Conventional mathematical models and data-driven deep learning (DL) methods are both tested for Human Decision-Making (HDM) in Complex Sequential Tasks (CST). These methods include complicated cognitive functions. Considering the use of adaptive timing mechanisms, the present research introduces a hybrid model that includes statistical cognitive dynamic models with Bidirectional Long Short-Term Memory (BiLSTM) attention networks. While DL serves as the basis for collecting temporal behavioural patterns, the mathematical model is vital for recording value changes, cognitive load, and decision selection by its application of classical Boltzmann logic. The mathematical model of contraction mapping serves as the basis for the invention of convergence guarantees. The recommended model achieves 84.7% prediction accuracy and improved cognitive state alignment (CSD=0.187) when tested on the Iowa Gambling Task dataset, involving 617 participants. The proposed model improves the accuracy of mathematical models by 12.4% and the validity of DL by 3.5%. This research has enabled the interpretation of artificial intelligence (AI) for human-centred decision support systems.