrimination while substantially compromising interpretability (10). The recently published TRIPOD+AI guideline therefore emphasizes head-to-head benchmarking between classical and ML approaches and recommends the use of model-agnostic interpretation tools—such as SHapley Additive exPlanations (SHAP)—when prediction is intended to inform clinical decisions (11, 12).

Against this background, the present study aimed to: develop a parsimonious clinical nomogram for MSARE prediction in EC patients receiving RT based on routinely available variables; assess its discrimination, calibration, clinical utility, and learning behavior through internal validation; quantify the contribution of each predictor using SHAP; and benchmark the logistic model against four commonly used ML classifiers—XGBoost, LightGBM, AdaBoost, and KNN—using stratified 5-fold cross-validation and the DeLong test. Our goal was to deliver a transparent, externally interpretable, and clinically deployable risk-stratification tool capable of informing proactive supportive-care planning during EC radiotherapy.

Materials and methods

Patients selection

Patients with histologically confirmed EC who received definitive RT or concurrent chemoradiotherapy (CRT) at our institution between January 2022 and December 2024 were retrospectively reviewed. Inclusion criteria were: (i) age > 18 years; (ii) pathologically confirmed EC; (iii) completion of thoracic RT; (iv) Eastern Cooperative Oncology Group (ECOG) performance status ≤ 1; (v) expected survival > 3 months; and (vi) complete baseline clinical and follow-up information. Exclusion criteria were: (i) prior thoracic RT; (ii) history of other active malignancy; and (iii) treatment interruption due to non-toxicity reasons. After application of these criteria, 151 patients were retained for analysis (Figure 1).

Figure 1

Sample size and number of candidate predictors

Among the 151 patients included, 81 developed MSARE. Following the conventional rule of approximately 10 outcome events per variable, this sample could reliably accommodate up to 8 candidate predictors in multivariable modeling (13, 14). LASSO regularization was used as the primary screening step, and the final model retained three predictors, yielding an events-per-variable ratio of 27. This is well within the limits recommended for robust logistic prediction model development.

Treatment and follow-up

All patients received intensity-modulated RT (IMRT) or three-dimensional conformal RT (3D-CRT) to a total dose of 40–70 Gy delivered in 20–35 daily fractions, five days per week. Treatment plans were generated from contrast-enhanced CT simulation scans extending from the cricoid cartilage to the costophrenic angle (slice thickness ≤ 5 mm; field extension was adjusted according to the extent of tumor invasion). The gross tumor volume (GTV) encompassed the primary tumor and involved nodal disease, and the elective nodal volumes were delineated according to ESTRO-ACROP principles. The clinical target volume (CTV) was generated by adding a 5-mm isotropic margin to the GTV, and the planning target volume (PTV) was obtained by adding a further 5–8 mm setup margin to the CTV. Dose constraints for organs at risk followed RTOG recommendations. Personalized CRT, delivered either concurrently or sequentially, was administered to 99 patients; the majority received etoposide (100–120 mg/m² on days 1–3) combined with cisplatin (75 or 60 mg/m² on day 1) or carboplatin (AUC-based dosing on day 1), in cycles repeated every 21 days as recommended by current NCCN guidelines.

Radiation esophagitis assessment

Patients were retrospectively analyzed for treatment-related toxicities at weekly intervals during RT and monthly during the first three months after RT completion. RE was diagnosed by the treating physician on the basis of symptoms, physical examination, and, when indicated, endoscopic findings, and was graded according to the RTOG/EORTC criteria. The highest observed grade during or within three months of RT was recorded for each patient. RE of grade ≥ 2 was defined as MSARE and represented the endpoint of the present study.

Nomogram construction and internal validation

The retained predictors were used to construct a nomogram with the R rms package. Each predictor was mapped to a partial-point scale (0–100), and total points were translated into a predicted probability of MSARE. Discrimination was quantified using the AUC. Calibration was assessed visually with calibration plots — including both apparent and bias-corrected (1,000 bootstrap resamples) curves—and statistically with the Hosmer–Lemeshow goodness-of-fit test. Clinical utility was evaluated using decision curve analysis (DCA, 15). To examine whether the available sample size was sufficient for stable model fit, a learning curve was generated by progressively enlarging the training set and recomputing training and validation AUC. To render the logistic model’s predictions transparent at both the cohort and the individual-patient level, SHAP values were computed for each patient and each input variable (12, 17). SHAP assigns each variable an additive contribution to the deviation between an individual’s predicted probability and the cohort mean prediction, with the sign indicating direction (positive SHAP increases predicted risk, negative SHAP decreases it). Global feature importance was summarized as the mean absolute SHAP value across the cohort, and the distribution of feature effects was inspected with a SHAP summary (beeswarm) plot.

