Generating subjective, objective, assessment, and plan (SOAP)-structured medication logs using DeepSeek-R1 through...

Area of Science:

• Artificial Intelligence in Pharmacy
• Clinical Informatics
• Large Language Models (LLMs)

Background:

• Clinical pharmacists manage complex medication logs, requiring efficient data processing.
• Standardized documentation, such as Subjective, Objective, Assessment, and Plan (SOAP) notes, is crucial for patient care.
• Large Language Models (LLMs) show potential for automating clinical documentation tasks.

Purpose of the Study:

• To evaluate the feasibility of using the open-source LLM DeepSeek-R1 for generating SOAP-format medication logs.
• To assess the model's potential to support clinical pharmacists in oncology settings.
• To compare the performance of single-source versus multi-source information inputs for log generation.

Main Methods:

• Utilized 30 oncology medication profiles to extract 80 days of logs, converted into simulated pharmacist-patient dialogues.
• Compared single-information-source (dialogue only) with multi-information-source (dialogue plus patient data, records, test results) inputs.
• Employed five prompts of increasing complexity and evaluated performance using BERT, ROUGE, and a blinded expert Seven-Dimension Index (7DI) metric.

Main Results:

• DeepSeek-R1 effectively generated structured SOAP logs with multi-source information and complex prompts (Prompts 4 and 5).
• Multi-source inputs significantly outperformed single-source dialogues, confirmed by machine scores (BERT-F1, ROUGE) and manual 7DI evaluation.
• Model output quality was dependent on input data completeness, requiring pharmacist review to correct errors for improved scores.

Conclusions:

• DeepSeek-R1 demonstrates utility in generating SOAP medication logs using prompt-engineered multi-source clinical information.
• The LLM can potentially enhance clinical pharmacists' efficiency, but output reliability necessitates thorough pharmacist review.
• Future research should address oncology-specific limitations, AI hallucinations, and explore comparisons with other LLMs and explainable AI.