SAMamba: Integrating State Space Model for Enhanced Multi-modal Survival Analysis

Xinyue Li

Abstract:

Survival prediction represents a challenging ordinal
regression task, involving modeling of intricate interactions
among various data modalities. The recent evolution of state
space models, Mamba in particular, has opened new vistas for
effectively processing sequence data, including genomic profiles
and gigapixel pathology Whole Slide Images (WSIs). In light
of these advancements, we propose Survival Analysis Mamba
(SAMamba), a novel approach that melds the Mamba framework
with multi-modal survival prediction. Specifically, we propose
a patch clustering layer to identify morphological prototypes
from the extensive collection of patches within WSIs and employ
gene set enrichment analysis to explore the biological associations
between pathways and gene sets for enhanced and robust feature
representation. Subsequently, we introduce Mamba structures to
capture the intrinsic relationships within pathology WSIs and
genomic profiles with linear computational complexity. Additionally,
we utilize multi-modal attention to seamlessly integrate
multi-modal data and design a self-attention pooling module to
further refine insights from each data modality for enhanced survival
outcome prediction. Extensive experiments on four public
TCGA datasets are conducted to validate the effectiveness of our
proposed SAMamba, using ablation studies, statistical analysis,
and visualization. The experimental results demonstrate that our
method achieves superior performance compared to state-of-theart
methods, highlighting the potential of the proposed SAMamba
for multi-modal survival outcome prediction. Our code will be
released at https://github.com/coffeeNtv/SAMamba.