Some experts say the tech isn’t quite ready for the mainstream yet. Both established tech companies and newer startups are pushing for the use of generative AI in healthcare. This type of AI can generate and analyze many forms of media, including photos, text, audio, videos, and more.
In order to create a more customized patient intake experience, Google Cloud and Highmark Health—a nonprofit healthcare organization based in Pittsburgh—are working together on generative AI tools. In order to identify “social determinants of health,” Amazon Web Services (AWS) is reportedly collaborating with anonymous clients to develop generative AI to examine medical information. Providence, a nonprofit healthcare network, is collaborating with Microsoft Azure to develop a generative AI system that can automatically sort patient communications delivered to doctors into priority order.
Among the most notable generative AI firms in the healthcare industry are Ambience Healthcare, that is developing an AI app for doctors; Nabla, which is an AI assistant for doctors; and Abridge, which makes analytics tools for doctors’ notes.
Investments in healthcare focused generative AI initiatives reflect the widespread enthusiasm for generative AI. Health investors agree that generative AI has made a huge impact on their investment strategies, and together, healthcare businesses using AI have secured hundreds of millions of dollars in venture capital. However, both doctors and patients aren’t sure if generative AI for healthcare is ready for the mainstream.
People might not be interested in generative AI. Roughly half of American consumers surveyed by Deloitte recently believed that generative AI may enhance healthcare in some way, whether by reducing wait times for appointments or making it more accessible. Just under half of those surveyed anticipated that generative AI would lead to lower healthcare costs.
The chief AI officer of the largest health system in the United States, run by the Department of Veterans Affairs, Andrew Borkowski, disagrees that the skepticism is unjustified. Borkowski expressed concern that the implementation of generative AI would be rushed because of its “significant” limits and questions about its effectiveness.
He explained that an important problem with generative AI is that it can’t deal with complicated medical questions or crises. “It is not appropriate for offering thorough medical advice or treatment recommendations due to its limited knowledge base, which includes out-of-date clinical information, and its absence of human expertise.”
Those arguments seem to have some basis in research. Some healthcare institutions have piloted OpenAI’s generative AI chatbot, ChatGPT, for limited use cases. A study published in the journal JAMA Pediatrics reported that 83% of the time, ChatGPT made errors diagnosing pediatric disorders. In testing, OpenAI’s GPT-4 as diagnostic assistance, doctors at Beth Israel Deaconess Medical Center in Boston discovered that the model ranked the incorrect diagnosis as the top answer approximately two out of three times.
Medical administrative activities are an integral component of clinicians’ everyday workflows, but today’s generative AI often struggles with them. In 35% of the cases, GPT-4 did not pass the MedAlign test, which measures the effectiveness of generative AI in tasks such as reviewing patient health records and searching across notes.
You shouldn’t take the medical advice given by OpenAI or any of the other generative AI suppliers seriously. However, Borkowski and others argue that they have room to improve. “Relying solely on generative AI for healthcare could lead to misdiagnoses, inappropriate treatments, or even life-threatening situations,” according to Borkowski.
Institute for AI in Medicine researcher Jan Egger of the University of Duisburg-Essen, who oversees AI-guided medicines and investigates how new technologies might improve healthcare, agrees with Borkowski that this is a serious problem. His current thinking is that using generative AI in healthcare requires a doctor to keep a careful check on it.
It is becoming increasingly difficult to be aware of the fact that “the results can be completely wrong,” as Egger put it. Undoubtedly, generative AI has several potential applications; one is the pre-writing of discharge letters. However, doctors must verify it before making the ultimate decision.
A common risk of generative AI is the reinforcement of harmful stereotypes. Generative AI in healthcare has the potential to make grave mistakes, one of which being the reinforcement of damaging preconceptions.
Researchers from Stanford Medicine evaluated ChatGPT and other generative AI-powered chatbots in 2023 by asking them questions on skin thickness, renal function, and lung capacity. The co-authors discovered that ChatGPT consistently gave incorrect responses. What’s more, some of the answers even perpetuated long-held misconceptions about biological distinctions between Blacks and whites, which have been known to cause medical professionals to incorrectly diagnose health issues.
The irony is that the very patients who would benefit most from generative AI in healthcare are also the ones who are most likely to suffer from discrimination against them. Deloitte found that individuals without health insurance, who are disproportionately persons of color (as shown in a KFF study), are more open to using generative AI to solve problems like locating a doctor or getting help with mental health issues. Disparities in treatment could worsen if AI suggestions are biased.
However, according to some scientists, generative AI is making strides in this area. Researchers claimed to have attained 90.2% accuracy on four difficult medical benchmarks using GPT-4, according to a Microsoft paper released in late 2023. This score was not achievable with Vanilla GPT-4. But the researchers claim that they improved the model’s score by 16.2% via prompt engineering, which is to say, making specific outputs for GPT-4 through the design of prompts.
Other Than Chatbots
Generative AI is useful for more than just conversing with chatbots, though. A number of academics have speculated that generative AI might have a significant impact on medical imaging.
A technique known as complementarity-driven deferral to clinical workflow (CoDoC) was introduced in a July article published in Nature by a group of scientists. This system’s main objective is to determine when it is more appropriate for medical imaging specialists to use AI for diagnosis rather than more conventional methods. The co-authors reported that CoDoC outperformed specialists while simultaneously lowering clinical workflows by 66%.
Panda, an artificial intelligence model for X-ray detection of possible pancreatic lesions, was demonstrated in November by a Chinese research team. Although surgical intervention is frequently not possible until these lesions have progressed too far, research demonstrated that Panda was very accurate in classifying them.
An expert in generative AI at the University of Oxford, Arun Thirunavukarasu, said that there is “nothing unique” about it that stops it from being used in healthcare settings.
He stated that In the short- and mid-term, there are more practical uses for generative AI technology. Some examples include text correction, automated note and letter documentation, and enhanced search capabilities for better EHR optimization. “There’s no reason why generative AI technology, if effective, couldn’t be deployed in these sorts of roles immediately.”
Rigorous Science
While generative AI has shown promise in some niches of medicine, specialists like Borkowski warn that there are still many regulatory and technological hurdles to jump before the technology can be relied upon as a reliable, all-encompassing healthcare assistant.
Using generative AI in healthcare raises significant privacy and security concerns,” Borkowski stated. Patients’ privacy and faith in the health care system have been compromised due to the delicate nature of their medical records and the possibility of their abuse or illegal access. Also, there are many unanswered concerns about responsibility, data protection, and the use of non-human entities to practice medicine, as well as the ever-changing legislative and legal landscape that surrounds generative AI in healthcare.
Despite his fervent support for generative AI in healthcare, Thirunavukarasu acknowledges the necessity for “rigorous science” to support patient-facing technologies.
To support the implementation of patient-facing generative AI, he emphasized the need for pragmatic randomized control trials that show clinical benefit, especially without direct medical oversight. “To prevent any unforeseen damage from occurring after a large-scale deployment, it is crucial to have good governance moving forward.”
The World Health Organization (WHO) has just issued new guidelines that call for human and scientific supervision of generative AI in healthcare settings, along with the implementation of new safeguards like audits, transparency, and impact assessments. According to the World Health Organization’s standards, the objective is to give a diverse group of individuals a chance to have their voices heard and offer input as generative AI for healthcare is being developed.
Borkowski warned that “the widespread implementation of medical generative AI may be… potentially harmful to patients and the healthcare industry as a whole” until issues are properly addressed and appropriate safeguards are implemented.
