Machine learning classifies predictive kinematic features in a mouse model of neurodegeneration

Abstract

Abstract Motor deficits are observed in Alzheimer’s disease (AD) prior to the appearance of cognitive symptoms. To investigate the role of amyloid proteins in gait disturbances, we characterized locomotion in APP-overexpressing transgenic J20 mice. We used three-dimensional motion capture to characterize quadrupedal locomotion on a treadmill in J20 and wild-type mice. Sixteen J20 mice and fifteen wild-type mice were studied at two ages (4- and 13-month). A random forest (RF) classification algorithm discriminated between the genotypes within each age group using a leave-one-out cross-validation. The balanced accuracy of the RF classification was 92.3 ± 5.2% and 93.3 ± 4.5% as well as False Negative Rate (FNR) of 0.0 ± 0.0% and 0.0 ± 0.0% for the 4-month and 13-month groups, respectively. Feature ranking algorithms identified kinematic features that when considered simultaneously, achieved high genotype classification accuracy. The identified features demonstrated an age-specific kinematic profile of the impact of APP-overexpression. Trunk tilt and unstable hip movement patterns were important in classifying the 4-month J20 mice, whereas patterns of shoulder and iliac crest movement were critical for classifying 13-month J20 mice. Examining multiple kinematic features of gait simultaneously could also be developed to classify motor disorders in humans.