Unanimous AI achieves 22% more accurate pneumonia diagnoses | Industry
Breaking Tech Industry news from the top sources
It’s no great mystery that artificial intelligence’s (AI) predictive prowess can significantly improve health care outcomes. AI’s been shown to outperform dermatologists in diagnosing melanoma, and a recent study by Google subsidiary DeepMind found that algorithms were on par with clinicians in detecting eye conditions.
But what if AI could perform even better with the aid of humans? That’s the pitch Unanimous AI, a startup headquartered in San Francisco, has been giving for the better part of four years. There’s merit to it: in a study conducted with the Stanford University School of Medicine, its system diagnosed pneumonia “significantly” more accurately — 22 percent — than a team of radiologists working alone, and reduced errors by 33 percent.
The results are being presented today at the Society for Imaging Informatics in Medicine conference in National Harbor, Maryland.
“Humans have experience, knowledge, wisdom, and tuition that’s not being represented in these big data sets,” Stanford-educated computer scientist and CEO Louis Rosenberg told VentureBeat in a phone interview. “It takes 12 years to become a radiologist — 12 years of talking to other doctors, observing and processing the world, and filling the databases inside their heads. There’s value in this.”
Unanimous AI’s platform leverages swarm intelligence — a biological phenomenon where groups of organisms amplify their intellect by forming real-time systems — to improve prediction accuracy.
“Animals use unique techniques they’ve evolved over millions of years with feedback loops that converge on an optimal combination of insights,” Rosenberg explained. “[Scout] bees … vibrate their bodies, and birds detect motions propagating through flocks.”
Unlike the swarm intelligence observed in nature, however, Unanimous AI adds AI to the mix in a process it calls artificial swarm intelligence. Here’s how it works: human participants — in this case radiologists at Stanford and other institutions across the country — log into Unanimous AI’s platform using a networked computer. Collectively, they attempt to move a cursor with a mouse, touchpad, or touchscreen toward a prediction (i.e., a diagnosis) while algorithms process their behavior in real time.
It’s a two-step process — participants converge on a coarse range of possibilities and then on a refined value — but one that’s quicker than you might expect. Radiologists in the study settled on 50 diagnoses in about 60 seconds on average (and as quickly as 33 seconds), with some exerting more influence than others. That’s by design — Unanimous AI’s algorithms infer swarm members’ convictions based on their cursor motions and weigh their contributions accordingly.
“Over the last four years, [the system] has learned from hundreds of thousands of questions,” Rosenberg said. “You can think of it as a system of feedback loops. It’s aggregating sentiment.”
In the end, the radiologists achieved 82 percent diagnostic accuracy overall compared to human experts’ 73 percent accuracy. “[S]warm-based technologies are quite promising for use in medical diagnosis,” the authors of the study wrote.
It’s the first time Unanimous AI’s tech has been applied to the medical field, but hardly the first time it’s made headlines. In 2016, it correctly predicted the results of the Kentucky Derby, beating betting odds of 542 to 1. It’s also forecasted the past three years’ worth of Oscar winners more accurately than any publication, as well as the Time Person of the Year, President Trump’s approval rating, and the exact score of last year’s Super Bowl.
For Rosenberg, who helped to pioneer one of the world’s first augmented reality systems for the U.S. Air Force in the early ’90s, swarm intelligence is a way to amplify human abilities in ways that AI alone cannot.
“Most AI companies are focused on finding patterns in big data, basically, and that only works for data you can put into a database in a very consistent way,” he said. “The question we ask is, how can we connect groups of people together and make them smarter using AI? How can we take groups of people and turn them into artificial experts?”
Unanimous AI currently has two products in its portfolio, both of which operate on a system-as-a-service model: Swarm Insight and Swarm AI. The former has an enterprise bent — one client used it to predict how customers might react to a television ad for pancakes, and another — beverage company Constellation Brands — had it gauge customer reactions to its investment in cannabis. Swarm AI, meanwhile, taps a company’s workforce for consensus; engineers at Boeing use it to aid in cockpit designs.