The contradiction between the medical services provided by China’s healthcare industry and the increasingly growing demand for medical care is becoming more prominent. “AI + healthcare” intelligent diagnosis is an effective way to alleviate this contradiction. Large Language Models (LLMs), with their exceptional natural language understanding and reasoning generation capabilities, have opened new pathways for deeply understanding complex medical texts. However, directly applying LLMs to the medical field may result in hallucinations due to the lack of domain-specific medical knowledge. Additionally, retraining LLMs to update their parameters is not only time-consuming and labor-intensive but also costly. To address this, we adopted the LoRA fine-tuning technique, using Chain-of-Thought (CoT) to design prompt templates for training and inference on Qwen2.5 and InternLM2.5. To further enhance prediction accuracy, we introduced a voting mechanism, selecting the top three models with the highest prediction scores for model fusion. During the post-processing stage, we leveraged the powerful capabilities of LLMs in understanding, generation, and correction to carefully evaluate the top two predictions selected after voting, ultimately determining the best answer option. Experimental results show that our method achieved an F1 score of 99.05% in the preliminary round, with an F1 score of 92.10% in the semi-final, securing the fourth place in the overall ranking, which fully validates the effectiveness of this approach.