.svg){kind=icon}
Over the last year, most large corporations either announced large Generative AI (Gen AI) initiatives or commissioned extensive studies to understand the impact of Gen AI on their businesses. Healthcare is no exception. The internet is awash with articles from big consulting companies about the transformative impact of Gen AI on healthcare and how their consultancies could help healthcare enterprises beat their competition to Gen AI riches.
Most of these articles are overly optimistic or misleading.
The runaway success of a few pioneering Gen AI applications (namely ChatGPT and Stable Diffusion) allowed futurists, consultants, journalists, and even the general public to speculate wildly about what could be done with this new technology. While we look forward to a future where Gen AI-powered robots would diagnose diseases with a Star Trek tricorder and operate on patients without human supervision, we are still stuck in an age where insurance claims are processed by armies of people in windowless cubicle farms, drugs are prescribed as randomly as a roulette bet in Las Vegas, and clinical trials are starving for patients while patients are starving for cures.
We believe Gen AI has a bright future in healthcare. However … Large Language Models (LLMs) do not offer substantive solutions to the grand challenges of healthcare.
Healthcare enterprises that have spent vast sums on Gen AI solutions over the last year do not have much to show to their Boards of Directors. Those who are familiar with the Gartner Hype Cycle realize that the Gen AI “peak of inflated expectations” is behind us and we are riding the “trough of disillusionment” toward a readjustment period.
Leap AI was not fooled by the inflated expectations from pundits and not discouraged by the proclamations of doom either. We understand that the progress of transformative technologies is never linear. It took over a hundred years for electric cars to dominate innovation in the automotive world, but we hope that the Gen AI transformation in healthcare would evolve much faster. Just don’t bet the entire farm on LLMs.
Healthcare is a wild beast. For those who are accustomed to the structured worlds of supply chain management, ERP, finance, or manufacturing, healthcare is too complex an endeavor for the limited inference capabilities of LLMs. Healthcare has a very unique combination of structured knowledge (e.g., insurance reimbursement), vast uncertainty (e.g., differential diagnosis), and exceptions to rules (genomics). Medical LLMs such as Google’s Med-PaLM are trained on medical corpora but are still not a replacement for trained healthcare personnel unless they are confined to very narrow tasks. Furthermore, developing a healthcare application using an LLM is not as simple as plugging ChatGPT behind a user-friendly app. Specialized knowledge is required to add proprietary knowledge layers, to incorporate context, to update the model as medical knowledge evolves, and to prevent the LLM from making up false information.
The biggest limitations of LLMs are not even about the technology itself. LLMs should not be confused with a general intelligence that could reason, learn, and adapt, but they still have remarkable capabilities and will undoubtedly find their way into many solutions and prove their value. The biggest limitations are in current implementations. Since the early 1970s, the holy grail of AI in healthcare has been the AI physician with skills to diagnose and treat patients. Several diagnostic AI systems were developed in those early days but found limited use in the real world due to various issues such as inability to understand context, legal implications, and the difficulty in adapting to changing medical practice. Furthermore, physicians did not appreciate the unwanted incursion into their core competence. Even if an AI physician would be acceptable in the future, it would merely lead to improvements in efficiency and cost. A more reasonable scenario is an AI assistant for healthcare staff to aid in charting, interpretation of test results, literature searches, and other fairly mundane tasks. Improvements in efficiency are important, but they are not true breakthroughs.
Fortunately, Gen AI offers much more than LLMs trained on textual knowledge bases. Gen AI foundation models span a wide spectrum from text to multimedia to highly-structured datasets. At Leap AI, we believe that the future of Gen AI in healthcare is based on specialized medical foundation models that are designed to address specific problems that are too complicated for traditional approaches to solve. These specialized models include:
- Metabolic pathways
- Pathophysiology
- Genomics
- Proteomics
- Pharmacodynamics
- Cancer biology
- Integrative histology, pathology, and imaging
With proper Gen AI approaches combined with extensive expertise in software development and deep scientific knowledge, it becomes possible to tackle many of the great obstacles in healthcare and achieve real breakthroughs in healthcare and medicine.
The future is within reach, and Leap AI is pioneering the next revolution in healthcare.
.svg)