GenAI Use Cases and Challenges in Healthcare with Dr. Girish Nadkarni

Dr. Girish Nadkarni discusses the safe and ethical implementation of AI in healthcare, emphasizing the importance of assurance testing for large language models (LLMs) to align with clinical practices. Generative AI is increasingly automating back-office tasks, which make up 85% of operations, such as scheduling and billing, while clinical applications are emerging. The impact of AI on physician-patient interactions is significant, as ambient AI can generate billable notes from conversations, allowing physicians to focus more on patient engagement.

AI plays a critical role in risk prediction and sepsis detection, with predictive algorithms being vital in life-and-death scenarios. Nadkarni highlights the need for careful deployment of AI tools, as inaccurate algorithms can have serious consequences. Assurance, monitoring, and privacy are essential, with randomized testing recommended before AI integration in clinical care. At Mount Sinai, the focus is on safety and ethics in deploying predictive AI algorithms.

The synergy between classical machine learning and generative AI is explored, particularly in data-limited scenarios. Generative AI can enhance communication among healthcare professionals by providing reasoning for its predictions, although concerns about hallucinations and accuracy persist. Hallucinations are viewed as a creative feature rather than a flaw, and while guardrails can reduce their occurrence, they cannot be entirely eliminated. A risk-based approach is necessary to determine which decisions can be automated versus those requiring oversight.

AI governance in healthcare is evolving, requiring diverse stakeholder perspectives. Effective governance should differentiate between back-office and patient-facing applications, with stricter requirements for clinical impacts. Key considerations include safety, effectiveness, and ethics, advocating for clarity of purpose and established oversight protocols.

The complexities of decision-making in healthcare are emphasized, with some decisions being irreversible. Tailored training for healthcare professionals and continuous feedback mechanisms are essential for adapting to the evolving landscape. AI's role in drug discovery and precision medicine is highlighted, with advancements aiding in the development of new antibiotics and personalized treatments. The potential for personalized medications tailored to individual genetic profiles is discussed, alongside the concept of a "Coke machine" for drugs.

The current healthcare system is critiqued for being reactive, with a call for personalized prevention methods. Promising use cases for GenAI in the next 12-18 months include clinical summarization and diagnostic imaging advancements. AI is seen as a transformative force for global healthcare access, particularly in underserved areas, although challenges remain regarding patient data input and regulatory frameworks.

The use of virtual agents in healthcare is also discussed, with an emphasis on ethical implementation, especially for vulnerable populations. Speculation exists about AI-driven robots potentially outperforming human surgeons, though the timeline for this capability is uncertain. Bottlenecks in scaling GenAI solutions include reliability, accuracy, and governance, with the integration of AI into existing workflows being crucial.

Transformational technologies like AI require time for societal adaptation, as illustrated by historical transitions. AI is currently enhancing efficiencies in healthcare by assessing readmission risks and identifying potential patient issues. While false positives may not always have severe consequences, quick insights can lead to better resource management. The complete integration of AI into complex medical procedures remains a future goal, with personalized medicine representing a promising advancement in transforming patient care. A key takeaway for healthcare leaders is to prioritize patient-centered approaches to increase the chances of successful technology implementation.