One of the most prevalent cancers is bladder cancer, and non-muscle-invasive bladder cancer (NMIBC) has a high recurrence rate, therefore early detection is essential for efficient patient care. This work combines longitudinal clinical data and histological pictures to provide a deep learning-based method for bladder cancer recurrence prediction. Convolutional neural networks (CNNs), which is fine-tuned VGG16 and long short-term memory networks (LSTMs), which is stacked bidirectional GRU-LSTM were combined in a hybrid model that was improved by an attention mechanism to collect temporal and spatial data. Large-scale datasets were used for training and validation, and the model performed better than conventional techniques, achieving 90% accuracy, 88% precision, 85% recall, and 86% F1-measure. In accordance with clinical findings, the model identified vital factors such as tumor size, recurrence intervals, and treatment protocols. Attention maps, which highlighted important visual areas and temporal points, substantially improved interpretability. By facilitating individualized treatment planning, this method helps physicians to optimize therapeutic treatments and stratify patients according to recurrence risk.