Prognosis prediction for radiotherapy patients with nasopharyngeal carcinoma using multi-omics fusion

Abstract

Nasopharyngeal carcinoma (NPC) poses considerable difficulties for prognosis prediction due to its complex tumor morphology, dynamic changes during radiotherapy, and variability in patient responses. Reliable prognosis prediction is essential for guiding personalized treatment strategies, particularly in evaluating adaptive radiotherapy (ART) eligibility and assessing distant metastasis risk. This dissertation develops advanced machine learning frameworks that integrate clinical, radiological, and multi-omics data to address these key objectives in NPC study. Despite progress in research on ART eligibility and distant metastasis prediction, several unresolved challenges remain. Single-omic models, such as those based on computed tomography (CT) or magnetic resonance imaging (MRI) alone from radiomics, are limited in their ability to capture the full scope of tumor heterogeneity and anatomical changes during treatment. While multi-omics information has demonstrated potential, current models often struggle to effectively integrate complementary information from different omics data, leading to suboptimal predictive accuracy and incomplete representation of tumor characteristics. To address these limitations, this dissertation introduces a supervised multi-view contrastive learning framework (MMCon) for ART eligibility prediction. By combining radiomics features (CT and MRI scans with radiotherapy) and dose information, MMCon enhances feature representation through a contrastive learning process that improves separability between patient groups. To further refine the embedding space, an additive margin is applied to the learning objective, ensuring the alignment of imaging features with clinical labels. This framework effectively models tumor heterogeneity and achieves high accuracy in ART prediction. A complementary study focuses on improving the generalizability of ART prediction models across clinical samples from different hospitals. This approach integrates radiomics, dosiomics, and geometric features from patient data, where radiomics includes contrast-enhanced CT, T1-weighted (T1-w) MRI, T2 MRI. Geometric features quantify spatial relationships between tumor volumes and nearby organs at risk, providing additional predictive context. Relevant features are selected using least absolute shrinkage and selection operator (LASSO) and combined with clinical parameters to construct a multi-omics nomogram, achieving robust performance. For distant metastasis prediction, this dissertation develops a multikernel-based framework to integrate radiomics, morphology, and dosiomics features extracted from the contralateral parotid gland. This method mitigates nonlinearity data distribution through radial basis function (RBF) mapping, which transforms features into a high-dimensional space linearity distribution for classification. Additionally, a label-refinement strategy is employed to improve model flexibility in handling high-dimensional data, enabling accurate metastasis risk prediction. The proposed methodologies significantly enhance performance across the three prediction tasks. For ART eligibility prediction, MMCon achieves AUC values between 0.92 (*100%) and 0.97 (*100%) across multi-omics datasets, demonstrating robust performance in multi-omics data analysis. The Genomap-based framework improves model transferability, achieving consistent accuracy in diverse patient populations with an external validation AUC of 0.96 (*100%). For distant metastasis prediction, the kernel-based method provides a reliable solution for integrating heterogeneous data sources and supports accurate metastasis risk stratification. The model demonstrates strong predictive performance, with an average AUC of 0.93 (*100%) across validation datasets. In conclusion, this dissertation explores key challenges in NPC prognosis prediction and proposes machine learning frameworks that aim to improve prediction accuracy, reliability, and applicability. These approaches integrate different medical imaging and multi-omics data to provide practical tools for personalized treatment planning and disease management. The proposed methodologies establish a strong foundation for future research in predictive oncology and enable the development of tailored therapeutic strategies for NPC patients.