Artificial and augmented intelligence in dermatology
Where do things currently stand with machine learning — and what does the future hold?
Feature
By Emily Margosian, Assistant Editor, July 1, 2022
“I think the beauty of machine learning is that symbiosis and integral interplay between humans and the machine. We are not obsolete. We’re not going to be replaced by robots,” said Ivy Lee, MD, FAAD, deputy chair of the Academy Augmented Intelligence Committee. “We as dermatologists should be proactive and optimistically critical of machine learning solutions to make sure they are safe, accurate, and fair. Our patients depend on us to make those judgments. That’s why it’s so important for all physicians now to have a basic understanding of machine learning, because it will be increasingly adopted in medicine as it is in other aspects of our lives.”
Currently in a period of rapid development, machine learning has garnered increased attention in the lay press and among physicians over the last five years. Once within the realm of science fiction, the area of augmented (AuI) and artificial intelligence (AI) has since found its way into the daily life of anyone who has made a request of Siri or Alexa. Beyond home virtual assistants, the expanding area of machine learning has posed new questions regarding the future practice of medicine and dermatologic care.
This month, DermWorld speaks with dermatologist machine learning experts to discuss the current landscape of AI and AuI in dermatology, its future potential, and current limitations.
AI vs. AuI
If you’ve ever consulted Google Maps when planning a route, you’ve utilized augmented intelligence: an AI-driven application used as an assistive tool by a human operator. However, despite its ubiquity, the distinction between artificial and augmented intelligence is not always clear.
“Artificial intelligence is basically any system with the ability to gauge and learn that can be used for various human-like functions. Augmented intelligence, on the other hand, carries the potential to assist humans to use artificial intelligence. So as opposed to AI, where it’s just the machine doing the work — augmented intelligence can help a physician work together with a machine to give the best outcome,” explained Carrie Kovarik, MD, FAAD, professor of dermatology at the University of Pennsylvania Perelman School of Medicine.
“Augmented intelligence really emphasizes the role and importance of humans in the loop, with humans being the predominant users of artificial intelligence. Often in the press you will hear that artificial intelligence can be autonomous — making decisions on its own — or assistive, and augmented intelligence refers to the assistive role of AI,” added Dr. Lee. “I think the augmented aspect is the one we really want to promote throughout medicine, because it aims to enhance human intelligence, and improve the physician-patient relationship, rather than replace it.”
Characteristics of high-quality AI and AuI
Most data scientists are familiar with the phrase, ‘garbage in, garbage out.’ According to Dr. Lee, the same concept applies to the creation of high-quality machine learning algorithms, particularly those for use in the highly visual field of dermatology. “It’s important for physicians to have a central role in the development of any data-driven tool or system. As experts in the skin, we know what questions to ask about the accuracy and clinical applicability of these tools. In a lot of cases, accuracy is dependent on how things are labeled. For example, if I’m designing a machine learning algorithm and label an image as eczema — and I diagnose and label the image incorrectly— everything that looks like that image going forward will be mislabeled as eczema. I think that’s where the value of partnership and collaboration with our data scientist and engineering colleagues is important,” she explained.
“It’s important for physicians to have a central role in the development of any data-driven tool or system. As experts in the skin, we know what questions to ask about the accuracy of these tools.”
Dr. Lee also stressed the importance of diverse, representative data in the development of AI and AuI systems. “Not only do we want data that are high quality in terms of crisp images that are properly labeled, but we also want to make sure that these tools are incorporating data that is representative of our patient populations,” she explained. “We know that skin disease can present differently across different skin tones, and most of the tools that we have seen thus far have not been very diverse in terms of input data. Most of the data being utilized features lighter skin tones from limited geographical regions. We must think critically about where that input data is being collected from and how representative it is, because that will determine the fairness and accuracy of these tools when applied in the real world.” (More discussion of this issue appears under the heading “Limitations of machine learning” later in this article.)
Current landscape of AI and AuI in dermatology
While AI and AuI has begun to wind its way into daily life, “I think people have the feeling that there are more usable AI applications right now in dermatology than there really are. In terms of diagnostic capability, we’re not exactly there yet,” said Dr. Kovarik. “However, if we’re talking about non-diagnostic capabilities that can be used in dermatology, I think there’s a lot of potential, and those are likely to be available to physicians closer in the future.”
In the U.S., there are currently no FDA-approved dermatology AI tools, although there are some that have received regulatory approval in Europe. “There are several companies trying to create consumer-facing skin cancer detection applications, and some are already on the market abroad,” said Roxana Daneshjou, MD, PhD, FAAD, clinical scholar in dermatology and postdoctoral scholar in biomedical data science at Stanford School of Medicine. “I try to educate patients that these apps do not have dermatologist-level performance in the real world and that it’s important to see a board-certified dermatologist for their concerns. I am excited for the further development of augmented intelligence or human-in-the-loop tools. Such algorithms, which are being developed, would provide support to non-specialists trying to triage patients.”
According to Dr. Kovarik, the FDA is currently refining its process for evaluating commercially available medical AI devices. “The FDA has been working on a formal way to validate these algorithms and these apps, but as you can imagine, it’s sort of a moving target,” she explained. “Unfortunately, the public can download and use dermatology AI apps that have not been validated or formally reviewed, which can have dangerous consequences.”
This year, the International Skin Imaging Collaboration (ISIC) released the first-ever guidelines for AI algorithms used in dermatology (JAMA Dermatol. 2022;158(1): 90-96). These guidelines, co-authored by Dr. Daneshjou, propose a broad range of recommendations for stakeholders to consider when developing and assessing image-based applications, including image quality, bias assessment, labeling technique, and others.
