Review of Visual Analytics in Healthcare Big Data: From CNNs to Multimodal Deep Learning Models

The rapid expansion of healthcare big data, driven by electronic health records, wearable devices, and advanced medical imaging, has created both opportunities and challenges for clinical practice. Deep learning models such as CNNs, RNNs, transformers, GANs, and GNNs demonstrate strong predictive capabilities, yet their interpretability remains a major concern for real-world deployment. This review explores visualization and interpretability techniques that bridge the gap between complex models and medical decision-making. We first analyze CNN-based methods like saliency maps and Grad-CAM for imaging tasks, followed by RNN and attention mechanisms that capture temporal patient trajectories. Transformer-based attention maps are examined for disease localization, while GANs and GNNs highlight data augmentation and structural interpretability. Applications in diabetic retinopathy classification, biomarker activation, and multi-lesion/multi-scale networks illustrate how visualization enhances clinical trust and diagnostic transparency. Despite these advancements, challenges persist, including lack of standardization, computational costs, qualitative-heavy evaluation, and real-time adaptability. We conclude by emphasizing the need for standardized protocols, multimodal integration, and real-time interpretability tools, envisioning a future where AI-driven visualization becomes an integral component of safe and reliable medical practice.