Background Electronic Medical Records (EMRs) play a pivotal role in training large-scale language models (LLMs) within the medical domain.

Objective To explore the value of electronic medical record data by studying a three-stage training paradigm based on a general large language model.

Methods Firstly, in the continued training phase, extensive EMR texts were employed to further train the pre-existing general model, thereby enhancing its medical-specific linguistic knowledge. Secondly, during the supervised fine-tuning phase, annotated EMR data were utilized to modify the model for specific clinical tasks such as medical named entity recognition and clinical trial screening, enabling the model to acquire specialized task-oriented skills. Finally, in the reinforcement learning phase, feedback from doctors was integrated to optimize the model's outputs, improving the accuracy and interpretability of decision-making.

Results Experimental results demonstrated that the model's performance in clinical tasks was significantly enhanced, the occurrence of hallucinations was mitigated, and the reliability of its outputs was improved.

Conclusion This study provides an effective approach for constructing standardized and trustworthy medical LLMs, offering substantial practical application value.