Performance evaluation of Machine Learning algorithms on System on Chips in Wearables for Healthcare Monitoring

Abstract

Compressed Machine (ML) and Deep Learning(DL) techniques are the emerging areas in research, the use of compressed ML and DL algorithms is recommended on System on Chip(SoC) platforms as the latencies of ML or DL algorithms will be extremely high. This paper explores the need to use compressed ML/DL techniques by analyzing the various popular ML algorithms by implementing them on different platforms including SoCs, in particular Snapdragon 410c which will be used on the wearable biomedical device.This paper looks at fall prediction using wearables. The wearable device uses 410c as the platform and data is collected using IMU and heart rate sensors. A data set was created by collecting IMU and heart rate sensor values across different fall and non fall Activities of Daily Living(ADL) activities, over 70k data points was collected of which 70 percent was used to train the algorithm and 30 percent was used as test data. The same data set was used across varying algorithms and platforms and the performance as well as the latencies were analyzed.It was found that the latencies was found to be greater than 120 sec for even K Nearest Neighbours(KNN) when run on SoCs. This clearly indicates that we cannot use classical ML algorithms. When a public dataset SmartFall which had 92,780 data points was used, Support Vector Machine(SVM) algorithm did not converge even after several hours of computing. In case of Random Forest it took 2235 seconds for the algorithm to converge. As the size of the dataset increased obviously there was an increase in latency with complex ML algorithms are unable to converge, hence this research indicates the necessity of compressed ML algorithms.