ESTRO 2024 - Abstract Book

S1361

Clinical - Head & neck

ESTRO 2024

By leveraging database analytics, a seamless integration of patients' cancer registration files with their outpatient medical records was achieved. This comprehensive data collection encompassed a wide range of factors. Host factors such as age, sex, body height, body weight, and lifestyle habits like alcohol consumption, betel nut chewing, and smoking were included. Tumoral factors covered aspects like clinical T, clinical N, clinical cStage, and histology type. Treatment-related factors spanned the radiotherapy dose, the gap between simulation and treatment, and the duration of radiotherapy treatment. Additionally, a thorough review of outpatient clinical records provided insights into patients' previous medical histories, highlighting conditions such as diabetes mellitus, hyperlipidemia, hypertension, and ischemic heart disease. Outcome data, crucial for survival and event analysis, detailed recurrence patterns, dates, survival status, and incidents of cerebrovascular accident (CVA). A comprehensive evaluation was conducted by incorporating all these machine learning models together. The aim was to compare their predictive efficiency, with models such as Random Forest, Naïve Bayes, KNeighbor, SVM, Decision Tree, Logistic Regression, XGBoost, and LightGBM In the analyzed cohort, 504 individuals were newly diagnosed with NPC. Males made up 67.26% of this group. A significant portion of the cohort, 51.2%, fell within the age range of 40-60 years. When looking at the stages of diagnosis, stage IV was the most prevalent at 40.7%, closely followed by stage III, which accounted for 25.8% of the patients. Upon applying various machine learning models to this data, different predictive accuracies emerged. The Random Forest, KNeighbor, and SVM models all achieved an accuracy of 0.9406. Naïve Bayes had an accuracy of 0.7936, while the Decision Tree model registered at 0.8020. Logistic Regression matched the accuracy of Random Forest with 0.9406, and XGBoost was slightly behind at 0.9307. Notably, the LightGBM model showcased the highest accuracy, achieving an impressive 0.9504, marking it as the best-performing model in this evaluation. Results:

Conclusion:

This multicenter retrospective data-driven deep learning prediction model informed us that LightGBM yielded the best predictive results. In the future, prospective clinical trials and external validation of this prediction model are needed to further reinforce our findings.

Keywords: CVA, NPC, prediction model

References:

1. Lee CC, Su YC, Ho HC, Hung SK, Lee MS, Chiou WY, Chou P, Huang YS. Increased risk of ischemic stroke in young nasopharyngeal carcinoma patients. Int J Radiat Oncol Biol Phys. 2011 Dec 1;81(5):e833-8. doi: 10.1016/j.ijrobp.2010.11.036. Epub 2011 May 11. PMID: 21570205. 2. Chen MC, Kuan FC, Huang SF, Lu CH, Chen PT, Huang CE, Wang TY, Chen CC, Lee KD. Accelerated Risk of Premature Ischemic Stroke in 5-Year Survivors of Nasopharyngeal Carcinoma. Oncologist. 2019 Sep;24(9):e891 e897. doi: 10.1634/theoncologist.2018-0747. Epub 2019 Mar 21. PMID: 30898891; PMCID: PMC6738310.