Could AI help diagnose rare diseases?

According to Venkataraman, "[d]abbling in self-diagnosis is becoming more robust now that people can go to chatbots powered by large language models scouring mountains of medical literature to yield answers in plain language..." Researchers are also finding that, if provided the right information, AI chatbots like ChatGPT are often correct about their diagnoses.

For example, one mother whose son had seen 17 doctors for chronic pain recently put his medical information into ChatGPT and received an accurate diagnosis of tethered cord syndrome. This then led a neurosurgeon to confirm an underlying diagnosis of spina bifida that could be alleviated by an operation.

Patients with rare diseases could benefit from AI in particular, Venkataraman writes. "Doctors are very good at dealing with the common things," said Isaac Kohane, chair of the department of biomedical informatics at Harvard Medical School. "But there are literally thousands of diseases that most clinicians will have never seen or even have ever heard of."

Patients with rare diseases often spend years without a diagnosis. Doctors may also miss diagnoses with clear symptoms if they don't recognize the condition. For example, Venkataraman's niece was not diagnosed with Prader-Willi syndrome until she was five years old even though she had classic symptoms at birth, such as floppy muscles and difficulty sucking. None of her doctors recognized the signs of the disease to give her a test to diagnose the disease until years later.

According to NIH's Undiagnosed Diseases Network, as many as 11% of its referred patients each year have illnesses that expert reviewers can diagnose by closely examining lab results and doctors' notes. To better help these patients, Kohane and Matt Might, a computer scientist whose son died from a rare disease, are training a large language model to examine patients' health records and help make diagnoses more quickly.

AI may also help patients with rare diseases connect with others like them. Recently, networks that link medical databases have been helping patients with rare gene mutations reach out to each other, but "one day soon, AI could help link patients with similar conditions even more easily, without knowing the genes causing their illnesses upfront," Venkataraman writes.

For example, a new project at Harvard University called SHEPHERD uses neural networks to assist in the diagnosis of rare diseases by finding similar patients based on existing medical information, identifying candidate genes after genome sequence analysis, and finding patients with the same causal gene or disease.

Precautions are still needed to ensure patient safety and privacy

Although AI has the potential to improve diagnoses in medical care, a problem with relying on the technology "is that we don't yet know how much we can trust it," Venkataraman writes.

Currently, the general public does not have access to the data sets being used to train the major consumer-facing AI models. This makes it difficult to determine whether the medical literature it's being trained on is skewed toward a certain population or includes bias, which can then be further amplified by AI.

To address this issue, Venkataraman recommends governments require commercial chatbots used for medical advice disclose the data sources they were trained on. AI systems that use patient health records should also be subject to government regulations to protect patients' privacy.

Recently, President Joe Biden signed an executive order establishing the first standards for AI in healthcare and other industries.

Under the order, HHS will establish an AI Task Force to develop policies and frameworks on how to responsibly use and deploy AI and AI-enabled technologies in health and human services. HHS, the Department of Veterans Affairs, and the Department of Defense will also develop a framework to help identify and capture clinical errors that occur with AI in healthcare settings. They will also create a central tracking repository to identify associated incidents that cause harm.

Outside of the government, independent researchers should also evaluate the medical advice AI models provide patients. Medical boards and associations should also certify specific AI models so that patients know which ones are credible and trustworthy, Venkataraman writes.

"Companies that want to do right by consumers must make chatbots that are more reliable and transparent," Venakataraman added. So far, these companies have refused to provide details on what data was used to train their models.

"Preventing a dystopian future hinges on discernment in how and when we use AI," Venkataraman writes. But "[w]e'd be wise to remember … that what people need most when they are sick is compassion, curiosity and care — which humans, at their best, still pull off better than machines." (Venkataraman, Washington Post, 11/15)

Generative artificial intelligence offers life science leaders an unprecedented opportunity to revolutionize the drug development process. Explore how life science leaders can leverage generative AI to increase diversity in clinical trials, identify better targets for drug discovery, and reduce drug development timelines — all while promoting equity and improved health outcomes.