Machine learning model benchmarking

To benchmark conventional logistic regression against algorithms capable of capturing non-linear interactions, four supervised ML classifiers were trained on the same three predictors: XGBoost (gradient-boosted decision trees), LightGBM (histogram-based gradient boosting), AdaBoost (adaptive boosting), and K-Nearest Neighbors (KNN, distance-based). Performance was evaluated under stratified 5-fold cross-validation, with AUC, accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1-score, and Cohen κ computed on both the training and validation folds. Pairwise differences in AUC were tested using the DeLong method (16). Clinical utility across models was further compared using DCA on the validation folds.

Statistical analysis

Continuous variables were summarized as mean ± standard deviation (SD) or median with interquartile range (IQR) and were compared between groups using the Student t-test or the Mann–Whitney U test, as appropriate. Categorical variables were summarized as counts (proportions) and compared with the χ² or Fisher exact test. Candidate predictors of MSARE were screened with LASSO logistic regression and five-fold cross-validation; variables retained at λ_min were entered into multivariable logistic regression to compute adjusted odds ratios (OR) and 95% confidence intervals (CI). Two-sided P values < 0.05 were considered statistically significant. Analyses were performed in R 4.1.1 and Python 3.10.

Results

Patient characteristics and feature selection

A total of 151 patients were enrolled, comprising 121 men (80.1%) and 30 women (19.9%), with ages ranging from 18 to 89 years (median 66; IQR 59.0–72.5, Figure 1). MSARE developed in 81 patients (53.6%) and was absent in 70 (46.4%). Baseline demographic, clinical, and laboratory characteristics in both groups are summarized in Table 1. Out of 26 candidate variables, three predictors—upper esophageal tumor, hypertension, and serum albumin (ALB)—were retained on LASSO with non-zero coefficients (Figures 2A, B), and they were entered into the multivariable logistic regression model.

Figure 2

Development of the nomogram

Multivariable logistic regression confirmed that upper esophageal tumor (OR 7.32, 95% CI 3.00–19.90), serum albumin (OR per g/L 0.85, 95% CI 0.79–0.91), and a history of hypertension (OR 2.35, 95% CI 1.02–5.58) were each independently associated with MSARE (all P < 0.05; Table 2). These three predictors were integrated into a clinical nomogram, in which higher total points correspond to higher predicted probabilities of MSARE (Figure 3).

Figure 3

Validation of the nomogram

Under stratified 5-fold cross-validation, the logistic-based nomogram showed strong and stable discrimination (Figures 4a, b), with a mean training AUC of 0.832 (SD 0.026; 95% CI 0.760–0.904) and a mean validation AUC of 0.816 (95% CI 0.676–0.949). The small gap between training and validation AUC (ΔAUC = 0.016) is consistent with minimal overfitting. Calibration was excellent: the bias-corrected calibration curve closely followed the diagonal of ideal calibration (Figure 4c), and the Hosmer–Lemeshow test confirmed good model fit (P = 0.423). At the Youden-derived optimal cut-off of 0.513, the nomogram achieved 81.5% sensitivity, 75.7% specificity, 83.5% overall accuracy, 79.5% PPV, and 77.9% NPV. Decision curve analysis demonstrated positive net clinical benefit relative to the strategies of “treat all” and “treat none” across threshold probabilities of approximately 10% to 95% (Figure 4d). The learning curve, which plotted training- and validation-fold AUC against incrementally enlarged training sizes, plateaued at approximately 60–80 patients with closely matched training and validation performance (Figure 4e), supporting the adequacy of the available sample for a stable three-variable logistic fit.

Figure 4

SHAP-based feature attribution

SHAP analysis was performed on the final logistic model to quantify the global contribution of each predictor. Serum albumin emerged as the single most influential feature (mean |SHAP| = 0.172; relative contribution 51.2%), followed by upper esophageal tumor (0.099; 29.3%) and hypertension (0.066; 19.5%, Figure 4f). The SHAP summary (beeswarm) plot (Figure 4g) confirmed the expected direction of effects: lower albumin values (red) were associated with positive SHAP values and therefore higher predicted MSARE risk, whereas the presence of upper esophageal tumor and hypertension each shifted predictions toward higher risk; the directionality of all three features was concordant with the multivariable OR estimates.

