By using large language models, this paper aims to present an optimized framework for extracting medical knowledge that is cross-lingual and disease diagnosis. It will address the language variance and terminological inconsistency by fine-tuning LLMs for medical cross-lingual text processing. We evaluate the model on a rich multilingual dataset with over 10 million clinical entries across five languages in terms of disease diagnosis accuracy of 92% and medical term extraction precision of 88%, which sets the baseline models well behind by nearly 15% in terms of diagnostic accuracy. Specific forms of adaptation, such as domain adaptation and cross-lingual training, were shown to be critical factors that improve performance across languages. As such, these results hold the promise of optimized LLMs to deliver specific, multilingual AI-driven healthcare solutions advancing inclusion in global applications for medical AI.