Dr Joseph Chervenak Discusses How Large Language Models Are Revolutionizing Reproductive Care

Personalized information tailored to recommended reading levels through large language models can address socioeconomic barriers when it comes to receiving reproductive care and improving patient outcomes, says Joseph Chervenak, MD, MBA, obstetrician and gynecologist at Montefiore Institute for Reproductive Medicine and Health.

Transcript

You are presenting on the topic of “Patient Centric & Access to Care.” Can you tell us more about what this topic is about and the importance of large language models?

Access to care, which I believe is the main category our work is submitted under, is about one of the central challenges in the field of infertility care. One of the aspects of the field that drove me to pursue fellowship training is the relentless focus on innovation and the dramatic improvement in what can be offered to patients in a short period of time.

This synergistic combination of improved success rates of assisted reproductive technologies, or ART, with improving access to services and general knowledge of infertility has driven a lot of the growth in the field that I think people are excited about.

However, difficult-to-address economic, social, and structural barriers to care have meant that the growth and thus the benefits of fertility-related care are then unevenly distributed. Attempts to address these barriers through strategies such as state insurance mandates have been met with mixed results.

So, while the conference is going to be a celebration of continuing cutting-edge research and some exciting developments, this topic remains paramount. We believe the development of large language models such as ChatGPT by OpenAI have the potential to make a real difference in fertility care—a field that is uniquely influenced by patient counseling.

So much of what we do as REIs [reproductive endocrinologists] is counsel on how human reproductive biology works. Patients for years have used the internet to get quick answers to their questions, and as a result, clinical content online has proliferated. However, to date, the only way to ask medically related questions and receive direct, reliable, and comprehensive answers has been to consult a health care professional, who is often the provider who digests this material and makes it comprehensible by patients. That’s what has the potential to change now with advancements in large language models and what we sought to investigate with our study.

What are large language models and how are practices such as Montefiore leveraging these models in reproductive medicine?

Large language models, put simply, are deep learning artificial intelligence algorithms that can perform language tasks. Tasks include answering questions and the ability in general to carry out language-based commands. An example would be to write a story in the style of Shakespeare, or write an article in the scientific style. Perhaps the most popular of these models, and I think what most people have heard about, is ChatGPT by OpenAI. This model is a chatbot. It’s accessible through a publicly available website. And it may be thought of as a more capable version of the software that we’ve been using on customer service websites for years that enable us as humans to have conversations with a computer interface. What’s changed, though, is the power of the model beneath that chatbot. The model that powers ChatGPT allows it to answer questions and follow commands to produce novel text with exceptional range and detail.

One of the initial language models that ChatGPT was based on, generative pre-trained transformer 3 or GPT-3, had a truly immense dataset that included 57 billion words and 175 billion parameters. So for reference, the average human will only read, speak, or write a billion words in our lifetime. So GPT-3, which was sort of one of the initial models, is already 57 times that. So, this really allowed for the major leap from power models and allowing the performance that a lot of people are excited about.

I’m proud to say that Montefiore has been at the leading edge of this research. In the September issue of Fertility and Sterility, the group involved with the study we will present at the conference—which include myself and my mentors, Dr Jindal [Sangita K. Jindal, PhD, professor, Montefiore Medical Center Institute of Reproductive Medicine & Health] and Dr Lieman [Harry J. Lieman, MD, professor, director of reproductive endocrinology and infertility division, Montefiore Medical Center Institute of Reproductive Medicine & Health]—we published the first paper to evaluate ChatGPT as a clinical knowledge tool in the field of fertility. Our paper demonstrated that the model proved to be pretty good at various fertility knowledge tasks, like responding to questions, validated surveys, and we thought this was pretty remarkable for a generally trained domain-agnostic model. So, we believe our work demonstrates the potential for many clinical applications of that technology and one of the most powerful would be to use these chatbots for low-cost, high-yield improvements in clinical outcomes on a population level.

Our field is very attracted to implement this technology. Optimal outcomes are critically dependent on patient knowledge of their own biology. For example, a female patient desirous of a child who presents with decreased ovarian reserve or egg count, needing potentially expensive treatments such as IVF [in vitro fertilization], may have averted this outcome by better understanding the relation between age and fertility. For many patients, there are several several evidence-based strategies that they can pursue on their own, they just need the knowledge and awareness to pursue them.