Comparison with machine learning models

On the training folds (Figure 5a), the five candidate models achieved AUCs of 0.876 (XGBoost), 0.868 (KNN), 0.849 (AdaBoost), 0.832 (logistic regression), and 0.640 (LightGBM). On the validation folds, the ranking shifted in favor of logistic regression: it yielded the highest mean validation AUC of 0.816 (95% CI 0.676–0.949), followed by XGBoost (0.804), AdaBoost (0.781), KNN (0.753), and LightGBM (0.594, Figure 5b, Table 3). Logistic regression also showed the smallest gap between training and validation AUC (ΔAUC = 0.016), whereas ensemble and distance-based ML models exhibited larger gaps (ΔAUC 0.07–0.12), indicating greater overfitting at the available sample size. Pairwise DeLong tests revealed no statistically significant differences in AUC between logistic regression and any of the other four models (all P > 0.05; Table 4). Therefore, these results should be interpreted as evidence of comparable internal discrimination. The validation AUC boxplot (Figure 5d) confirmed that logistic regression had the highest median validation AUC. Decision curve analysis on the validation folds (Figure 5c) showed marginally higher net benefit for XGBoost and KNN at threshold probabilities of approximately 30–60%, but logistic regression remained clinically useful across the relevant range.

Figure 5

Discussion

Definitive radiotherapy remains a cornerstone in the management of esophageal cancer, both as a curative-intent modality for unresectable disease and as part of trimodality therapy for locally advanced resectable tumors. Accurate, easily deployable prediction of MSARE is essential. In the present study, we developed a parsimonious nomogram based on three readily available clinical variables, demonstrated its robustness through internal cross-validation and learning-curve analysis, interpreted its mechanism through SHAP, and showed that it performs at least showed no statistically significant AUC difference compared with four commonly used ML algorithms on the same data.

Among the three predictors retained in the final model, upper esophageal tumor location carried the largest adjusted OR (7.32). This is anatomically and biologically plausible. The cervical and upper-thoracic esophagus lies in close proximity to the trachea, thyroid gland, and great vessels, which together reduce the geometric flexibility available for radiation-sparing beam arrangements and frequently result in incidental dose to the highly radiosensitive squamous mucosa of the laryngopharynx (18, 19). The proximal esophagus also drains into an unusually dense cervical lymphatic network; tumors of the upper third metastasize to cervical nodes in up to 60% of cases, mandating elective irradiation of the supraclavicular and lower-cervical nodal regions and consequently an elongated irradiated mucosal length (20). Experimental and clinical data further suggest that the proximal esophagus has a steeper dose–toxicity relationship than the distal segments, with proposed mechanisms including a higher density of sensory afferent fibers, a thinner muscularis mucosa, and a more limited submucosal vascular reserve in this region (21, 22). These mechanistic considerations align with our findings and reinforce the clinical strategy of contralateral esophagus-sparing planning in patients with upper esophageal tumors (18).

Hypertension was associated with a more than two-fold increased risk of MSARE in our cohort (OR 2.35). The biological basis for this association lies in the chronic vascular and inflammatory remodeling that characterizes long-standing hypertension. Endothelial dysfunction, characterized by reduced nitric oxide bioavailability, increased reactive oxygen species generation, and impaired endothelial-dependent vasodilation, predisposes vascular and perivascular tissues to amplified injury when challenged by ionizing radiation (23, 24). Hypertensive small-vessel disease limits the perfusion required to repair radiation-induced mucosal damage (25–27). Activation of the renin–angiotensin–aldosterone system in hypertensive individuals additionally promotes a pro-fibrotic and pro-inflammatory tissue microenvironment through TGF-β and NF-κB signaling, potentiating the inflammatory cascade triggered by radiation (28). In the context of concurrent chemoradiotherapy, these mechanisms may converge with chemotherapy-induced endothelial toxicity to disproportionately raise the risk of MSARE (29). Although our dataset did not permit subgroup analysis by antihypertensive class, experimental data have suggested that inhibitors of the renin–angiotensin axis may attenuate radiation-induced normal-tissue injury—an interaction worth exploring formally in future prospective cohorts.

