Using Deep Learning to Understand Lack of Exercise
Exercise is generally considered a healthy activity. For example, walking can reduce risk of a cardiovascular event, such as a heart attack. Unfortunately, studies report that roughly half of those over 65 years or older do not regularly exercise. To address this issue, Aetna has programs to encourage exercising. In support, we have built a deep learning model that predicts the probability that a member is exercising and, more importantly, to provide clues as to why a person is not exercising. A sequential model with an attention mechanism estimates the probability of exercise as well as highlights prior events in a person's medical record that most contributed to the estimate. The findings demonstrate good predictive performance, such as 77.2% positive predictive value among the top scoring 10%. Commonly found drivers of not exercising include chronic obstructive pulmonary disease, asthma medication, and depression. Examination of individual cases often reveals clear narratives that help to tailor the exercise program, such as special dance classes for those with a portable oxygen concentrator. More broadly, the work shows how deep learning finds revealing events in a person's healthcare journey with an eye towards providing more effective care.
Mark Homer is Head of Healthcare AI at Fidelity. He has been developing machine learning algorithms and solutions in healthcare for over a decade. Mark received his BS and MS from MIT, PhD in biomedical engineering from Brown University, and an MMSC from Harvard Medical School where he won a National Library of Medicine fellowship award. He has over 20 years of experience in AI, working in a variety of domains including drones, bioreactors, and brain-computer interfaces.