Fall detection in elderly people: impact of sensor position and feature selection

The most predictive sensors, sensor positions, and features for detecting falls early in elderly people were investigated using feature dataset from three-axis accelerometers placed on the head, pelvis, right and left shank, and pressure-sensing insoles. A feature database containing records of 100 older people (76 non-fallers and 24 fallers) was used. The three different feature selection algorithm was used, and most predictive feature vector obtained. An SVM classification model was developed for each feature vector. As a result, the best classification accuracy was observed for features determined by the feature selection algorithm using the chi-square test. In the classification using ten features of the acceleration signal recorded from the head, 80.17% (±8.33) accuracy, 43.99% (±21.93) sensitivity, and 91.59% (±9.57) specificity were obtained. These results demonstrate that a sufficiently accurate model can be developed using a sensor and activity when the correct feature selection and classification algorithm is determined.