Feature selection for the classification of fall-risk in older subjects: a combinational approach using static force-plate measures

Abstract

IntroductionFeature selection prevents over-fitting in predictive models. This study aimed to present an effective feature selection method that leads to a reliable classification of fall-risk in older subjects using static force-platform data across four conditions only.\n\nMethod528 features were generated from a publicly available dataset of force-plate signals from 45 low-risk and 28 high-risk subjects. Subjects were classified as high- or low-risk if they recorded [≥]1 falls in the prior 12 months and/or were rated as high-risk on the FES. The feature selection protocol included SVM-RFE, GA and ReliefF and finally SAFE. Several machine-learning models were then used to evaluate classification performance.\n\nResults67 features were identified after the three-fold process which was further reduced to 18 features after SAFE. The MLP achieved the highest average classification accuracy of 80%. All classification models evaluating this final subset displayed high variance across all performance metrics, especially in terms of sensitivity to high-risk subjects.\n\nInterpretationAn optimal feature set of static force-plate measures was insufficient in creating a reliable classifier of fall-risk. This was due potentially to the limited information about fall-risk that could be provided by such measures leading to under-fitting/over-fitting being unavoidable and appeared to be centered around an insensitivity to high-risk subjects.\n\nConclusionStatic stability measures have shown some usability in fall-risk classification however feature sets limited to such measures are inadequately sensitive to high-risk subjects. The utilized feature selection methods demonstrated their ability to identify relevant stability measures and could be used successfully on dynamic measures.