Eric Topol, MD answers his own startling question, contrasting medical imaging with decision support for both clinicians and patients. His recent Substack ‘Ground Truths’ article (link below, free access) will make you think harder about what is being sold as ‘medical AI’ and what has actually been validated through multiple studies. 

Imaging AI is the Undiscovered–but Mapped Out–Country. Deep Learning (DL)-based AI models developed using medical imaging have substantial validation over more than a decade, and they are accelerating. There have been multiple validated studies using information from retinal scans as predictors of future medical conditions such as Parkinson’s, heart disease, stroke, and Alzheimer’s Disease. The retina is apparently a diagnostic gateway to nearly every organ; many studies have focused on it as scans are fairly routine. Other AI-assisted models have used deep learning to detect multiple health conditions: thymus, cardiovascular conditions, through mammography, colonoscopy, and importantly, detecting pancreatic and other cancers from computed tomography (CT) images done for other reasons. “Opportunistic AI” alone is being used in detection for a long list of health conditions. Dr. Topol’s point is that none of these new diagnostic methods have made it into standard practice, despite being used in other countries like China (PANDA) and with at least four companies developing uses for retinal AI to detect specific diseases.

Medical LLMs and Generative AI, on the other hand, are building what may be Castles In The Air.  Seemingly everyone is developing, funding, and selling a LLM-based chatbot, LLM-aided diagnosis, management, patient triage, and direct patient use. Unfortunately, they’re being sold without real, continuous evidence through rigorous studies over time. What studies there are, are generally simulations, small-scale studies, or individual case studies which need further real-world validation. The clinical trials, the infrastructure, and the monitoring for safety, effectiveness, and cost are simply not there yet, and it’s past time. (Raj Manrai quoted in Science). In addition, generative AI keeps changing making studies harder to track results over time. Dr. Topol’s conclusion: “In summary, there is very little evidence for LLMs benefiting patients or doctors for health outcomes.”

That is not to say, as Dr. Topol does, that AI won’t grow in usefulness in areas such as medical research and chart summaries, discharge instructions, translations, administrative work such as documentation of billing codes, clinical workflow, and insurance authorization. AI has already worked its way into RCM where no respectable company does not have an AI-enabled tool. The American Medical Association (AMA) study he cites indicates both current use and growing acceptance by physicians. (To this Editor, it resembles the telehealth usage graphs of a decade ago, and she expects the same progress.) 

He calls it a paradox between imaging AI and LLMs. This Editor calls it a shame that healthcare technology and investment keep chasing what’s easy, ‘sexy’, and can generate fast revenue/ROI. Not what is more difficult but proven, and that can have a potential huge impact on health outcomes.

Dr. Topol’s closing is fitting:

Let’s fix this paradox of medical AI implementation. It’s a two-fold and major undertaking. Amping up the use of medical AI where it’s proven and performing the clinical trials required to justify wide-scale adoption where pivotal evidence is lacking.