Mitigating the Risk of Health Inequity Exacerbated by Large Language Models