One day in 2017, computational biologist Aristotelis Tsirigos, PhD, strolled into the kitchen near his office in NYU Langone Health’s molecular pathology lab to grab a cup of coffee, and met a new faculty member. When Dr. Tsirigos introduced himself as director of clinical informatics, his new colleague, Narges Razavian, PhD, perked up. She was also a computer scientist, albeit one with a slightly different focus.
While Dr. Tsirigos was applying genomic data to uncover some of the complex cellular processes underlying disease, Dr. Razavian, assistant professor of population health and of radiology, was experimenting with ways to use artificial intelligence, or AI, in the clinic as a diagnostic tool.
But like most researchers at the intersection of AI and medicine, her big challenge was finding enough patient data to properly train her software to recognize the telltale markers of pathology. As it happened, pulling together large databases of patient data was one of Dr. Tsirigos’s specialties—and so a fruitful partnership was born.
The result of that serendipitous encounter is an experimental AI program that can automatically diagnose two of the most common types of lung cancer with 97 percent accuracy, a rate that bests human pathologists.
Even more remarkable, the system has taught itself to identify specific genetic mutations associated with lung cancer by analyzing pathology images alone—a process that ordinarily relies on expensive genetic tests and can take weeks. “Delaying the start of cancer treatment is never good,” says Dr. Tsirigos, associate professor of pathology at NYU Langone’s Laura and Isaac Perlmutter Cancer Center. “Using AI, we will soon be able to instantly determine cancer subtypes and possibly the mutational profile to start targeted therapies sooner.”
Last September, Dr. Tsirigos and Dr. Razavian published their work online in Nature Medicine. Their paper, along with hundreds of others published in recent years chronicling the rapid ascent of AI in medicine, signals a looming paradigm shift in which intelligent machines are poised to make healthcare safer and more efﬁcient.
It’s a prospect that could dramatically reduce the estimated 10 million medical errors that occur every year in the United States, all while mitigating harmful delays and unnecessary testing. “The partnership between human clinician and machine can potentially be remarkably richer than either one alone,” says Daniel K. Sodickson, MD, PhD, vice chair for research in the Department of Radiology and director of the Center for Advanced Imaging Innovation and Research.
AI is already helping gastroenterologists to spot hidden polyps during colonoscopies, ophthalmologists to diagnose retinal pathologies from eye scans, and psychiatrists to identify post-traumatic stress disorder from voice recordings. Some experimental AI systems are scanning electronic medical records and flagging undiagnosed conditions like heart failure and diabetes. Indeed, machine-learning algorithms are now 150 times as fast as humans in finding signs of stroke and other neurological events in CT scans.
What’s more, the machines will only get faster as developers gain access to more clinical data. The so-called “neural networks” that power AI learn to make decisions autonomously by searching for patterns in vast pools of data—the more data, the better. Indeed, last February, the federal government launched a program called the American AI Initiative to facilitate data sharing among federal agencies and universities. The hope is to spur the development of machine-learning software in the United States to keep pace with AI advances in China, where the absence of privacy laws gives the government unhindered access to hundreds of millions of medical records.
Among the many reasons AI projects are booming at NYU Langone is that researchers benefit from the medical center’s trove of medical records stored in Epic, its electronic health records database. Stripped of any information that might reveal patients’ identities, these records, along with medical scans, pathology reports, and even voice recordings, are now powering numerous projects intended to leverage neural networks for clinical good.
Some look to be especially useful for diagnostic screening, for prompting a closer look at cases when a system’s decision is at odds with a clinician’s, and for flagging risks and problems that sometimes fly under the radar. With computers backing up clinicians rather than replacing them, patients are likely to end up with better diagnoses and treatment decisions, potentially at lower cost.
“The system can support what we do or help us pay more attention to something we may have overlooked,” says Leora Horwitz, MD, associate professor of population health and of medicine, and director of the Center for Healthcare Innovation and Delivery Science, who is helping to develop several machine-learning systems at NYU Langone. “But we still have to rely on our experience and what we see in the patient.”
Read More About Artificial Intelligence
See the different ways NYU Langone experts are harnessing artificial intelligence to transform medicine and advance clinical health and research.
- A Faster, Clearer Way to Program Cochlear Implants
- Using Artificial Intelligence as a Second Set of Eyes to Scope Out Precancerous Lesions
- Giving Pathologists Supervision to Diagnose Lung Cancer Faster & More Precisely
- Tapping Facebook’s Artificial Intelligence Expertise to Make MRIs Speedier & More Comfortable
- Can a Computer Have a ‘Eureka!’ Moment?
- Hearing the Signs of Trauma in a Patient’s Voice
- The Muscle Behind Artificial Intelligence’s Brainy Insights
- Spotting Hospital Patients Who Require an Extra Layer of Vigilance
- What Happens to the Brain After a Concussion? AI May Be the Only Way to Find the Answer