Evaluation of LLM-generated narratives based on patient-reported outcome measures

Patient-reported outcome measures (PROMs) are instruments, often validated questionnaires, that capture patients’ experiences of symptoms, functioning, and quality of life. While useful, patients often receive little or no personalized feedback based on their PROMs responses, limiting their potential to support understanding, reflection, and management of their health. One promising way to provide such feedback is through patient narratives, which can translate structured questionnaire data into coherent stories that may enhance interpretability and emotional resonance (Boomstra et al., 2024). In practice, however, providing personalized narratives for each individual PROMs profile is not feasible at scale due to time and resource constraints. Given recent advances in large language models (LLMs) and their potential of generating natural-language narratives from structured numeric input, the goal of this work is to examine if LLMs can generate personalized patient narratives that both accurately represent PROMs data and are of sufficient quality to be useful and meaningful. The present study therefore investigates whether and how LLMs can generate narratives based on PROM data, focusing on the accuracy and quality of the generated texts. This study aims to provide evidence on the feasibility of using LLM-generated narratives as personalized PROMs feedback.