Assessing equitable use of large language models for clinical decision support in real-world settings: fine-tuning and internal-external validation using electronic health records from South Asia