Temporal–contextual contrastive learning with importance-guided feature fusion using an improved particle swarm optimization algorithm

The classification of cancer subtypes is essential for individualized therapy. Subtype prediction may be accomplished by using multimodal data gathered from subjects. Multimodal cancer data contains concealed joint information that existing vector based fusion algorithms cannot sufficiently exploit. We propose a MultiModal Temporal Contextual Contrastive Learning with Importance-Guided Feature Fusion (MMTCLIF) network, which employs a singular value decomposition bidirectional long short term memory autoencoder (BiLSTM-AE) to achieve excellent portrayal of sequential relationships among multimodal cancer data. An enhanced Particle Swarm Optimization (PSO) algorithm is used to tune the training hyperparameters of the MMTCLIF network to increase cancer classification performance. Our artificial intelligence based method effectively extracts data characteristics and integrates self-supervised contrastive learning, optimizing data use without the need for significant labeled data. Extensive experiments on five The Cancer Genome Atlas (TCGA) multi-omics cancer datasets demonstrate that MMTCLIF achieved classification accuracies ranging from 0.8824 to 0.9341 and macro Area Under the Curve (AUC) values between 0.9550 and 0.9869, consistently outperforming competing methods. Ablation studies verify the effectiveness of each major component of the proposed framework, confirming that the integration of importance guided feature selection, temporal and contextual contrastive learning, and adaptive hyperparameter optimization collectively contributes to superior cancer subtype classification performance.