Can Generative AI Deliver Patient-Friendly Summaries? A Case Study Using NICE Guidance for Spinal Muscular Atrophy
Author(s)
Manuel Cossio, MPhil, MS1, Ramiro Eugenio Gilardino, MHA, MSc, MD2.
1Director, Artificial Intelligence Lead, Cytel, Dubendorf, Switzerland, 2Independent, Dubendorf, Switzerland.
1Director, Artificial Intelligence Lead, Cytel, Dubendorf, Switzerland, 2Independent, Dubendorf, Switzerland.
OBJECTIVES: This study evaluated the feasibility of using Google Gemini, a generative AI model, to produce plain language summaries (PLS) for patients from health technology assessment (HTA) evidence documents, using a NICE Highly Specialized Technologies (HST) guidance as a case example. The aim was to explore AI’s potential to support more equitable and meaningful patient involvement in HTA by automating the production of these summaries.
METHODS: A prompt template was adapted from the Scottish Medicines Consortium’s Summary Information for Submitting Patient Groups (v4, 2017). Iterative prompt refinement (4 rounds) was applied to optimize clarity, structure, and completeness. The AI-generated PLS was evaluated across eight document sections using 18 criteria (scored 1-10), assessing readability, relevance, and patient-friendliness. A human-in-the-loop reviewer provided expert feedback on language, bias, and patient comprehensibility. The evaluation also considered the PLS’s alignment with goals of informed participation, transparency, and regulatory equity expectations (e.g., EU HTA Regulation).
RESULTS: The model created a well-structured, 8-page (2,570-word) PLS from the NICE guidance on onasemnogene abeparvovec for spinal muscular atrophy in only 15 seconds (https://tinyurl.com/y5vrpupy). The average quality score was 8.27/10, with top scores in treatment explanation (9.2) and readability (8.9). Lower performance was noted in comparator clarity (6.8) and articulation of unmet need (6.5). The output met CEFR B1 readability and was largely jargon-free. Reviewers noted the summary aligned well with PLS guidance, though minor contextual refinements would be required to meet real-world HTA standards for submission.
CONCLUSIONS: Generative AI shows strong potential to support PLS development, enhancing transparency, efficiency, and inclusion in HTA processes. Broader validation and integration into HTA workflows are recommended to promote consistent, scalable, and patient-centered communications.
METHODS: A prompt template was adapted from the Scottish Medicines Consortium’s Summary Information for Submitting Patient Groups (v4, 2017). Iterative prompt refinement (4 rounds) was applied to optimize clarity, structure, and completeness. The AI-generated PLS was evaluated across eight document sections using 18 criteria (scored 1-10), assessing readability, relevance, and patient-friendliness. A human-in-the-loop reviewer provided expert feedback on language, bias, and patient comprehensibility. The evaluation also considered the PLS’s alignment with goals of informed participation, transparency, and regulatory equity expectations (e.g., EU HTA Regulation).
RESULTS: The model created a well-structured, 8-page (2,570-word) PLS from the NICE guidance on onasemnogene abeparvovec for spinal muscular atrophy in only 15 seconds (https://tinyurl.com/y5vrpupy). The average quality score was 8.27/10, with top scores in treatment explanation (9.2) and readability (8.9). Lower performance was noted in comparator clarity (6.8) and articulation of unmet need (6.5). The output met CEFR B1 readability and was largely jargon-free. Reviewers noted the summary aligned well with PLS guidance, though minor contextual refinements would be required to meet real-world HTA standards for submission.
CONCLUSIONS: Generative AI shows strong potential to support PLS development, enhancing transparency, efficiency, and inclusion in HTA processes. Broader validation and integration into HTA workflows are recommended to promote consistent, scalable, and patient-centered communications.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
PT21
Topic
Health Technology Assessment, Patient-Centered Research
Topic Subcategory
Instrument Development, Validation, & Translation
Disease
Neurological Disorders, Rare & Orphan Diseases