The FDA is looking at new rules to govern AI, but is the rulebook outdated?

October 08, 2019
by Sean Ruck, Contributing Editor
Rik Primo, the principal of Primo Medical Imaging Informatics Inc., has built a career in healthcare over decades. Still, he had to go back long before his career began, in fact a couple of centuries before he was born, to provide a background on today’s artificial intelligence in healthcare. Part one’s history lesson about AI and its fundamentals sets the stage for part two, published next issue, which will focus on FDA activities and actions with “Software as a Medical Device.” There will be a few timeline jumps, so pay attention!

Rik Primo
We start in the mid-1700s to early 1800s. Mathematician Jean-Baptiste Joseph Fourier developed his mathematical physics models like infinite mathematical series in his book “The Analytical Theory of Heat”. They would become the foundation of the Fourier analysis and Fourier transform. The latter is a mathematical process now used for breaking down MR signals to different frequencies, phases and amplitudes of sine waves. Much of Fourier’s work is used today in mathematics, physics and geometry, finding its way into computing by the 1960s, and playing an extremely important role today in digital signal processing and analytics, artificial intelligence and other domains.

The work of Pierre-Simon, Marquis de Laplace from the same era also heavily figures into today’s healthcare technology. As Primo explains, “The algorithm, known as a Laplace operator or Laplacian, is effective in detecting edges in images. AI applications in imaging are dependent on being able to detect edges.” Primo says a third man, Simon Bayes, also figures in today’s AI decision-making through the Bayesian theorem. Bayes’ work dates back to the early 1700s.

In 1956 at a conference at Dartmouth College, the term "artificial intelligence" was mentioned for the first time. Prominent scientists John McCarthy and Marvin Minsky, known as the “Fathers of AI” fueled the continuous progress in AI research and development.

Today, Primo points to Snapchat, a current social media application that uses AI technology. “In Snapchat, the application performs facial recognition which is accomplished by digitizing and quantification of an image – with pixels and grayscales and with recognizing different colors. When your face is exposed to the camera and the picture is taken, machine learning comes in,” he says. “There have been hundreds of thousands of images reviewed by people who have clicked and marked the eyes, the eyebrows, the nose, the mouth, and so on. The application is then using the Bayesian inference theorem, Boolean algebra, Laplacian edge detectors, and after a while, the system can clearly and autonomously identify, even under different angles, the location of the nose, eyes, hair from the user in the picture. Thus, Snapchat is based on machine learning, which is an AI mechanism. An important difference with AI deployments in healthcare is that kids intuitively use Snapchat on their tablets or smartphones without the need for extensive training.

From 2011 (the year of Snapchat’s introduction) we jump back to the 1960s and 1970s. In 1962, sci-fi author Arthur C. Clarke’s first adage was published that would become known as Clarke’s Three Laws. By 1973 the third in the trio saw the light of day. Primo explains that the first law has to do with “when a distinguished and elderly scientist states that something is possible, he is almost certainly right. When he states that something is impossible, he is very probably wrong.”

Primo says the second law states that the only way to discover the limits of a possibility is to venture a little bit past them into the impossible. “The third law, any sufficiently advanced technology is indistinguishable from magic, is where we are today. We see it with Snapchat, where the application does seemingly impossible things with images. AI has come a long way and where it will end is probably never,” he says.

The final leap takes us back to present day. The FDA faces a regulatory challenge with medical devices based on software only, Primo says. These devices are actually software applications and better-known under the term “Software as a Medical Device” (SaMD). This terminology was defined by the International Medical Device Regulators Forum and these devices don’t necessarily involve hardware. To be effective, the FDA needs to be in step with those developing medical devices and software based on AI. When a manufacturer has developed the AI-based SaMD, the manufacturer’s organization needs formal FDA approval before marketing and commercializing this SaMD application. The FDA will make the determination about whether this specific SaMD is performing as safe and effective, as proposed. Is it safe for the healthcare provider and patients? Under which risk classification should the SaMD be categorized. Class I, II or III?

HCBN will cover these topics in the second installment, coming next month!