Genetic-algorithm-based feature-selection technique for fall detection using multi-placement wearable sensors

For a machine-learning-based fall-detection approach using wearable sensors, having a high number of features can not only cause a reduction in the detection rate because of irrelevant features, but it can also cause a high computational cost. Therefore, the number of features needs to be reduced through a feature-selection technique. However, current studies in fall-detection only consider features that can give an optimum detection rate without considering their computational cost. Having features with a high computational cost on wearable devices can cause their battery to drain fast. This paper presents a genetic-algorithm-based feature-selection technique that can search for a subset of low-computational-cost features from different sensor placements, where those features can give a relatively good detection rate in terms of F-score. The experimental results show that our technique is able to select a subset of low-computational-cost features that can achieve up to 97.7% of F-score on average. Compared to the SelectKBest and Recursive Feature Elimination (RFE) techniques (a filter and an embedded feature-selection technique, respectively), our approach is able to select features that can give a comparable F-score and significantly lower computational cost.