An adaptive language model-based intelligent medication assistant for the decision support of antidepressant...

Depression, a pervasive mental disorder affecting millions worldwide, requires a holistic and personalized approach to treatment. Combining therapy and antidepressant medication is crucial, but selecting the right medication remains a complex task due to the variety of options available, each with unique, contraindications and drug interactions. While intelligent clinical decision support systems exist, there is a gap in medication-specific solutions. This paper proposes an innovative intelligent medication assistant, leveraging retrieval-augmented generation and large language models to provide personalized recommendations considering patient records and current health situation.

The proposed assistant tailors recommendations based on individual context by injecting patient-specific documents into large language models, enhancing efficiency and minimizing adverse effects. Our approach adapts to various query formats and dynamically incorporates relevant sources, empowering healthcare providers to make informed decisions. The analysis covers 40 clinical cases of patients requiring psychiatric monitoring. The patient's basic biography, clinical context, and ongoing situation were presented for each model, followed by questions about antidepressant pharmacological treatment.

The case study used 40 mental health patient clinical cases, and the results showed that the proposed assistant represents a significant advancement for the health provider, enabling the support, fast search, and consultation of medication information leaflets in almost real-time. A total of ten large language models with two configurations each were evaluated. Two commercial models and eight open local models were used.

The commercial model gpt4o had a total score of 4.10/5.00, while the open model llama3 had a total score of 2.92/5.00. The results showed that commercial models present better results than open models, providing more accurate answers, considering the patients' context.