A data mining approach for determining biomechanical adaptations in runners who experienced and recovered from patellofemoral pain syndrome

Abstract

ABSTRACT Patellofemoral pain (PFP) is a common musculoskeletal pain disorder experienced by runners. While biomechanics of those with PFP have been extensively studied, methodological considerations may omit important adaptations exhibited by those experiencing and recovered from pain. Instead of a priori selection of discrete biomechanical variables, a data mining approach was leveraged to account for the high dimensionality of running gait data. Biomechanical data of runners symptomatic for, recovered from, and who had never experienced PFP were collected at the 1st (M1) and 21st (M21) minutes of a treadmill run. Principal component analysis and a logistic regression model were used to classify healthy and symptomatic runners, and a feature ranking process determined the important features. The M1 model achieved an accuracy of 82.76% with features related to knee flexion angle, hip abduction moment and gluteus maximus activation, while the M21 model required an additional nine features to achieve an accuracy of 79.31%. Data for recovered runners were projected onto the models, resulting in five and seven out of twelve symptomatic classifications at M1 and M21, respectively. Following the onset of pain, a greater number of features were required to classify runners with PFP, suggesting they may experience individual pain adaptation strategies.