Morehouse neuroscientists are hopeful that data-driven strategies can help doctors more swiftly detect a disease that kills one American every four minutes.
How would you know if you were having a stroke? The American Heart Association counsels us to think FAST—if you notice Face drooping, Arm weakness, and Speech difficulty, it’s Time to call 911. But the outward signs of a stroke are sometimes so subtle that they go unrecognized, even by the trained eye of a physician. In those cases, reports a new study by neuroscientists at the Morehouse School of Medicine and Emory University, a computer algorithm might be our saving grace.
The team, led by Morehouse’s Roger Simon, showed that computational might can be harnessed to identify a stroke’s genetic “fingerprint.” Of the 20,000-odd genes that make up our genome, only a fraction are expressed at any given moment. Those “on” genes–or more specifically, the messenger RNA that they encode–are known as the transcriptome. During a stroke, some genes that are normally “off” turn “on,” and others that are normally “on” turn “off,” producing subtle changes in the transcriptome. The exact nature of those changes varies from person to person, but Simon and his colleagues were able to train a computer to recognize certain universal patterns.
The process, a type of machine learning, works a lot like image-recognition software. As that software sees more and more pictures of, say, cats and dogs, it starts to pick out distinguishing features. Eventually, it can look at a picture it’s never seen before and tell whether that picture is of a cat or dog. No one ever has to tell it that cats have whiskers and dogs have floppy ears.
Simon and his coworkers trained their recognition software not on cats and dogs but on transcriptomes collected from patient blood samples. After being trained on a few dozen subjects, half of whom had just had a stroke, the computer could detect strokes with about 90% accuracy. The study focused exclusively on African-Americans, who are nearly twice as likely as Caucasians to suffer a stroke.
The new study adds to a growing body of exploratory work on diagnostic machine learning, including studies that used similar algorithms to detect cancer and auto-immune diseases. But researchers still have one big hurdle left to surmount: They need a lot more data.
Sophisticated image recognition algorithms require many thousands of training photos to give reliable results. Clinical studies, like the one done by Simon and his coworkers, usually involve only a few dozen or so participants. With so small a sample, it’s impossible to know whether a clinical study’s success can be replicated in the public at large. Results that appear rosy might well be flukes.
Still, Simon and his colleagues are optimistic that machine learning can one day help doctors more swiftly identify and treat stroke victims. If so, the AHA’s famous FAST acronym might take on new meaning: Face drooping, arm weakening, and difficulty speaking might mean it’s time not only to call 911 but to boot up the nearest computer.