Friday, August 19, 2016

Artificial Intelligence, Machine Learning, and the FDA

(A version of this column, with hyperlinks, was published by Forbes.)

In July, the Food and Drug Administration issued guidance on three topics important to the future of medical innovation. These welcome guidelines demonstrate the FDA is doing the best it can to ensure it does not interfere inappropriately with advances in medical technology that rely on processing information.

However, the guidelines also show the FDA will be limited in its ability to respond effectively to future innovations. Current law does not really define the FDA’s powers to regulate devices that depend on advances in artificial intelligence and machine learning, as applied to health care. Guidelines give medical entrepreneurs some comfort the FDA will not impose an undue regulatory burden on them, but they are no substitute for legislation precisely defining the FDA’s powers in the digital age.

Fortunately, new legislation that moves in the right direction - the 21st Century Cures Act - has passed the House of Representatives and will hopefully finish its passage through the Senate quickly enough that reconciliation between the two chambers can take place, and a good bill will pass before the 114th Congress adjourns.

Consider the first new guidance, which addresses “general wellness products.” These include “audio recordings, video games, software programs and other products that are commonly, though not exclusively, available from retail establishments.” The good news is the FDA confirms it will not regulate such products as medical devices, as long as they meet two factors: They are intended for only general wellness; and present low risk to users.

Such a device may claim that it “may help to reduce the risk of” or “may help living well with” certain chronic diseases. An acceptable claim for a software product might be that it “coaches breathing techniques and relaxation skills, which, as part of a healthy lifestyle, may help living well with migraine headaches.” The product’s value derives from information, rather than doing something directly to the body.

Information goes in the other direction, too. The second guidance addresses the use of “Real World Evidence” in research. “RWE” derives from data collected outside clinical trials. Although not usually used to win approval of a new device, RWE can be used to gain the FDA’s permission for a device to be used for more indications that the one for which it was originally approved. What is the source of data to build the evidence? “The data is typically derived from electronic systems used in health care delivery, data contained within medical devices, and/or in  tracking patient experience during care, including in home-use settings.”

Finally, the third guidance addresses adaptive design of clinical trials supporting the FDA’s approval of new medical devices. “Adaptive” refers to “a clinical study design that allows for prospectively planned modifications based on accumulating study data without undermining the study’s integrity and validity.”

If poorly executed, adaptive design risks moving the goalposts in the middle of the game, posing hidden risks to patients. If well executed, adaptive design can reduce the time and cost of clinical research. The guideline describes how to implement good adaptive design in clinical trials. If sponsors follow these guidelines, new devices will get to patients quicker. How is this data going to be collected? See previous paragraphs!
And the ocean of data will not just flood between patients, doctors, device and drug makers, and regulators. Data is also stored in the cloud, to be used for diagnosis and treatment decisions. In 2001, it cost $100 million to sequence the human genome. By 2015, the cost had dropped to around $1,500, according to the National Human Genome Research Institute.

However, not all sequencing has to be done at the highest level of accuracy, nor does the entire genome have to be sequenced to produce useful information. Limiting the sequencing to the protein-coding oregions of a genome reduces the cost to under $1,000. This is making genome sequencing increasingly available in forward-looking  medical facilities. Illumina, which makes genome-sequencing technology, is enjoying double-digit revenue growth that analysts expect to continue.

President Obama has invited one million people to contribute their personal health data to the Precision Medicine Initiative. If promoted correctly, patients’ sense of public service will overcome their privacy concerns, and the million-person cohort will furnish scientists with an unprecedented trove of real world data.

IBM has already shown significant success with “Watson for Oncology.” Watson, IBM’s artificial intelligence system (which became a pop culture celebrity by beating Jeopardy! champions),  is being used in hospitals worldwide to better inform oncologists’ decisions. A recent news report goes so far as to claim Watson diagnosed a Japanese woman’s rare leukemia at the University of Tokyo after months of fruitless effort by doctors. Within 10 minutes, Watson had reviewed 20 million research papers and recommended the right course of treatment.

This is just the beginning. As Forbes contributor Todd Hixon has described, there is a flood of interest in machine learning, with large companies having acquired dozens of start-ups within the last five years. Yet, terms like "artificial intelligence" or "machine learning" are not used in current FDA guidance.

Congress and the president have an opportunity - indeed, a duty - to amend the FDA’s regulatory powers for the 21st century so patients, doctors, scientists, and medical entrepreneurs can be confident these new technologies can come to market in an environment of regulatory clarity.

No comments: