Development of a Multimodal Predictive System for Skin Cancer: A Systematic Review of Deep Learning and Data Fusion Techniques

Skin cancer, particularly melanoma, remains one of the most lethal yet preventable malignancies worldwide, necessitating timely and accurate diagnosis to improve patient survival. Traditional diagnostic methods such as visual inspection and dermoscopy are often limited by subjectivity and access to specialist care, especially in resource-constrained settings. Recent advances in artificial intelligence (AI) and deep learning (DL) have demonstrated significant potential in augmenting diagnostic accuracy by analyzing dermoscopic images. However, image-only models overlook crucial contextual factors in clinical metadata such as age, sex, lesion location, and family history. This systematic review synthesizes current research on multimodal predictive systems integrating dermoscopic imagery with structured clinical metadata using deep learning frameworks. A total of 30 eligible studies published between 2019 and 2024 were critically analyzed, highlighting the use of datasets such as ISIC and HAM10000, and exploring architectures including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), hybrid models, and attention-based mechanisms. Data fusion strategies, especially late fusion and attention-enhanced models, proved most effective, with several models achieving classification accuracies exceeding 90%. Despite promising outcomes, challenges such as limited availability of multimodal datasets, lack of model interpretability, and inadequate externa lvalidation hinder real-world clinical translation. The review underscores the need for explainable AI, equitable dataset representation, and federated approaches to ensure ethical and scalable deployment. This work provides a consolidated reference for future research on robust, interpretable, and clinically applicable AI-driven skin cancer diagnostic systems.