Highlights:
- Generative AI is a subset of deep learning that creates content based on different data and prompts.
- Although generative AI has the potential to benefit obstetrician-gynecologists, improvements are needed regarding accuracy, incompleteness, and bias.
SAN FRANCISCO — AI has evolved rapidly over the past few years, and with it, generative AI, which creates potentially useful content for healthcare based on short prompts and rich data, has emerged — but it also comes with some trade-offs.
“Unless you've been living in seclusion for the past year and a half, you know that generative AI has been in the news a lot.” Nicole Younglin, MD, MBA, FACOG, “This is a very exciting wave of new technology with potential benefits, but also trade-offs,” the women's health clinical director at Google and clinical assistant professor at the Stanford University School of Medicine said in a presentation at ACOG's Annual Clinical and Scientific Meeting.
Traditional and generative AI models are widely used among consumers, and many patients are presenting AI-driven medical diagnoses and information to their healthcare providers. Therefore, while the benefits of AI in healthcare are many, there are still some caveats to be aware of, and it is important for healthcare providers to understand and learn AI tools.
What is generative AI?
AI is the concept of machines performing tasks that typically require human intelligence. With AI, machine learning algorithms learn from patterns and relationships in data without being explicitly programmed, according to Young-Lin.
There are three main types of machine learning: supervised learning, unsupervised learning, and reinforcement learning. According to Young-Lin, supervised learning is where algorithms learn from labeled training data, usually from humans, unsupervised learning is where algorithms learn from unlabeled training data and discover relationships based on the data alone, and reinforcement learning is where algorithms learn to make decisions by interacting with the environment and rewards.
Additionally, a subset of machine learning uses artificial neural networks called deep learning, which can be used in obstetrics and gynecology to read mammograms, for example, Young-Lin said. Currently, companies and various academic centers use AI to help radiologists diagnose cancers and abnormal lesions before they become cancer or are misdiagnosed. Generative AI is a subset of deep learning that can generate audio, video and other data.
“When we think about generative AI, we're taking in a lot of data and using it to train what's called foundational models. These are foundational models that you've heard about in the news, and they stand for Generative Pre-Trained Transformers, or GPT,” Young-Lin says. “In that training process, especially in larger companies with lots of engineers with expertise, you can do something called human feedback augmentation, where a human evaluates some of the responses and retrains that model at a foundational level.”
Once these basic models are in place, they can be tweaked and adapted to the tasks required, according to Young-Lin.
“This is powerful because it removes the need to explicitly train a model for just one task,” Young-Lin says. “You can use the same model to perform a variety of tasks. That's the power of generative AI.”
Currently, generative AI models can provide more than 90% accuracy on the U.S. medical licensing exam, compared with just 50% accuracy a year and a half ago, Younglin said.
“There's a common saying in Silicon Valley and the technology world: 'If you move fast, you break things,' but in healthcare, of course you want to move fast, but you definitely don't want to break anything,” Younglin says. “We're seeing more technology adoption in our field, and we still have a lot to learn.”
Generative AI in Healthcare
Currently, companies are working on applying generative AI in biomedical and natural language processing for protein and drug design for digital twins, biosimulation, and synthetic patient data, Younglin said. Generative AI is also being used in the clinical world for radiology image enhancement, electronic health records, coding, and more.
Generative AI has many positive aspects in healthcare, including the potential to simplify administrative tasks such as scheduling, referral processes, coding, and billing. Additionally, work is underway to incorporate generative AI into clinical documentation, inbox management, clinical diagnostics, treatment, decision support, and error and anomaly detection for healthcare providers. For patients, generative AI has the potential to improve patient education, care summaries, language translation, personalized care and experiences, and affordable access to healthcare.
In medical education, generative AI can assist both learners and educators. For learners, generative AI can improve learning efficiency, personalize learning and feedback, and engage with dynamic content while creating lesson plans and materials and providing real-time feedback, and for educators, it can improve student assessments. Generative AI can also be useful in research, such as drug development, predictive modeling, connecting patients with clinical trials, conducting literature reviews, and drafting abstracts and papers, Young-Lin said.
Finally, in the field of public health, generative AI can help analyze data for population health management, improve health workforce skills, develop public health campaigns, and optimize resource allocation.
“Obviously, it's not all smooth sailing,” Younglin says. “We live in a world of AI, not just generative AI, so there's a lot to think about, especially as a provider.”
Important aspects to consider include privacy and security, ethical and legal issues, and over-reliance. In healthcare, sensitive data requires security and privacy, but currently there is a lack of transparency for users and patients about what their data is being used for. Young-Lin also mentioned ethical and legal concerns with generative AI, particularly around intellectual property, copyright, regulation, liability and ownership. There are concerns that over-reliance could lead to generative AI leading to de-skilling, dehumanization and circumvention of the learning process.
Moreover, generative AI models have downsides: Young-Lin said they may fabricate information to provide a response, provide incomplete responses, or omit important information, which could be dangerous when dealing with patients and their care.
“We also know that bias and discrimination are rampant in the medical world,” Younglin says. “These models are trained and encoded with existing data that is inherently biased. And these models perpetuate and amplify these biases in their responses.”
In one study published in January, researchers evaluated the use of a GPT-4 model in the same 17 clinical scenarios and observed changes in recommendation and referral patterns when race changed.
“We have a unique opportunity and responsibility to be at the forefront of this change. To really have a voice, we need to understand it, and the first step is to try it,” Younglin said. “Only once we understand it can we actually guide development and ensure that these tools and technologies serve the best interests of patients and women's health.”
