Three clinical variables predicted moderate-to-severe esophagitis during esophageal cancer radiotherapy

KEY POINTS

• This single-centre retrospective cohort included 151 patients with esophageal cancer who completed thoracic radiotherapy between 2022 and 2024. 81/151 patients (53.6%) developed grade ≥2 moderate-to-severe acute radiation esophagitis during treatment or within three months afterward.
• Patients received 40–70 Gy in 20–35 fractions using IMRT or 3D conformal radiotherapy. 99 patients (65.6%) also received chemotherapy, most commonly etoposide with cisplatin or carboplatin.
• From 26 candidate variables, LASSO retained only three independent predictors: upper esophageal tumor location (OR 7.32, 95% CI 3.00–19.90; p<0.001), hypertension (OR 2.35, 95% CI 1.02–5.58; p=0.047) and baseline serum albumin (OR 0.85 per 1 g/L increase, 95% CI 0.79–0.91; p<0.001).
• The internally validated logistic nomogram achieved a mean validation AUC of 0.816 (95% CI 0.676–0.949) versus 0.832 in training folds. Calibration was acceptable, with a Hosmer-Lemeshow p=0.423.
• At the Youden-derived threshold of 0.513, the model achieved 81.5% sensitivity, 75.7% specificity, 83.5% reported overall accuracy, 79.5% positive predictive value and 77.9% negative predictive value. Decision-curve analysis suggested net benefit across a broad range of risk thresholds.
• SHAP analysis identified albumin as the dominant contributor, accounting for 51.2% of mean feature attribution, followed by upper tumor location at 29.3% and hypertension at 19.5%. Patients who developed esophagitis had median albumin 31.2 g/L versus 39.0 g/L among those who did not.
• More complex machine-learning methods did not improve discrimination. Validation AUCs were 0.804 for XGBoost, 0.781 for AdaBoost, 0.753 for K-nearest neighbors and 0.594 for LightGBM, versus 0.816 for logistic regression; none differed significantly from logistic regression by DeLong testing.

CLINICAL TAKEAWAY

SOURCE

Frontiers in Oncology