S553
Clinical – Head & neck
ESTRO 2026
explainable machine learning model for predicting muscle loss after radiotherapy and identifying its contributors. Material/Methods: This study included 1024 patients with oral cavity cancer (derivation cohort, 636 patients; external validation cohort, 388 patients) who underwent surgery and adjuvant radiotherapy between 2010 and 2021. Data from the derivation cohort were divided into 70% training dataset and 30% internal validation dataset. Skeletal muscle mass was measured using pre- and post-radiotherapy computed tomography scans at the C3 vertebral level. A decrease in muscle mass ≥ 4.2% after radiotherapy was identified as the threshold for increased all-cause mortality in the previous study; therefore, this threshold was applied to define “muscle loss.” Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost) models were trained to predict muscle loss. Model performance was evaluated using the area under the curve (AUC). The SHapley Additive exPlanations (SHAP) method was used to interpret the best performing model. Results: Of the 636 and 388 patients in the derivation and external validation cohorts, 166 (26.1 %) and 98 (25.3 %) experienced muscle loss, respectively. The RF model achieved the highest AUC compared to the XGBoost and CatBoost models in the validation cohorts (internal validation: AUC: 0.960, 0.948, and 0.954; external validation: 0.913, 0.892, and 0.904, respectively). Mini Nutritional Assessment score, chemotherapy, and mean dose to the superior pharyngeal constrictor muscle (PCM), supraglottic larynx, and middle PCM were the five most important features for predicting muscle loss. The SHAP dependence plots showed a positive nonlinear relationship between mean dose to the swallowing structures and the risk of muscle loss (Figure 1). The SHAP force plot provided a personalized model prediction interpretation for each patient (Figure 2).
total dose of 70 Gy in 35 fractions. Skin was assessed before and at 3, 6, 12, and 18 months after radiotherapy. All 18 patients were followed up to 12 months, and 14 up to 18 months. Cervical skin elasticity was measured at six points (three on each side of the neck) using a Cutometer® DUAL MPA580 (Courage + Khazaka, Germany). R2 (gross elasticity) was the main index. Skin induration and lymphedema were graded by CTCAE v5.0, and their relationships with R2 were analyzed. Results: The median follow-up was 18.7 months (range, 12– 24.3). Within 12 months, Grade 1 skin induration appeared in 4 patients (22%). Lymphedema occurred in 14 (78%), mostly Grade 1, appearing in 12 of 18 cases (67%) at 3 months. The median R2 decreased from 63.4% at baseline to 57.3% at 12 months. In the 12-month cohort, the Friedman test showed significant differences among time points (p=0.002), and Wilcoxon analysis confirmed a significant reduction between baseline and 12 months (p=0.002). In the 18-month subgroup, no significant difference was observed between 12 and 18 months (p>0.05). Within 12 months, no correlation was found between induration grade and R2, but patients with lymphedema tended to show lower R2 values. Conclusion: Cutometer-based measurement provides a useful, non-invasive tool for assessing late radiation-induced skin changes in head and neck cancer. R2 reflected treatment-related alterations not directly correlated with induration but associated with lymphedema tendency. This quantitative approach may facilitate early recognition and monitoring of radiation-related skin and soft-tissue toxicity. Further studies are warranted to validate its long-term clinical relevance. Keywords: Head and neck, Skin toxicity, Elasticity Digital Poster Highlight 923 Explainable machine learning model for predicting muscle loss after radiotherapy in oral cavity cancer Jie Lee 1,2 , Jhen-Bin Lin 3 , Ning-Hsiang Weng 4 , Kun-Pin Wu 4 1 Radiation Oncology, MacKay Memorial Hospital, Taipei, Taiwan. 2 Medicine, MacKay Medical University, New Taipei City, Taiwan. 3 Radiation Oncology, Changhua Christian Hospital, Changhua, Taiwan. 4 Biomedical Informatics, National Yang Ming Chiao Tung University, Taipei, Taiwan Purpose/Objective: Skeletal muscle loss after radiotherapy is associated with poor survival outcomes in patients with oral cavity cancer. This study aimed to develop an
Made with FlippingBook - Share PDF online