Post-fall Detection Using ANN Based on Ranking Algorithms

Abstract

The purpose of this study was to develop an accurate and efficient algorithm for post-fall detection. Thirty healthy male subjects were recruited and asked to perform 23 movements, comprising 14 activities of daily living and nine fall motions. The algorithm was developed using the ANN toolbox provided with MATLAB and inertial measurement unit (IMU) data were used to distinguish between fall and non-fall cases. An IMU sensor was located at the center between the left and right anterior superior iliac spines. A total of 32 feature vectors were extracted from 3-axis acceleration and angular velocity signals. Based on the five different ranking algorithms (relief-F, T-score, correlation, Fisher score, and minimum redundancy maximum relevance) used, feature vector subsets comprising the feature vectors were created and subsequently evaluated. Accuracy was compared according to the number of feature vectors constituting the subset, which were based on rank-lists. The results showed that the subset comprising all the feature vectors showed the best accuracy (99.86%), but a similar accuracy could be obtained with a subset comprising fewer feature vectors. The T-score was found to be the most optimal among the five ranking algorithms. Furthermore, T-score with two feature vectors achieved an accuracy of 99.17%. The results of this study are expected to assist in the construction of subsets of feature vectors based on ranking algorithms for post-fall detection with high accuracy and less computational cost.