Big Data Driven Multimodal Classification Using Clipped RBMs and Cross Modality Attention in MMDBN

Advancements in medical imaging and data acquisition have led to an exponential increase in high-dimensional, heterogeneous cancer data, necessitating scalable and intelligent diagnostic frameworks. In response, we propose a Modified Multimodal Deep Belief Network (MMDBN) architecture, built entirely upon Clipped Restricted Boltzmann Machines (CRBMs) with modified Contrastive divergence and augmented with a Cross-Modality Attention Fusion (CMAF) mechanism. This architecture is optimized for distributed big data environments, enabling real-time, high-throughput analysis of brain, breast, and bone cancer modalities. The Clipped RBM layers ensure bounded activation dynamics for robust unsupervised feature learning, mitigating instability and overfitting in large-scale training scenarios. CMAF adaptively weighs modality-specific representations per instance, improving generalization and interpretability, especially under incomplete or noisy modality conditions. Fine-tuning of the stacked network leverages supervised learning to optimize discriminative capacity across modalities. Empirical evaluation on benchmark medical datasets demonstrates the superiority of the proposed model, achieving 96.78% classification accuracy, with an AUC-ROC of 94.80, outperforming conventional DBN, CNN, and SVM-based baselines. This work highlights a significant advancement in deep multimodal learning for oncology, bridging the gap between data-intensive computation and clinically relevant cancer diagnosis.