S2435
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Poster Discussion 2051
Multigrade prediction model for xerostomia following radiotherapy in head and neck cancer Emmy Dalqvist 1,2 , Tiziana Rancati 3 , Anna Embring 2,4 , Gabriella Alexandersson von Döbeln 4,5 , Eva Onjukka 1,2 1 Department of Nuclear Medicine and Medical Physics, Karolinska University Hospital, Stockholm, Sweden. 2 Department of Oncology Pathology, Karolinska Institutet, Stockholm, Sweden. 3 Data Science Unit, Fondazione IRCCS Instituto Nazionale dei Tumori, Milan, Italy. 4 Department of Radiotherapy, Karolinska University Hospital, Stockholm, Sweden. 5 Department of Clinical Science, Intervention and Technology, Karolinska Institutet, Stockholm, Sweden Purpose/Objective: To develop a cumulative dose–response model for xerostomia grading using repeated longitudinal assessments from a clinical quality registry. Material/Methods: Over the past decade, a local quality registry has collected clinician-reported xerostomia according to the RTOG/EORTC Radiation Morbidity Scales (grades 0 – 4). A cohort of 392 patients (1615 follow-ups) was analyzed, including head and neck cancer patients with prescribed dose ≥ 50 Gy, no re-irradiation or adaptive replanning, clinically contoured salivary glands, and ≥ 3 follow-up assessments (excluding <5 months to avoid influence from acute effects). A cumulative logit mixed model (ordinal logistic regression with random intercepts) was fitted to account for repeated measurements [1], with patient ID as the random grouping factor. Investigated predictors included mean doses to the parotid glands, submandibular glands, and oral cavity, as well as clinical factors (concomitant and induction chemotherapy, smoking, age, sex and time since radiotherapy). Model fitting was performed in R (clmm, package ordinal). Discriminative performance was evaluated using the Polytomous Discrimination Index (PDI) [2]. Ninety-five percent confidence intervals were derived from 10,000 patient-clustered bootstrap Xerostomia severity was recorded on an ordinal scale (grade 0 – 3) with the following distribution: grade 0, 21%; grade 1, 53%; grade 2, 23%; grade 3, 2%; and grade 4, 0%. Significant fixed effects included time since radiotherapy and mean doses to the parotids, the submandibular glands and the oral cavity. The odds ratios were: ORTime (per year) = 0.84 (0.75–0.91), ORP (per Gy) = 1.04 (1.01–1.07), ORSMG (per Gy)= 1.03 (1.01–1.05) and OROC (per Gy)= 1.04 (1.03–1.06). Figure 1 shows how the probabilities of different samples. Results: grades change with time and dose. Random intercept variance was 2.1 (SD = 1.4), which indicates that 39% of
Among the 229 patients in the validation cohort, Kaplan-Meier curve analysis showed that 116 patients were predicted to be low-risk and 99 were classified as high-risk. The Log-rank test yielded a statistically significant difference between the two groups (p=0.001<0.005). C-index reached 0.68, demonstrating the statistical significance of high/low-risk prediction in the validation cohort. Univariate Cox regression analysis revealed a strong association between the risk stratification and clinical outcomes (HR=2.27, 95%CI:1.36-3.79, p=0.001). Conclusion: This study demonstrates that AI model-extracted pathological features can effectively predict postoperative survival and facilitate precise adjuvant therapy for patients with LARC. References: 1. Chen G, Jin Y, Guan WL, Zhang RX, Xiao WW, Cai PQ, Liu M, Lin JZ, Wang FL, Li C, Quan TT, Xi SY, Zhang HZ, Pan ZZ, Wang F, Xu RH. Neoadjuvant PD-1 blockade with sintilimab in mismatch-repair deficient, locally advanced rectal cancer: an open-label, single-centre phase 2 study. Lancet Gastroenterol Hepatol. 2023;8(5):422-431. https://doi.org/10.1016/S2468- 1253(22)00439-3. Epub 2023 Mar 1. PMID: 36870360.2. Shao, Z., et al. "Transformer-Based Correlated Multiple Instance Learning for Whole Slide Image Classification." arXiv preprint arXiv:2106.00908, 2021.3. Chen, R.J., Ding, T., Lu, M.Y. et al. Towards a general- purpose foundation model for computational pathology. Nat Med 30, 850–862 (2024). https://doi.org/10.1038/s41591-024-02857-3 Keywords: Deep Learning, Whole-slide images
Made with FlippingBook - Share PDF online