Serum albumin emerged as both an independent predictor in the multivariable model (OR per g/L 0.85) and, on SHAP analysis, as the single most influential feature, contributing 51.2% of the model’s average attribution. Clinically, serum albumin is now recognized as more than a static index of nutritional reserve: it is a negative acute-phase reactant, downregulated in response to systemic inflammation through cytokine-mediated suppression of hepatic synthesis and capillary leak (30, 31). In radiation oncology, low pretreatment albumin reflects the dual burden of malnutrition—which is present in 60–85% of EC patients (32)—and chronic tumor-driven inflammation, both of which amplify mucosal vulnerability to radiation. Consistent with our findings, a retrospective study in EC patients showed that pretreatment albumin below 35 g/L was associated with a substantially increased risk of grade ≥ 2 RE (33). Mechanistically, albumin itself possesses antioxidant and free-radical-scavenging properties (34); under oxidative stress, it can undergo structural modification to form oxidized albumin, which loses these protective functions and contributes to pro-inflammatory signaling, microthrombosis, endothelial injury, and intensified local inflammation (35–37). Because albumin, unlike tumor location, is a potentially modifiable variable, our findings support the integration of pre-RT nutritional optimization and structured anti-inflammatory measures into the management of high-risk EC patients identified by the nomogram.

In clinical practice, for patients predicted to be at high risk, several interventions can be considered. From the radiotherapy-planning perspective, a high predicted risk may prompt review of plan quality and consideration of esophagus-sparing optimization when target coverage can be maintained. Additionally, practical interventions include pre-RT nutritional assessment and optimization, early dietary and dysphagia counseling, timely analgesic and mucosal-protective therapy, hydration or enteral nutritional support when needed, and closer symptom monitoring during treatment.

SHAP analysis adds an important interpretive layer to the nomogram. While the nomogram provides a population-level scoring system, SHAP values quantify how each variable shifts the predicted probability of an individual patient relative to the cohort average, rendering the model’s reasoning transparent at the level of the individual case. The dominance of serum albumin in SHAP attribution indicates that the strongest day-to-day variation in MSARE risk in our cohort is driven not by the binary anatomical factor of tumor location but by a continuous, modifiable nutritional–inflammatory biomarker. This observation is clinically actionable. The finding that conventional logistic regression showed no statistically significant AUC difference compared with the discrimination of XGBoost, LightGBM, AdaBoost, and KNN in our cohort is consistent with an accumulating body of evidence on the performance of ML in low-dimensional clinical datasets. In a systematic review of 71 studies comparing ML and logistic regression for clinical prediction, Christodoulou and colleagues reported no consistent advantage of ML when the sample size was modest and the number of candidate variables was limited (10). Our analysis supports this view, but it does not establish statistical superiority or formal non-inferiority of logistic regression. Together with its inherent interpretability and the convenience of being translatable into a nomogram, these considerations justified our decision to retain logistic regression as the primary model and to relegate ML results to a benchmarking role.

Several limitations should be acknowledged. First, this is a single-center retrospective analysis, and despite the application of strict inclusion criteria and validated grading scales, residual selection and information biases cannot be excluded. Second, dosimetric parameters and planning-CT-based radiomic features were not included; these variables may further refine prediction and should be explored in future iterations of the model. Third, internal validation by bootstrap and cross-validation, while informative, does not replace external validation in independent cohorts, which we are now actively pursuing through prospective multi-center collaboration.

Conclusions

We developed and internally validated a parsimonious, interpretable nomogram for the prediction of MSARE in patients with esophageal cancer receiving radiotherapy. The model relies on three routinely available clinical variables—upper esophageal tumor location, hypertension, and serum albumin, which demonstrate promising internal discrimination, good calibration, and clear net clinical benefit on decision curve analysis. External validation is required before broad clinical implementation.

References

• 1

  BrayFLaversanneMSungHFerlayJSiegelRLSoerjomataramIet al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. (2024) 74(3):229–63. doi: 10.3322/caac.21834

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  YangHLiuHChenYZhuCFangWYuZet al. Long-term efficacy of neoadjuvant chemoradiotherapy plus surgery for the treatment of locally advanced esophageal squamous cell carcinoma: The NEOCRTEC5010 randomized clinical trial. JAMA Surg. (2021) 156(8):721–9. doi: 10.1001/jamasurg.2021.2373

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  CoxJDStetzJPajakTF. Toxicity criteria of the radiation therapy oncology group (RTOG) and the european organization for research and treatment of cancer (EORTC). Int J Radiat Oncol Biol Phys. (1995) 31:1341–6. doi: 10.1016/0360-3016(95)00060-C

