Transforming Surgical Consent: AI and Human Collaboration

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Mar 19 2024

In an article recently published in the journal Npj Digital Medicine, researchers proposed an artificial intelligence (AI)-human expert collaboration approach to effectively address the challenges of readability and specificity of surgical consent forms without sacrificing clinical detail to ensure truly informed consent.

Challenges of administering consent forms

In healthcare, informed consent is a key legal requirement and fundamental ethical principle ensuring that sufficient information is provided to the patients to enable them to make informed decisions regarding their treatment. Consent forms act as both legal documentation that protects both physician and patient and an educational tool.

Thus, a well-designed surgical consent form offers patients comprehensible, concise, and clear information about the alternatives, risks, and benefits of a proposed surgical procedure. However, the specificity and readability of consent forms are the biggest challenges to achieving truly informed consent, as they impede patients' comprehension.

For instance, studies have demonstrated that the reading level of most surgical consent forms exceeds the comprehension of the average patient. Additionally, procedural consent forms are typically written in a format generalizable to all potential procedures, which is another major challenge as such a format does not discuss characteristics that are specific to a procedure, like benefits, steps, and risks, with adequate nuance.

Although surgeons can address the generic consent forms' problem using distinct procedure-specific patient consent or education literature or by offering verbal supplementation, implementing such measures faces multiple barriers, like institutional requirements for maintaining generic consent. Quality improvement solutions that can address these twin challenges require significant investment of resources, human capital, and time.

The proposed AI-human collaborative approach

In this study, researchers introduced a novel AI-human expert collaborative approach/framework to resolve the issues of readability and specificity of consent forms. Newly developed AI systems, specifically large language models (LLMs), have demonstrated their ability to adjust, reformulate, and summarize text, making them highly relevant to this task.

Researchers qualitatively and quantitatively investigated the use of generative pre-trained transformer 4 (GPT-4), a general LLM, to simplify surgical consent forms. Specifically, GPT-4 was utilized to evaluate and transform the consent forms into an accessible reading level in a standardized, effective, and efficient manner.

They also developed an extensible and streamlined legal and medical review framework to ensure that the content of simplified and original consents remains the same. Moreover, the feasibility of using GPT-4 to generate de novo procedure-specific and highly readable consents that can meet expert-level scrutiny was determined. Consent forms obtained from 15 large academic medical centers were evaluated for readability and then simplified using GPT-4.

Nonparametric tests were performed to compare the post-simplification and pre-simplification readability metrics, while a malpractice defense attorney and medical authors conducted independent reviews. Additionally, the potential of GPT-4 to generate de novo procedure-specific consent forms was investigated by evaluating the GPT-4-generated forms using a validated eight-item rubric and expert subspecialty surgeon review.

Significance of the study

The analysis of consent forms from 15 academic medical centers displayed substantial reductions in average reading time, passive sentence frequency, and word rarity following GPT-4-facilitated simplification. The original consent forms contained a median of 651.0 words and 3976.0 characters, requiring 3 minutes and 15 seconds of reading time on average.

However, the median number of characters and words decreased to 2485.0 and 483.0, respectively, and the reading time was reduced to 2.42 minutes after consent processing by the LLM/GPT-4. Additionally, the median characters per word and words per sentence were decreased from 5.4 to 4.5 characters and from 21.6 to 19.0 words, respectively, post-simplification.

Moreover, the percentage of sentences in passive voice decreased substantially from 38.4% to 20.0% after consent processing by the LLM. Readability was improved from an average college freshman level to an 8th-grade level, which is the reading level of an average American.

Before LLM processing, the surgical consent forms displayed a median Flesch-Kincaid reading level and Flesch reading ease score of 13.9 and 35.3, respectively. However, the reading level decreased to 8.9 and the reading ease improved to 63.8 post-simplification. Legal and medical sufficiency was also confirmed.

The GPT-4-generated procedure-specific consent forms for five different surgical procedures achieved an average sixth-grade reading level. They also achieved perfect scores on a standardized consent form rubric and withstood scrutiny during expert subspecialty surgeon review.

To summarize, this study's findings demonstrated the feasibility of the proposed AI-human expert collaboration approach to significantly improve the consent process by providing specific, clear, and comprehensible information to patients, promoting truly informed consent.