AI-Based Skin Disease Prediction System Using ConvNeXt-Tiny with Explainable Visualization and Phase Guidance

Skin diseases are common, but seeing a dermatologist is not easy for everyone, especially in under-served regions. Many deep-learning classifiers are heavy, hard to interpret and stop at a class label. We present MediCare AI, a lightweight screening system that sorts skin images into seven disease categories by fully fine-tuning an ImageNet-pretrained ConvNeXt-Tiny with AdamW, cosine learning-rate decay, label smoothing and early stopping. Gradient-weighted Class Activation Mapping shows which parts of the image drove each prediction. A Streamlit app accepts uploaded photos and live camera frames, reports the confidence and a rough three-level visible phase, and gives treatment suggestions and preventive advice through a generative service, with a local offline fallback. A single network reached roughly 80%–90% validation accuracy. The tool is meant for education and early screening, not as a replacement for professional diagnosis. Index Terms—skin disease classification, ConvNeXt, transfer learning, Grad-CAM, explainable artificial intelligence, Streamlit.