In this study, we investigated a hybrid framework that integrates large language models (LLMs) with conventional machine learning for early-stage skin lesion assessment using the UCI dermatology dataset as a proxy for early skin cancer detection. We first developed a baseline model using only structured clinical and histopathological attributes and trained classical classifiers, with a gradient boosting model achieving an accuracy of 0.89, macro-averaged F1-score of 0.87, and macro-AUC of 0.93. We then generated textual summaries for each patient case and used an LLM to derive high-level semantic features, such as inferred risk level and lesion-type descriptors, which were added to the structured feature space. This structured-plus-LLM-features configuration improved performance to an accuracy of 0.92, macro-averaged F1-score of 0.91, and macro-AUC of 0.96, indicating that LLM-derived features captured clinically meaningful abstractions not fully exploited by the baseline model. Finally, we implemented a hybrid decision-refinement approach in which a primary gradient boosting classifier handled most cases, while low-confidence predictions were escalated to the LLM for refined diagnostic suggestions. This hybrid model achieved the best results, with an accuracy of 0.94, macro-averaged F1-score of 0.93, and macro AUC of 0.97, and demonstrated fewer misclassifications across challenging classes. These findings suggest that LLMs can enhance structured-data models both as semantic feature generators and as second-stage reasoning engines, offering a promising and interpretable pathway for embedding AI-driven decision support into dermatology workflows aimed at earlier and more reliable skin lesion risk stratification.