TL;DR
Researchers have developed an AI-human collaborative framework using GPT-4 to bridge the "literacy gap" in surgical informed consent. By simplifying complex institutional forms and generating new, procedure-specific ones, the team improved readability from a college-freshman level to an 8th-grade level (and as low as 6th-grade for new forms) without sacrificing medical or legal accuracy.
What Problem Does This Paper Address?
Informed consent is a cornerstone of medical ethics, yet the documents used to obtain it are notoriously difficult to understand. Most surgical consent forms are written at a college level, while the average American adult reads at an 8th-grade level. This "literacy gap" creates significant barriers to patient autonomy and is a known driver of health disparities.
Furthermore, many hospitals use "one-size-fits-all" generic forms that lack the nuance of specific procedures. While surgeons often supplement these with verbal explanations, the written record—the legal safety net—remains opaque to the very people it is meant to protect. The authors sought to determine if Large Language Models (LLMs) could automate the simplification process while maintaining the rigorous clinical and legal standards required for surgery.
Methods
The study employed a two-pronged approach:
- Simplification of Existing Forms: The team collected generic consent forms from 15 major academic medical centers across the US. They used GPT-4 to simplify the text with a standardized prompt: "While still preserving the same content and meaning, please convert to the average American reading level."
- De Novo Generation: They prompted GPT-4 to create entirely new, procedure-specific consent forms for five diverse operations (ranging from brain surgery to skin cancer removal).
The "Human-in-the-Loop" Validation: To ensure the AI didn't "hallucinate" or omit critical risks, the researchers implemented a multidisciplinary review process:
- Medical Review: Three physician authors independently compared original and simplified forms.
- Legal Review: A medical malpractice defense attorney verified that the simplified versions maintained legal sufficiency.
- Rubric Scoring: The de novo forms were evaluated using a validated 8-item rubric (the Spatz rubric) and scrutinized by subspecialty surgeons for clinical accuracy.
Datasets
- Institutional Forms: Consent documents from 15 diverse academic medical centers (7 Northeast, 2 Mid-Atlantic, and others across the US; 6 public, 9 private).
- Validation Procedures: Five specific surgical cases were used for de novo generation:
- Deep Brain Stimulation (DBS)
- Endoscopic lumbar discectomy (PELD)
- Laparoscopic appendectomy
- Coronary artery bypass grafting (CABG)
- Mohs micrographic surgery
Key Insights and Ideas
- The "Sweet Spot" of Complexity: The study found that GPT-4 could reduce "word rarity" (replacing jargon like "respiratory failure" with "losing the ability to breathe") and passive voice frequency, which are major hurdles for readability, without losing the "medical soul" of the document.
- AI as a "Floor," Not a "Ceiling": The researchers don't suggest AI should replace the process, but rather act as a high-quality draft generator. The "AI-human expert collaboration" is presented as a necessary framework to mitigate the risks of LLM hallucinations in high-stakes environments.
- Specificity Matters: Generic forms are often used because creating specific ones for every procedure is labor-intensive. This paper suggests AI can solve the "scalability" problem of personalized medicine by generating high-quality, procedure-specific documents in seconds.
Key Results
- Readability Leap: For institutional forms, the median Flesch-Kincaid Reading Level dropped from 13.9 (college level) to 8.9 (8th grade) ($P = 0.004$).
- Efficiency: Average reading time for patients was reduced from 3.26 minutes to 2.42 minutes.
- Superior De Novo Performance: The forms generated from scratch by GPT-4 were even more accessible, averaging a 6th-grade reading level and receiving perfect 20/20 scores on the standardized consent rubric.
- Expert Approval: In all cases, surgeons and legal counsel found the AI-generated or simplified text to be clinically accurate and legally sufficient, requiring zero wording changes after the initial AI output.
Limitations and Future Work
The authors acknowledge several important caveats:
- The "Reading" Problem: Improving the form doesn't guarantee a patient will actually read it. Research shows many patients sign without reading, regardless of complexity.
- Verbal vs. Written: The study focuses only on the written document; the verbal "consent discussion" remains a separate, critical component that AI does not yet address here.
- Small Sample: While the 15 institutions represent a range of regions, they are all large academic centers. Results might vary with forms from smaller community hospitals.
- Future Directions: The authors suggest this framework could be expanded to other patient-facing materials, such as discharge instructions, clinical trial enrollments, and real-time communication within Electronic Health Records (EHRs).
Summary updated