Google Health unveiled an open-access data set of real-world dermatology images in partnership with Stanford Medicine.
Thousands of people contributed their images of skin, hair and nail conditions to the Skin Condition Image Network (SCIN) for a diverse data set reflecting various skin tones, ages, settings and condition severities. Dermatologists and research teams then helped identify diagnoses on each image and labeled them based on two skin-tone scales.
Today, scientists and doctors can use the data set to help them develop tools to identify dermatological concerns, conduct dermatology-related research and expose medical students to more diverse examples of skin conditions and how they manifest across skin types.
SCIN was highlighted on Tuesday at Google Health’s annual Check Up event in New York City, where executives laid out the latest developments in AI and strategies for addressing health equity.
During her introduction, Karen DeSalvo, M.D., the company’s chief health officer, called the current moment an “inflection point in AI.” The company’s goal is to leverage the technology by building it into the products people already use every day. But AI will never replace doctors, Salvo acknowledged: “We must remember that AI is just a tool and at the end of the day, health is human.”
Responsible AI
Speaking about the need to train algorithms responsibly, Google’s chief health equity officer Ivor Horn, M.D., stressed the importance of including underrepresented communities. “Without diverse, representative data, AI models can do more harm than good,” Horn said during her presentation.
As part of its effort to build out more diverse and accessible health information online, Google rolled out a new feature in Google Lens last year, allowing users to take a picture of their skin and search the web for visually similar matches that come from reputable sources. The AI-powered feature is available in more than 150 countries.
Google has made several changes to its search function, partnering with the World Health Organization to include authoritative information on conditions and “knowledge panels.” It has also added diagrams from what it calls high-quality web sources to help users visually understand their symptoms or conditions. Google is working to make these visual results available on phones and plans to roll out this update globally in the coming months.
Equity is embedded deep in the core of Google’s strategy across all teams, Horn told Fierce Healthcare at the event. “This is the first time in really doing this work where I walked in and people were like ‘let’s do this, let’s go,’” Horn said. She joined the company in 2020.
“My favorite part of my job is when I don't have to talk about equity and someone else is talking about it,” Horn said. Horn’s team has worked across departments, alongside engineers, product teams and cloud partners to build out every product with equity in mind. “It is very much a part of what we do; it is not something that sits on the side,” Horn said.
Various teams, including researchers, clinicians, bioethics and AI researchers, have been working together to develop a framework for building AI that avoids unfair bias dubbed HEAL. It is built on a process that includes quantifying preexisting outcome disparities, assessing the likelihood that AI will perform equitably and measuring its performance with various subpopulations.
The company is also publishing research on identifying biases in large language models (LLMs) and offers a collection of seven adversarial testing data sets as a guide. The tools are based on existing literature on health inequities, actual model failures and participatory input from equity experts. Google is using these measurements to evaluate its own LLMs.
Many of these innovations are still in collaborative research phases, Horn noted, with the goal of “being intentional and being evidence-based in how we move forward.” Google has a team of partners like HCA Healthcare to bring generative AI pilots to hospitals and measure their impact. Google is also part of the Coalition for Health AI, which is building out guidelines for the responsible use of AI.
Recognizing language barriers, Google is leveraging an AI-driven dubbing tool called Aloud to dub videos at no cost to creators on YouTube. Mass General Brigham is among the organizations doing pilots with Aloud to dub first-aid videos from English to Spanish. Those efforts are being expanded to videos on chronic conditions like COPD and cancer.
Google will soon put out a new animation-style course on the Stanford Medicine for Continuing Medical Education YouTube channel, available in Spanish for free. Dubbed using Aloud, it helps health professionals recognize and address implicit bias in clinical practice.