The deployment of AI in the medical sector has long been a topic of intense debate. Across the globe, people are being warned off attempting to self-diagnose using generalized chatbots like ChatGPT and Google Gemini. These tools now come with explicit warnings that the medical information they provide could be inaccurate. Yet we live in a world where AI has been promised as the means for triggering a revolution in medical care. One of the first barriers to achieving that seems to be: how can we trust it?
Diagnosis has forever proven to be the most complex step in the medical process. Doctors, especially general practitioners, often need to interpret a limited set of symptoms and imperfect patient testimony to determine a diagnosis. It doesn’t take a medical expert to recognize that symptoms such as headaches, nausea, and fatigue – three common ones – could indicate hundreds of ailments. Yet in most cases, this is all doctors have to go on when attempting to diagnose, at least in the first instance.
AI has long been heralded as the tool that could swoop in and revolutionize this. By providing an extra brain of sorts that can immediately draw on infinitely more information than an individual person ever could, diagnosis can become quicker and more accurate. At least that is the theory. The reality has proven far messier, with generalized LLM tools regularly offering misdiagnosis, even when used by trained medical professionals. The result is a deepening mistrust of AI’s use in medicine.
How can this issue be resolved? First, it is important to remember why it must be resolved. According to the latest statistics, last year alone approximately 371,000 Americans died because of medical misdiagnosis. This doesn’t account for those who suffered life-altering injuries and disabilities. Tackling misdiagnosis is therefore, arguably, modern medicine’s biggest challenge.
It would be far too simplistic and inaccurate to simply blame medical professionals. It is not reasonable to expect them to immediately know the answer to a range of complex, unclear patient conditions. And time is, of course, always of the essence. What they need are tools to support them. AI tools that can rapidly pull together relevant information and quickly focus attention towards the correct answer. Generalized tools are too susceptible to misinformation to perform this role and have already lost the trust of millions through misuse.
The answer lies in specialist, purpose-built and trained models that depend only on evidence-based medical literature and studies. OpenEvidence, founded by tech entrepreneur Daniel Nadler, is one such example. The tool, which the company claims is already used by approximately 45% of physicians in the US, is not trained on the mass of information available on the World Wide Web.
Instead, it is specifically trained only on high-quality, peer-reviewed medical and scientific literature and databases, ensuring far higher quality information. The platform has already struck partnerships with the likes of Nature, the New England Journal of Medicine, and the National Comprehensive Cancer Network to train the model on their data. It therefore combines the speed of mass market LLMs with the accuracy of a medical university library.
The platform has been explicitly backed and supported by some of the world’s leading medical institutions, including the Mayo Clinic, the American College of Cardiology, the American Diabetes Association and the British Society for Hematology, amongst several others. This provides OpenEvidence with its most crucial ingredient for success: credibility. It tells users that this is a tool that can provide the rapid response benefits of AI whilst avoiding the problems and consequences of inaccuracy.
Daniel Nadler has previously spoken of the tragic misdiagnosis of his grandfather acting as his inspiration for OpenEvidence. The statistics speak for themselves. One of the greatest improvements, in terms of patient outcomes, in modern healthcare would be a dramatic reduction in misdiagnosis. AI undoubtedly holds the key, and OpenEvidence may be the type of tool that can help unlock the problem.
