Large language models in clinical psychiatry: Applications and optimization strategies

Table 1 Technical features of large language models

Technologies	Characteristics
PT	Large-scale medical knowledge learning: Through training on massive medical literature and case data, the model captures medical language patterns and basic pathological features
	Reducing annotation dependency: The model can utilize unannotated medical texts (such as electronic health records and papers) for initial training
	Versatility foundation: It provides a general medical semantic understanding capability for subsequent tasks (such as diagnosis and report generation)
SFT	Precise task adaptation: Optimize model performance for specific medical tasks, such as disease classification and image recognition
	High accuracy: Enhance the reliability of the model in specialized areas through professionally annotated data, such as cases labeled by doctors
	Enhanced compliance: Adjust model outputs to meet privacy or ethical requirements, such as anonymization processes
Agent	Automated processes: Performing repetitive tasks such as medical record organization and appointment reminders to enhance healthcare efficiency
	Multimodal interaction: Enabling patient-doctor communication and report interpretation through a combination of voice, text, and image
	Real-time decision support: Dynamically providing diagnostic and treatment suggestions, such as drug titration, in conjunction with a rule engine
RAG	Real-time knowledge integration: Incorporate the latest medical databases, such as PubMed and clinical guidelines, to prevent outdated knowledge within the model
	Evidence traceability: Generate results accompanied by references to facilitate verification of reliability by medical professionals
	Mitigation of hallucination risk: Generate content based on authoritative knowledge bases to minimize the likelihood of the model fabricating medical information
PE	Output controllability: Structured instructions guide the model to generate standardized results
	Flexible domain adaptation: Adjusting prompt words can quickly switch application scenarios
	Reduced training costs: Optimizing performance on specific tasks (such as improving the accuracy of rare disease descriptions) without retraining the model

PT: Pre-training; SFT: Supervised fine-tuning; RAG: Retrieval-augmented generation; PE: Prompt engineering.