AHDSN: an attention-enabled hybrid deep sequential network for cancer survivability prediction from multi-omics data

Cancer is the leading threat to human health and lifespan. Every day, the number of deaths caused by cancer continues to rise. Therefore, accurately predicting survivability from cancer has become an important area in cancer research. In predicting survivability, multi-omics data is advantageous as it provides information from different molecular levels of human biological processes, encompassing different omics such as genomics, epigenomics, transcriptomics, proteomics, and metabolomics. In this article we introduce a novel method called Attention-Enabled Hybrid Deep Sequential Network (AHDSN) which utilizes Long Short-Term Memory, Bidirectional Gated Recurrent Unit, and the attention mechanism to extract latent features from multi-omics data and Dense layers with softmax activation function for classification. Unlike conventional approaches that predict survival at a fixed time point (e.g., 5-year survival), the proposed AHDSN method predicts overall survival across the complete follow-up period using each patient's survival time and censoring status. We evaluated the proposed AHDSN method against several state-of-the-art approaches to assess their relative performance in survivability prediction from multi-omics data. To address class imbalance, both Random Oversampling (ROS) and Synthetic Minority Oversampling Technique (SMOTE) are applied during preprocessing to ensure a more balanced distribution of samples across classes. The experimental results show that the proposed AHDSN method surpassed other state-of-the-art methods in terms of accuracy, precision, recall, and [Formula: see text]-score across five multi-omics cancer datasets, Glioblastoma, Colon, Breast, Kidney, and Lung, achieving accuracies of 98.33%, 96.00%, 97.14%, 88.24%, and 80.00% when using ROS, and 97.12%, 96.00%, 96.22%, 85.18%, and 80.00% when using SMOTE respectively. Confidence Interval test also demonstrates the superiority of the proposed AHDSN method compared to other existing methods in producing the lowest error rate and the smallest error bound for all five multi-omics datasets. Additionally, SHapley Additive exPlanations analysis and heatmaps are employed to explain feature importance and illustrate how individual omics features contribute to model classification. Furthermore, the ablation study confirms the synergistic benefit of the proposed hybrid architecture and validates the importance of each component.