Healthcare is increasingly incorporating generative AI, driven by both Big Tech firms and startups. Google Cloud is collaborating with Highmark Health on generative AI tools for personalizing patient intake. Amazon’s AWS division is working on using generative AI to analyze medical databases for “social determinants of health.” Microsoft Azure is helping build a generative AI system for Providence to triage messages from patients.
Prominent generative AI startups in healthcare include Ambience Healthcare, Nabla, and Abridge. These startups have collectively raised tens of millions of dollars in venture capital, indicating the broad enthusiasm for generative AI in healthcare.
However, there is mixed opinion on whether generative AI is ready for prime time in healthcare. In a recent Deloitte survey, only 53% of U.S. consumers believed that generative AI could improve healthcare. Concerns about its limitations and efficacy have been raised by professionals like Andrew Borkowski, Chief AI Officer at the VA Sunshine Healthcare Network. He warns that generative AI’s finite knowledge base and lack of human expertise make it unsuitable for comprehensive medical advice.
Studies support these concerns. OpenAI’s generative AI chatbot, ChatGPT, was found to make errors diagnosing pediatric diseases 83% of the time. Physicians testing OpenAI’s GPT-4 as a diagnostic assistant also found that it ranked the wrong diagnosis as its top answer nearly two-thirds of the time. On the MedAlign benchmark, GPT-4 failed in 35% of cases for tasks like summarizing patient health records.
Generative AI can also perpetuate stereotypes, as shown in a Stanford Medicine study. The study found that ChatGPT frequently provided incorrect answers and reinforced untrue beliefs about biological differences between Black and white people, leading to potential misdiagnoses.
However, some experts argue that generative AI is improving. In a Microsoft study, researchers achieved 90.2% accuracy on medical benchmarks using GPT-4. Through prompt engineering, they were able to boost the model’s score by up to 16.2 percentage points.
Generative AI also has potential in medical imaging. Systems like CoDoC have shown promise in determining when specialists should rely on AI for diagnoses. Panda, an AI model, has demonstrated high accuracy in classifying potential pancreatic lesions in X-rays.
While generative AI shows promise in specific areas of medicine, there are technical and compliance roadblocks to overcome. Privacy and security concerns surround the use of generative AI in healthcare, and the regulatory and legal landscape is still evolving. Experts emphasize the need for rigorous science, clinical trials, and proper governance to ensure patient safety and address any potential harms.
The World Health Organization has released guidelines advocating for science, human oversight, auditing, transparency, and impact assessments in the development of generative AI for healthcare. These guidelines aim to encourage diverse participation and address concerns throughout the process.
Until these concerns are adequately addressed and safeguards put in place, the widespread implementation of medical generative AI may be potentially harmful to patients and the healthcare industry as a whole.