Survival analysis, is a widely used technique for analysing time-to-event data, inclusive of censored data. Even though, numerous survival analysis approaches have performed well but still have some underlying assumptions and other limitations. To overcome these assumptions and limitations, a novel encoder-based transformer model “ \(E_{no}UTSurv\) ” model, with dynamic adaptive computation time to predict the risk of heart failure has been proposed. The proposed model has better calibration and discriminative performance when compared to the state-of-the-art survival models. Additionally, it exhibits significant reductions in memory requirements (over 50%) and execution time (over 70%) when compared to transformer-based models, while maintaining or surpassing their performance, thus tackling the high computational requirements of the transformer architecture. To evaluate the scalability of the proposed model, its performance has been evaluated on the augmented dataset and the proposed model showcased similar enhanced performance. Thus, our experiments show that the proposed model has enhanced efficiency, optimal computational resource requirements and is scalable.
