Using Machine Learning to Predict Ground Reaction Forces

Johns Hopkins

Mentor: Dr. Ryan Roemmich

Center for Movement Studies at the Kennedy Krieger Institute, Johns Hopkins University

Every year there are almost 800,000 stroke victims in the US alone, about 80 percent of whom lose some walking abilities and require physical therapy. It is important for therapists to be able to track a patient’s recovery using gait metrics that can indicate how far a patient has progressed. One measurement that has been shown to correlate with the severity of a victim's loss in walking abilities, and could possibly help track progress in physical therapy, is ground reaction forces, the force you exert on the ground when you walk. Measuring GRFs requires expensive equipment, like pressure pads, and trained staff, making it hard to use in a clinical setting. This study explores the use of machine learning to predict GRFs from video, which could allow them to be found with little to no equipment or training. Six different algorithms were trained with data from the Center for Movement Studies at the Kennedy Krieger Institute, found using motion capture and pressure pads. The most accurate of the algorithms was the random forest, with 90% accuracy. This shows the viability of machine learning in predicting GRFs, though more testing is needed to see if it can work in a clinical setting.