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  MurroDJakateS. Radiation esophagitis. Arch Pathol Lab Med. (2015) 139:827–30. doi: 10.5858/arpa.2014-0111-RS

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  HowellDD. The role of radiation therapy in the palliation of gastrointestinal Malignancies. Gastroenterol Clin North Am. (2006) 35:125–30. doi: 10.1016/j.gtc.2005.12.011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  BradleyJMovsasB. Radiation esophagitis: Predictive factors and preventive strategies. Semin Radiat Oncol. (2004) 14:280–6. doi: 10.1016/j.semradonc.2004.06.003

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  NiederCImingenKSMannsåkerBYobutaRHauklandEet al. Risk factors for esophagitis after hypofractionated palliative (chemo) radiotherapy for non-small cell lung cancer. Radiat Oncol. (2020) 15(1):91. doi: 10.1186/s13014-020-01550-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  HiranoYKanekoHKonishiTItohHMatsudaSKawakuboHet al. Impact of body mass index on major complications, multiple complications, in-hospital mortality, and failure to rescue after esophagectomy for esophageal cancer: A nationwide inpatient database study in Japan. Ann Surg. (2023) 277(4):e785–92. doi: 10.1097/SLA.0000000000005321

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  ZhengXGuoWWangYZhangJZhangYChengCet al. Multi-omics to predict acute radiation esophagitis in patients with lung cancer treated with intensity-modulated radiation therapy. Eur J Med Res. (2023) 28(1):126. doi: 10.1186/s40001-023-01041-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  ChristodoulouEMaJCollinsGSSteyerbergEWVerbakelJYVan CalsterBet al. A systematic review shows no performance benefit of machine learning over logistic regression for clinical prediction models. J Clin Epidemiol. (2019) 110:12–22. doi: 10.1016/j.jclinepi.2019.02.004

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  CollinsGSMoonsKGMDhimanPRileyRDBeamALVan CalsterBet al. TRIPOD+AI statement: Updated guidance for reporting clinical prediction models that use regression or machine learning methods. Bmj. (2024) 385:e078378. doi: 10.1136/bmj-2023-078378

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12

  LundbergSMLeeSI. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. (2017) 30:4765–74.

  • Google Scholar

• 13

  VittinghoffEMcCullochCE. Relaxing the rule of ten events per variable in logistic and Cox regression. Am J Epidemiol. (2007) 165:710–8. doi: 10.1093/aje/kwk052

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 14

  RileyRDEnsorJSnellKIEHarrellFEMartinGPReitsmaJBet al. Calculating the sample size required for developing a clinical prediction model. Bmj. (2020) 368:m441. doi: 10.1136/bmj.m441

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 15

  VickersAJElkinEB. Decision curve analysis: A novel method for evaluating prediction models. Med Decis Making. (2006) 26:565–74. doi: 10.1177/0272989X06295361

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 16

  DeLongERDeLongDMClarke-PearsonDL. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach. Biometrics. (1988) 44:837–46. doi: 10.2307/2531595

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  LundbergSMErionGChenHDeGraveAPrutkinJMNairBet al. From local explanations to global understanding with explainable AI for trees. Nat Mach Intell. (2020) 2(1):56–67. doi: 10.1038/s42256-019-0138-9

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  Al-HalabiHPaetzoldPSharpGCOlsenCWillersHet al. A contralateral esophagus-sparing technique to limit severe esophagitis associated with concurrent high-dose radiation and chemotherapy in patients with thoracic Malignancies. Int J Radiat Oncol Biol Phys. (2015) 92(4):803–10. doi: 10.1016/j.ijrobp.2015.03.018

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19

  ChiricaMKellyMDSiboniSAiolfiARivaCGAstiEet al. Esophageal emergencies: WSES guidelines. World J Emerg Surg. (2019) 14:26. doi: 10.1186/s13017-019-0245-2

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  ChenJCaiWZhengXChenYLinYChenXet al. The pattern of cervical lymph node metastasis in thoracic esophageal squamous cell carcinoma may affect the target decision for definitive radiotherapy. Radiother Oncol. (2017) 123(3):382–6. doi: 10.1016/j.radonc.2017.04.011

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 21

  PanYBrinkCKnapMKhalilAANyhusCHMcCullochTet al. Acute esophagitis for patients with local-regional advanced non small cell lung cancer treated with concurrent chemoradiotherapy. Radiother Oncol. (2016) 118(3):465–70. doi: 10.1016/j.radonc.2016.01.007

