<p>The early diagnosis of chronic diseases is pivotal for improving patient outcomes, yet it remains challenging due to the subtle and complex nature of early-stage pathological patterns in heterogeneous biomedical data. This review systematically critiques the paradigm shift from manual feature engineering to smart feature extraction using deep learning (DL) for this purpose. We provide a structured analysis of how convolutional networks (CNNs), recurrent models (RNNs), transformers, and autoencoders autonomously learn hierarchical, clinically relevant features directly from raw data modalities, including medical images, physiological signals, and electronic health records. A core contribution of this work is a critical comparative analysis of these architectures, evaluating their trade-offs in terms of performance, computational efficiency, and applicability across major chronic diseases such as cardiovascular disorders, diabetes, and neurodegenerative conditions. Furthermore, the review synthesizes the pathway for integrating these extracted features into next-generation Medical Decision Support Systems (MDSS), offering an in-depth discussion on the pivotal challenges of real-world deployment. These include ensuring model explainability (XAI), achieving robustness across diverse clinical populations, and navigating regulatory and interoperability hurdles. Bridging the gap between algorithmic innovation and clinical translation, this review serves as a comprehensive guide for developing scalable, trustworthy, and effective AI-driven diagnostic systems.</p>