ESTRO 2024 - Abstract Book

S1231

Clinical - Head & neck

ESTRO 2024

6. Ludwig R, Pouymayou B, Balermpas P, Unkelbach J. A hidden Markov model for lymphatic tumor progression in the head and neck. Sci Rep. 2021;11(1):12261.

900

Digital Poster

Development of Prognostic Models for Small Sample Laryngeal Cancer: A Multi-Omics Integration Study

Sixue Dong, Zian Yao, Xiaoming Sun, Jiazhou Wang, Xiaomin Ou

Fudan University Shanghai Cancer Center, Department of Radiation Oncology, Shanghai, China

Purpose/Objective:

This study aims to present and evaluate two methods for the process of multi-Omics integration and develop prognostic model for laryngeal cancer from a small cohort of patient cases.

Material/Methods:

A training cohort containing 47 patients with laryngeal cancer from Fudan University Shanghai Cancer Center was included in this study, the other set of 30 patients from another cancer center were chosen as the external validation cohort, postoperative radiotherapy and chemotherapy were administered to all the selected patients and the follow-up periods were performed at least two years. We retrospectively collected MR imaging data (T1, T2, and T1+C), clinical data, and the dose distributions. The ROIs including gross tumor volume (GTV), planning target volume (PTV) were contoured by a head and neck radiologist with the title of associate chief physician and 17 years of clinical practice. The workflow was presented in Figure 1, the image features and texture features of dose distribution in PTV were extracted by using the Pyradiomics (Version 3.1.0) library in Python (Version 3.11.4), moreover, dose-volume histogram (DVH) features including V5~V50, Dmax and Dmean in PTV were recorded[1-3]. For the image features from T1, T2, and T1+C sequences, Student’s t -test was firstly employed to select the features exhibiting significant differences by limiting a P-value < 0.05, then the features from T1 and T2 sequences with intra class correlation coefficient ≥ 0.75 were filtered[4]. Dosimetric and clinical features were selected by Student’s t -test with a P-value < 0.05. The training cohort was randomly separated into a training set (70%) and test set (30%) and kept the P-value > 0.05 simultaneously. Monomics models including radiomics model, dosiomics model and clinical model were trained with a Random Forest to predict the 2-year progression-free survival (PFS), and the hyperparameters were tuned for each model. Subsequently, the features of three monomics models were integrated to train the first multi-omics model with a Random Forest (multi- RFfea); Additionally, ‘soft voting’ and ‘stacking’, the machine learning -based integration techniques were independently used to integrate the three monomics models into two separate multi-omics models (multi-RFvote and multi-RFstack)[5-7]. The Harrell Concordance Index (C-index) was analyzed as the performance of each prognostic model, and the area under the receiver operating characteristic (ROC) curve (AUC) was calculated to evaluate the performance during the external validation.