Implementation of Ensemble Algorithm with Data Pruning on Qualcomm Snapdragon 820c

Falls, especially those left unattended for an extended period are serious health issues for the geriatric population. Several efforts have been made in the use of IoT and ML algorithms in fall detection. These systems have issues in terms of sensor placements and connectivity since most of the data analysis is done on the cloud. Sensors are usually placed on the thigh or the torso this may prove to be very uncomfortable for the elderly. Also, constant disconnects from the network and high latencies may delay a Fall alert. Therefore, we have built a wrist-worn end device that uses the concept of Dew computing, where the computing is done on the end device itself. In order to have good prediction accuracies, we have also developed three ensemble algorithms; the first one is based on stacking, we call this Stack (A), the other two are based on voting, and we term them as Variable Weight Ensemble (B) and (C). This ensemble algorithm runs on the end device which is equipped with IMU and heart rate sensors. Qualcomm Snapdragon 820c is the SoC around which the end device is built. To execute the algorithms on the end device we use feature extraction as well as feature selection to prune the data. We obtained an accuracy between 97.5% to 98.1% for various algorithms and a sensitivity of almost 100%. In this paper we present a complete analysis of the Fall detection system (with and without pruning), we have analysed the accuracies and the latencies on 820c while using ensemble algorithms specifically. We have compared the performance of our algorithms with XGBoost, AdaBoost and RF ensemble techniques.