Dual-attention bidirectional LSTM with feature genomic analysis improves prognostic survival prediction in colorectal cancer patients

The increasing incidence and mortality rates of colorectal cancer necessitate accurate prediction of patients' prognostic survival time for better management, early screening, and extended lifespan. This study uses the TCGA public dataset to conduct differential analysis on lncRNAs in 39 diseased tissues and their normal counterparts from 413 colorectal cancer patient samples, identifying 458 differentially expressed lncRNAs (DELncRNAs). Univariate Cox regression analysis revealed 23 DELncRNAs significantly associated with overall survival (OS). These 23 DELncRNAs were further refined using the LASSO algorithm to determine their feature coefficients. An adaptive mining approach with dual-attention mechanisms was employed to explore the correlative properties between various factors and survival time. A bidirectional long short-term memory (BiLSTM) neural network was established for survival prediction. The model was validated using the Jiangnan University colorectal cancer dataset, demonstrating reliable predictions for patient survival and valuable support for clinical decision-making. The AUC values for patient survival prediction during the 3-year, 3-6 year, and 6-year periods were nearly 1.00, significantly outperforming other comparative trials.