Example clocks, to scale, from the research dataset (Click image to enlarge)
Update 5/04/16: The work detailed here won the 2016 INFORMS Innovative Applications in Analytics Award.
When patients who might suffer from Parkinson’s disease or Alzheimer’s disease visit a doctor, they’re often asked to draw a clock, hands pointing to a specific time of day. The clock drawing test has been used to measure a variety of aspects of cognitive function for more than 50 years.
But quietly over the past nine years researchers at Lahey Hospital and Medical Center outside Boston, as well as other clinics, have administered the test using a digitizing ballpoint pen that captures the drawing—and the drawing process—with considerable accuracy.
Based on this sizable collection of data, researchers at Lahey, MIT, and elsewhere have created a software system that analyzes the data and demonstrates accuracy substantially better than systems currently in use. The research behind the software [PDF] will soon be published in Machine Learning.
The new system, called the “digital Clock Drawing Test,” detects and analyzes information that was previously imperceptible: hesitations, uneven lines, and other behaviors once unseen. Such precise data may allow doctors to diagnose and treat illnesses like Alzheimer’s disease and Parkinson’s disease sooner, leading to better quality of life and reduced health care costs.
The software takes advantage of the fact that the data is time-stamped, allowing it to detect and analyze pauses, such as those between drawing the numbers and starting to draw the clock hands. The system also determines the amount of time people spend holding the pen rather than drawing, which is referred to as “think time.”
“Our system allows us to take a long-used, well-accepted test and extracts much more information from it; then the machine learning methods help in determining what’s important in the data coming from the pen,” says Professor Randall Davis of the MIT Computer Science and Artificial Intelligence Laboratory, an author of the study.
The researchers used data from some 2,600 tests to train their system to recognize signs of cognitive impairment. To do this, they used powerful machine learning techniques to build models and tested those models against the standard tests used by physicians. This showed that the new models were more accurate.
“Some of the machine learning techniques were designed to be transparent, providing us some insight into what factors in the drawing were important in the classification task,” says MIT Sloan Professor Cynthia Rudin, an author of the study.
For example, while healthy adults spend more “digital Clock Drawing Test” time thinking—with the pen off the paper—than inking, memory-impaired subjects spend even more time thinking than inking.
The system also analyzes significant latencies, such as the pause between drawing numbers and then drawing hands. Where drawing numbers might be rote, drawing the hands measures explicit reasoning, Davis explains. Longer latencies in making the shift between these tasks appear to suggest cognitive impairment.
Finally, the system can also detect minute drawing patterns, such as “hooklets,” tiny, sharp turns at the end of one stroke that point toward the beginning of the next. Hooklets may indicate that a person is anticipating what to draw next—meaning they’re planning ahead. As planning ahead is one sign of cognitive health, producing no hooklets might be an early sign of cognitive impairment.
The new system could remove an element of subjectivity for doctors who must now rely on checklists that ask such things as whether a clock circle has “only minor distortion.”
Davis and Rudin point out that the system is still in a research phase but could soon be a useful tool for doctors, assisting them in arriving at a better diagnosis.
“Given the greying of the population worldwide, we’re facing an enormous health care problem in the coming years,” Davis says. “In the absence of an effective drug, by 2050 there could be 14 million people with Alzheimer’s, with health care costs running to one trillion dollars [PDF].”
“So it’s crucial to treat them sooner and slow the progression,” he says. “This is a case where artificial intelligence and machine learning methods can make an important difference.”
The paper, Learning Classification Models of Cognitive Conditions from Subtle Behaviors in the Digital Clock Drawing Test, was written by William Souillard-Mandar, Davis, and Rudin of MIT; Rhoda Au of Boston University School of Medicine; David J. Libon of the Drexel Neuroscience Institute; Rodney Swenson of University of North Dakota Medical School; Catherine C. Price of University of Florida, Gainesville; Melissa Lamar of University of Illinois, Chicago; and Dana L. Penney of Lahey Health.