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22

  LiJWeiGHHuangHLanYPLiuBLiuHet al. Nerve injury-related autoimmunity activation leads to chronic inflammation and chronic neuropathic pain. Anesthesiology. (2013) 118(2):416–29. doi: 10.1097/ALN.0b013e31827d4b82

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23

  BakerJEMoulderJEHopewellJW. Radiation as a risk factor for cardiovascular disease. Antioxid Redox Signal. (2011) 15:1945–56. doi: 10.1089/ars.2010.3742

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 24

  DuranteAMazzapicchiABaiardo RedaelliM. Systemic and cardiac microvascular dysfunction in hypertension. Int J Mol Sci. (2024) 25:13294. doi: 10.3390/ijms252413294

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  VågesjöEParvKAhlDSeignezCHerrera HidalgoCGiraudAet al. Perivascular macrophages regulate blood flow following tissue damage. Circ Res. (2021) 128(11):1694–707. doi: 10.1161/CIRCRESAHA.120.318380

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  LemariéCATharauxPLLehouxS. Extracellular matrix alterations in hypertensive vascular remodeling. J Mol Cell Cardiol. (2010) 48:433–9. doi: 10.1016/j.yjmcc.2009.09.018

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27

  McMasterWGKiraboAMadhurMSHarrisonDGet al. Inflammation, immunity, and hypertensive end-organ damage. Circ Res. (2015) 116(6):1022–33. doi: 10.1161/CIRCRESAHA.116.303697

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 28

  Muñoz-DurangoNFuentesCACastilloAEGonzález-GómezLMVecchiolaAFardellaCEet al. Role of the renin-angiotensin-aldosterone system beyond blood pressure regulation: Molecular and cellular mechanisms involved in end-organ damage during arterial hypertension. Int J Mol Sci. (2016) 17(7):797. doi: 10.3390/ijms17070797

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 29

  De RuysscherDVan MeerbeeckJVandecasteeleKOberijeCPijlsMDingemansAMet al. Radiation-induced oesophagitis in lung cancer patients. Is susceptibility for neutropenia a risk factor? Strahlenther Onkol. (2012) 188(7):564–7. doi: 10.1007/s00066-012-0098-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  EvansDCCorkinsMRMaloneAMillerSMogensenKMGuenterPet al. The use of visceral proteins as nutrition markers: An ASPEN position paper. Nutr Clin Pract. (2021) 36(1):22–8. doi: 10.1002/ncp.10588

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  DonBRKaysenG. Poor nutritional status and inflammation: Serum albumin: Relationship to inflammation and nutrition. Semin Dial. (2004) 17:432–7. doi: 10.1111/j.0894-0959.2004.17603.x

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 32

  MarietteCDe BottonMLPiessenG. Surgery in esophageal and gastric cancer patients: What is the role for nutrition support in your daily practice? Ann Surg Oncol. (2012) 19:2128–34. doi: 10.1245/s10434-012-2225-6

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  DongJZhangWZhangTChenXZhaoJZengYet al. Baseline nutritional status could be a predictor for radiation esophagitis in esophageal cancer patients undergoing radiotherapy. Ann Transl Med. (2020) 8(18):1148. doi: 10.21037/atm-20-4078

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  LinYLinYYueJZouQ. The neutrophil percentage-to-albumin ratio is associated with all-cause mortality in critically ill patients with acute myocardial infarction. BMC Cardiovasc Disord. (2022) 22(1):115. doi: 10.1186/s12872-022-02559-z

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  ColomboGClericiMGiustariniDRossiRMilzaniADalle-DonneIet al. Redox albuminomics: Oxidized albumin in human diseases. Antioxid Redox Signal. (2012) 17(11):1515–27. doi: 10.1089/ars.2012.4702

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 36

  ArquesS. Human serum albumin in cardiovascular diseases. Eur J Intern Med. (2018) 52:8–12. doi: 10.1016/j.ejim.2018.04.014

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  LunaCAliqueMNavalmoralENociMVBohorquez-MagroLCarracedoJet al. Aging-associated oxidized albumin promotes cellular senescence and endothelial damage. Clin Interv Aging. (2016) 11:225–36. doi: 10.2147/CIA.S91453

  • Pubmed Abstract
  • CrossRef
  • Google Scholar