Pathway-Aware Multimodal Transformer for Explainable Survival Prediction in Bladder Cancer

Authors:

Mr Gowtham Raj M¹,Arun Chavan²,Bhuvan B³ ,Charan Sai Tej K V⁴ ,Charan Teja G S⁵ Assistant Professor, Dept of CSE, KSIT, Karnataka, India¹

Student, Dept of CSE, KSIT, Karnataka, India^2-5

***

ABSTRACT – Cancer survival prediction requires effective integration of heterogeneous biomedical modalities such as gene expression profiles and histopathological whole-slide images (WSIs) to capture both molecular and morphological characteristics of tumors. Existing approaches often rely on unimodal learning or coarse multimodal fusion, limiting their ability to model fine-grained genotype-phenotype interactions. This paper presents a survey and system-oriented analysis of a pathway-aware multimodal framework for bladder urothelial carcinoma (BLCA) survival prediction based on the Pathway-Aware Multimodal Transformer (PAMT). The framework organizes gene expression data into biologically meaningful pathway representations while processing WSIs using patch-based feature extraction and transformer-based encoding. A contrastive alignment mechanism and pathway-to-patch cross-attention fusion module enable biologically guided multimodal interaction learning between genomic pathways and localized tissue regions. The integrated representation is utilized for survival risk estimation using a Cox proportional hazards-based objective, while attention-based interpretability mechanisms provide insights into prognostically significant pathways and tissue structures. The survey findings indicate that pathway-aware transformer architectures offer improved multimodal interaction modeling, enhanced interpretability, and strong potential for clinically explainable cancer survival prediction systems.

KEY WORDS: Multimodal Learning, Survival Prediction, Bladder Cancer, Transformer Architecture, Explainable AI