However, beyond strictly diagnostic applications, many dermatologists are already likely utilizing digital assistant AuI programs in their practices, according to Dr. Lee. “I think where we probably encounter augmented intelligence is through our administrative tasks. For example, it can be used in documentation, with the help of natural language processing technology. You also see it being used for scheduling, and I think that is where we will likely integrate it in our offices first. What garners the most press is AI’s use in medical decision-making — clinical decision support and image analysis — that has not reached our offices yet, and that’s because there are no FDA-approved solutions at this time.”
Future potential
While machine learning has yet to be fully integrated into everyday clinical practice, recent studies have suggested that AI has future potential as a powerful tool in doctors’ diagnostic arsenal. A 2020 study published in Nature found that clinicians paired with AI demonstrated higher diagnostic accuracy in identifying skin cancer than either clinicians or AI alone.
However, in practice, such tools would require a rigorous degree of testing prior to deployment, cautioned Dr. Daneshjou. “If an algorithm can improve the performance of dermatologists, such an algorithm-human combo may be implemented in clinic. However, we would need a prospective, randomized clinical trial to prove that. I think we are still 5-10 years off from having such a diagnostic algorithm,” she explained. “There is a lot of opportunity to use AuI beyond just diagnostic tasks, however. For example, we have developed an algorithm at Stanford called TrueImage, which helps patients take higher-quality photos for telemedicine. Algorithms like this can help improve workflows and save time for patients and physicians.”
Limitations of machine learning
Despite the possibilities offered by machine learning, many proponents are quick to point out that current technology still faces several limitations, particularly within the context of medicine. While AI and AuI may play an assistive role, machine learning has yet to supplant human knowledge, with no FDA-approved algorithms capable of replacing a dermatologist’s expertise.
In terms of performance, questions have been raised regarding both the accuracy and equitable deployment of AI algorithms across all skin tones. “This hints at the issue of biased AI algorithms, particularly those that we’re thinking of using in dermatology for diagnostic and clinical decision support,” said Albert Chiou, MD, MBA, FAAD, clinical associate professor of dermatology at Stanford School of Medicine. “Many of these algorithms, especially those that are being developed commercially, utilize clinical image datasets which are often datasets of convenience. Meaning, developers will work with partners to obtain historical photographs from clinical practices or health care systems. Often, when scrutinizing these larger datasets, we’ve seen a glaring lack of representation of patients of color, particularly Black patients.”
“AI is almost like a mirror that reflects the health care system it’s developed in. There’s concern that when these algorithms are applied to Black patients or patients with darker skin tones, they won’t perform as well.”
While the lack of diversity in dermatology clinical imaging goes beyond AI, failure to utilize representative datasets during the development stage of an application can have far-reaching consequences. “AI is almost like a mirror that reflects the health care system it’s developed in. There’s concern that when these algorithms are applied to Black patients or patients with darker skin tones, they won’t perform as well. You can imagine the stakes are probably highest in cases involving potential skin cancers,” said Dr. Chiou, who serves as a young investigator on a Melanoma Research Alliance (MRA) grant focused on using AI as an early triage tool for identifying melanoma. “A lot of recent work has looked at the use of AI as a decision-support tool for non-dermatologist physicians. If a primary care doctor is trying to decide if a lesion is concerning enough to warrant a referral to see a dermatologist, and a biased AI does not perform well in identifying a potential skin cancer in a Black patient, you can imagine the harm arising from a delayed referral or diagnosis.”
Improving diversity among dermatology clinical image datasets is a multi-step process, involving cross-institutional cooperation and patient buy-in. “Actionable things that we can do now include making greater efforts to share images between institutions, which is actually a pretty complex business, because they oftentimes are not completely de-identifiable, or there’s high risk of identifiable features being present within clinical images,” said Dr. Chiou. “At my institution, we really support efforts like the International Skin Imaging Collaboration (ISIC) to pool images. We know that for rare diseases like melanoma, a single institution alone may not have the volume of images needed of these rare conditions within their patients of color to have the algorithms perform at par with other better-represented groups. Pooling these datasets would allow us to intentionally over-sample for rare diseases to enhance algorithm performance.” There’s a patient engagement piece as well, explained Dr. Chiou. “I think it’s worth explaining to our patients of color about the known lack of representation in dermatology AI research. Our hope is that they might be partners in participating in the types of prospective studies now that are being done to try to obtain more diverse image sets.”
Machine learning and the physician-patient relationship
In an increasingly automated world, some dermatologists have speculated on what impact the introduction of AI into clinical practice will have on the physician-patient relationship. Regarding patient-facing applications, “I think it’s important for us to educate patients on the state of AI in dermatology and provide some digital health literacy, so that if they come across these platforms, they can make more educated choices,” said Dr. Kovarik.
Regarding criticism that the introduction of AI technology may erode the physician-patient relationship, “For most practicing dermatologists, the patient-physician relationship is what brings us the most joy,” said Dr. Lee. “The term ‘augmented intelligence’ represents our hope that this technology augments our human capabilities and capacity to empathize, communicate, and personalize care for our patients. This uniquely human partnership with my patients is the most fulfilling part of my job. If I can apply machine learning in other aspects of delivering care — running an office, coding, scheduling, triaging messages, and charting — I think that would improve my efficiency and gift me more time and energy to be a better physician for my patients.”
