“Take two aspirin and call me in the morning” is the punch line to decades’ worth of doctor-diagnosis jokes, but San Francisco software company Enlitic doesn’t see the humor in it. Enlitic is building a computer system to help doctors make faster, more accurate diagnoses, and it’s gotten its first major customer in the Australian radiology provider Capitol Health.
Capitol is also leading Enlitic’s $10 million Series B round of funding. The medical imaging firm, with 51 clinics, will begin feeding its archives—MRIs, X-rays, ultrasounds, CT scans, and the like—into Enlitic’s system, which will be learning on the fly how to spot tumors, nearly undetectable fractures, and other medical conditions. In 12 to 18 months Capitol could begin rolling out the system in a few of its clinics to help its radiologists make various diagnoses, according to Enlitic CEO Jeremy Howard (pictured).
Enlitic is using what’s known as “deep learning,” or a type of programming meant to mimic how the brain sorts out and learns from incredibly complicated layers of input, like the varied objects in busy photos or the intricacies of individual voices and speech patterns.
Deep learning networks are behind facial recognition on Facebook, the navigation for Google’s self-driving cars, and perhaps even this article—a part of which I wrote using instant speech-to-text conversion in my browser. Health centers such as MD Anderson Cancer Center in Houston and the Cleveland Clinic are using IBM Watson’s artificial intelligence system to help with medical decisions. (IBM aims to add medical imaging to Watson’s brain with the recent purchase of Merge Healthcare.)
Can deep learning help make better medical diagnoses? Howard cites two recent studies to bolster Enlitic’s case.
First, he says the Enlitic system did a 50 percent better job detecting malignant lung nodules than thoracic radiology experts. After training on images from 6,000 people—1,000 people with lung cancer, 5,000 people without—Enlitic used images from about 100 patients to test its system versus the human experts.
Second, Enlitic says its software fared much better than radiologists at finding hard-to-detect fractures in the wrist. It used an old study of about 100,000 images as its baseline. In 200 X-rays where the Enlitic system’s diagnosis diverged from the original report, two radiologists did a close inspection and found that the algorithm was correct 75 percent of the time, said Howard. “This shows we were more accurate overall than radiologists, although a combination of radiologist plus algorithm would be best of all, which is what we provide,” he said.
Those results were retrospective and have not been published in peer-reviewed journals, although Howard said the company has “a strong desire to do so.”
How Enlitic’s system saves time, money, and provides better health care in real world situations has not yet been proven, leaving the company’s lofty rhetoric rather unfulfilled. Howard said the system will be able to help doctors diagnose problems with “every affliction in every part of the body,” and the VP of radiology Rodney Sappington said in a press release, “We have yet to see this level of improvement in radiology since Roentgen’s application of X-rays to medicine.”
When pressed on the difference between good results with controlled data sets and helping new patients get healthier faster, CEO Howard said there would be “a number of steps to get through this transformation process” with the Capitol Health leadership and that Enlitic would “continue to work on more clinical validation retrospectively. That counts for a lot.”
And the radiologists on the ground? “They have to get comfortable with it,” Howard said. “They have to learn to trust it. Initially we expect to see a great increase in accuracy, then an increase in efficiency over months as they learn to trust the system more and more.”
The partners hope to publish health outcomes data once the system is up and running.
Howard says Enlitic’s system won’t be limited to a few indications. A new lung scan that enters the system will certainly be analyzed for suspicious nodules, but the program should also flag, for example, previously undetected aneurysms—sections of a blood vessel wall that balloon out and threaten to rupture.
Previously a management consultant, advisor to Khosla Ventures, and president of a data-science community and competition platform called Kaggle, Howard has no background in medicine, which makes him “entirely well qualified to start a new medical company,” as he told a TED audience, tongue somewhat in cheek, during a talk in Brussels last year.
He is also Australian—a coincidence, he says. A mutual friend introduced him and the Capitol managing director John Conidi some time ago as two people with “crazy ideas about what computers can do in medicine.”
