Multimodal Deep Learning in Healthcare Recommender Systems: A Review of CNN-Based Architectures

Abstract

Recent multimodal deep learning approaches are changing how healthcare systems process and integrate widely different types of patient information. Based on this, in this review we investigate the Convolutional Neural Network CNN-based architectures in Healthcare Recommender Systems HRS, which incorporate a variety of data from different sources, such as medical imaging, Electronic Health Records EHRs, wearable sensor streams, clinical narratives. At the core of this ecosystem, CNNs learn hierarchical representations from high-dimensional visual and other unstructured medical data and are increasingly the feature-extraction backbone of modern healthcare recommendation pipelines. We first discuss CNN-based architectures for HRS and then examine their combination with traditional recommendation schemes such as collaborative filtering, hybrid recommendation techniques and other multimodal fusion methods. The review furthermore covers new research developments to leverage CNN-based HRS, including vision–language models, reinforcement learning, causal inference, fairness-aware optimization, federated learning and edge/IoMT deployment. This review reports methodological advances and recent challenges. These challenges include generalization, interpretation , privacy preserving, and clinical integration. Taken together, these analyses summarizes the progress of recent research and also provide open challenges. The goal is to guide the design of transparent and scalable systems. These systems should be also ethically sound. They focus on CNN-based healthcare recommender systems for next-generation personalized medicine.