Implementation of Stacking-Based Algorithms with Data Pruning on Qualcomm Snapdragon 820c

Falls, especially those left unattended for an extended period, are serious health issues for the elderly. Due to the current advancements in the Internet of Things(IoT) and Machine Learning (ML), several fall detection-based systems have been proposed. Many of these fall detection systems are not practical in the real world since either the sensor used is the ambient sensor that captures the images or sounds surrounding the elderly or are wearable sensors worn on the thigh, torso, waist or a combination of these. Ambient sensors confine the movement of the elderly to a single room and are a very costly solution. The wearable sensors on the thigh, torso or waist maybe extremely 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 a Stacking-based ensemble algorithm, which we term Stack (A). This ensemble algorithm runs on the end device with IMU(inertial measurement unit) and heart rate sensors. Qualcomm Snapdragon 820c is the System on Chip (SoC) around which the end device is built. We use feature extraction and feature selection to prune the data to execute the algorithm on the end device. 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; 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.