Manually-Curated Versus LLM-Generated Explanations for Complex Patient Cases: An Exploratory Study with Physicians

Multimorbdity guideline-based clinical decision support systems (MGCDSSes) have emerged to optimize outcomes for multimorbid patients by generating personalized treatment plans that consider many clinical data sources. The success of these systems relies on their ability to explain treatment rationale, fostering trust in their outcomes among physicians. While traditionally developing treatment explanations required significant manual effort from physicians, the emergence of large language models (LLMs) offers potential to automate and simplify this process. LLMs like Meditron70B have shown promise in generating treatment explanations, saving time and resources for physicians. However, questions remain regarding the accuracy and depth of LLM-generated explanations. In this work, we evaluate the performance of Meditron70B in generating treatment explanations within our MitPlan MGCDSS using a physician-focused survey. We highlight both the promise and potential limitations of using LLMs for this purpose.