S649
Clinical – Head & neck
ESTRO 2026
made of existing machine learning models for studying recurrences. Material/Methods: Total of 119 patients with 105 variables were available for the analysis including dose-volume metrics, confounding factors such as smoking, alcohol intake and tobacco chewing.Clinical characteristics of the patients such as TNM status, co-morbidities, weight loss, adjuvant chemotherapy and adaptive radiotherapy status were also evaluated.Missing values were imputed by average/most frequent methods for numerical and categorical variables respectively.Models studied included Logistic Regression, SVM, KNN, Neural Networks, Naive Bayes, Random Forest, Gradient Boosting, AdaBoost and ensemble of these models was also used. K-fold cross- validation is used to evaluate the models with ROC analysis.36 patients experienced recurrence with a minimum follow up of 5 years. Results: AdaBoost and Gradient Boosting methods had the highest AUC=1, classification accuracy (CA), F1, precision and recall values compared to other models. (The Area Under the Curve (AUC) is a metric used to evaluate the performance of a binary classification model. Larger area under the curve represents a more accurate model).Ensemble model had the same AUC as AdaBoost and Gradient Boosting methods, but the rest of the parameters were reduced due to averaging of the inferior models (CA, F1, precision and recall=0.992).The most poorly performing model was KNN with values for AUC, CA, F1, precision and recall values of 0.761,0.739,0.671,0.770 and 00.739
Digital Poster Highlight 4111
From constraint to comfort: 2 years of H&N precision radiotherapy without a face mask. Marion Essers 1 , Lennart Mesch 2 , Richard Hamelink 2 , Remke Lamers 2 , Vera Elshout 2 , Maaike Beugeling 2 , Eva de Wee 2 , Robert Poorter 2 , Willy de Kruijf 1 1 Medical Physics, Institute Verbeeten, Tilburg, Netherlands. 2 Radiotherapy, Institute Verbeeten, Tilburg, Netherlands Purpose/Objective: The standard method for setup and immobilization during irradiation of head and neck cancer (H&N) patients is using a 5-point thermoplastic face mask. However, face masks can be distressing or even intolerable for patients. In addition, when anatomy changes, replacement of the mask and replanning is required. In previous studies we showed that patient treatment without face mask using Surface Guided Radiotherapy (SGRT) results in improved patient comfort with the same dose delivery precision as using a face mask (1,2). In this study, we will describe our continued results and improvements in maskless H&N treatment. Material/Methods: After the simulation and implementation studies (1,2), between August 2024 and September 2025, exactly 100 curative H&N patients were irradiated, if possible, positioned in the dorsal shell of an individual head base (DSPS-Prominent® posterior shell, Macromedics, Moordrecht, The Netherlands, Figure 1), and irradiated on a TruebeamTM equipped with Identify (Varian,Palo Alto,CA,USA). Patients were setup using SGRT, a CBCT was acquired, on-line setup correction was performed, and if necessary, a second CBCT to check the motion during setup correction was acquired. Before treatment, a new surface was captured. Treatment was interrupted if SGRT intrafraction motion exceeded a threshold th2 of 2mm/2 ⁰ . The number of patients starting with a dorsal shell, with a face mask, the number changing to a face mask, and the number of fractions with motions > th2 during setup and treatment was determined. To optimize patient setup procedures, strategies aimed at reducing the frequency of CBCT verifications were explored, such as individualized adjustments of the region of interest (ROI) and the application of a spreadsheet-based approach for systematic comparison between CBCT and SGRT setup deviations.
respectively. Conclusion:
AdaBoost and Gradient Boosting models had the highest AUC =1 for predicting tumor recurrences amongst the models studied. References: 1. Demsar J, Curk T, Erjavec A, Gorup C, Hocevar T, Milutinovic M, Mozina M, Polajnar M, Toplak M, Staric A, Stajdohar M, Umek L, Zagar L, Zbontar J, Zitnik M, Zupan B (2013) Orange: Data Mining Toolbox in Python, Journal of Machine Learning Research 14(Aug):2349 − 2353.2.Hindocha S, Charlton TG, Linton- Reid K, Hunter B, Chan C, Ahmed M, Robinson EJ, Orton M, Ahmad S, McDonald F, Locke I, Power D, Blackledge M, Lee RW, Aboagye EO. A comparison of machine learning methods for predicting recurrence and death after curative-intent radiotherapy for non- small cell lung cancer: Development and validation of multivariable clinical prediction models. EBioMedicine. 2022 Mar;77:103911. doi: 10.1016/j.ebiom.2022.103911. Keywords: Carcinoma nasopharynx, modelling, recurrence
Made with FlippingBook - Share PDF online