
CHICAGO (360Dx) – With the in-store release of the iPhone 8 and, iPhone X, and Apple Watch Series 3 set for Friday — and the new Apple iOS 11 already available for download — comes the debut of Core ML, a framework for machine learning built into new Apple mobile devices. One of the first companies to take advantage of that new technology is VisualDx, maker of an image-based clinical decision support system.
Rochester, New York-based VisualDx this week introduced DermExpert, a mobile app aimed at primary care physicians that helps iPhone users analyze images of skin lesions within a second, according to VisualDx CEO Art Papier
VisualDx got a prerelease boost from a CNBC story last month about machine learning in the new Apple products. DermExpert was among the initial batch of Core ML-enabled apps to hit iTunes Tuesday, which was Day 1 of iOS 11 availability.
DermExpert is aimed at any medical issue that presents on the skin, and not necessarily those seen by dermatologists. "When people think of dermatology, they think of warts and acne and skin cancer and psoriasis — common things. But there are many, many systemic diseases and infectious diseases that present on the skin," Papier noted.
"If you have Lyme disease and it starts with a rash and your doctor calls it something else, you are going to end up with arthritis a year or two later," he said. "The skin is often a window to systemic illness, infectious disease, drug reactions — all kinds of different problems."
With DermExpert, medical professionals can take a photo of a skin abnormality and have the app classify the type of lesion nearly instantaneously to help non-dermatologists make an accurate diagnosis. It is a technology that VisualDx has been thinking about for years, but the on-phone machine learning finally made it possible.
"People over the last decade have said to us, 'Art, when can we point the cell phone at the rash or the lesions on the skin and get the software to analyze it?'" Papier commented. He always responded that that was something on the radar for the future.
VisualDx began as a way to help non-dermatologists diagnose skin conditions; the company said that 65 percent of medical complaints about skin issues are seen by medical professionals other than dermatologists. VisualDx has since branched out into clinical decision support for other diagnostic areas, but it maintains an extensive collection of skin images.
A VisualDx software update released in June included translations into Spanish, French, German, and Mandarin. "There's a lack of specialists in many foreign countries," Papier noted. "We have a lot of specialists here in America, but many countries are relying on their GPs" for skin conditions. "We need to augment the brains of the general practitioners. There's just too much for one individual to know."
Within the last two years, the company started working on training machine-learning algorithms on its image bank. "We made a lot of progress last summer [2016]. We had the ability to point the camera at a rash and it would classify the type of lesion," Papier said.
"It would say the type of lesion, not the diagnosis, but the type of lesion. Is it an ulcer? Is it a vesicle? Is it a macule? That is key to getting a correct differential diagnosis, accurately describing what you see on the skin," he explained.
During initial testing, though, photos would be sent to a VisualDx cloud server for analysis. That was problematic. "We felt our customers would not want [patient-specific] images being sent up to the cloud to be analyzed at a third party because of HIPAA," said Papier, a dermatologist and medical informaticist.
At the 2017 Apple Worldwide Developers Conference in June, though, the iPhone maker announced the forthcoming Core ML. "You could run these machine-learning models on the phone itself without having to send the image. We became really excited," Papier recalled.
This allowed VisualDx to embed its models in the locally stored app. "What that means is that you have a doctor taking a picture of a patient on their smartphone and the image doesn't go anywhere," Papier said. "We [at VisualDx] never see the image. The image gets analyzed on the phone, and that's it."
Users then can take the image classification and run it against the VisualDx knowledge base. "The user verifies the analysis because the analysis is not going to be 100 percent correct, so there's actually a training effect where the user is learning the classifications of lesions as they use this tool. Then it blends in and merges with the diagnostic support that VisualDx has always addressed," Papier said. "It really ensures patient confidentiality."
The company has not made any decisions on whether to expand this new feature to Android smartphones. "Apple is out in front with these new tools to put machine learning on the phone," Papier said. Unlike iOS, Android supports devices from dozens of hardware manufacturers, so those phone makers will need to put machine learning in their devices before VisualDx can develop an Android version of DermExpert.
With or without DermExpert, VisualDx interfaces with electronic health records from Cerner and Epic Systems and, according to Papier, is "fully compatible" with the Fast Health Interoperability Resources standard — commonly called FHIR — to encourage interoperability with other EHRs.
Papier is passionate about addressing the issue of diagnostic error, as highlighted by a 2015 report from the US National Academy of Medicine — formerly known as the Institute of Medicine — that said that most Americans will be victim to at least one "meaningful" diagnostic error in their lifetimes, either due to a delayed or inaccurate diagnosis.
Papier is a board member of the Society to Improve Diagnosis in Medicine and he is a disciple and former student of Larry Weed, the recently deceased medical informatics pioneer who had been pushing for clinical decision support since the 1960s.
Don't call VisualDx or DermExpert diagnostic software, though. "We're not going direct to diagnosis. We're just augmenting the physical exam," Papier said.
This classification helps the company avoid the need for clearance from the US Food and Drug Administration, which has said that standalone clinical decision support — technology not tied to a specific medical device — does not require premarket review as a diagnostic device.
"Obviously, we're going to follow the FDA guidance very carefully. So far, there's nothing that said we need to do something specific for augmenting the physical exam," Papier noted.