Ai-generated personalized informed consent for elective spinal surgery: a comparative study using retrieval-augmented...

Informed consent (IC) documents in spine surgery frequently lack procedure-specific risk data, quantitative complication rates, and discussion of alternative treatments. These deficiencies impair patient comprehension and constitute a leading basis for malpractice litigation, with failure to obtain adequate consent frequently cited among the most common allegations in spine surgery claims. Large language models (LLM) have shown promise in generating readable consent forms, yet their output may include inaccurate statements. Retrieval-augmented generation (RAG), which ground model output in a user-defined knowledge base may overcome this limitation, but have not been applied to IC document production. This study aimed to generate RAG-based personalized IC documents for elective spinal surgery and to compare their quality with standard consent forms.

Standard consent forms for eight spinal procedures obtained from the national neurosurgical society served as the control group. For the study group, consent documents for the same procedures were generated using NotebookLM, a RAG platform that grounds output in user-defined sources with automatic citation. Three strategies ensured the documents were evidence-based, personalized, and readable: (1) a knowledge base combining North American Spine Society guidelines with procedure-specific literature retrieved from PubMed; (2) three patient archetypes per procedure whose clinical variables were incorporated into generation prompt; and (3) a prompt instruction targeting sixth-grade reading level. Quality was compared using the Spatz instrument, scored by three blinded evaluators. AI documents were further assessed for personalization using a four-domain rubric, content accuracy through hierarchical verification of all factual claims, and readability using Flesch indices.

AI-generated documents scored significantly higher than standard forms on the modified Spatz instrument (14.75 ± 0.35 vs. 10.17 ± 0.69 out of 15; p < 0.001), with 75% achieving the maximum score. Mean personalization score was 7.44/8 (93.1%). Across 661 verified factual claims, 89.0% carried automatic citations; citation accuracy was 98.1%, reporting accuracy was 99.7%, and the fabrication rate was 0.76%. Mean Flesch-Kincaid Grade Level was 7.76, with 95.8% of documents meeting the recommended reading level threshold. Inter-rater reliability was good (ICC = 0.844).

When grounded in a curated evidence base and guided by a structured prompt, a RAG-based system produced consent documents that were superior in content quality to standard forms, personalized to individual patient profiles, traceable to their cited sources, and written at an appropriate reading level. This workflow may reduce the effort needed to prepare appropriate IC documents while providing a verifiable disclosure record for medicolegal documentation.