Personalized fall detection monitoring system based on user movements

Abstract

High accuracy, fall detection systems is a fundamental requirement among the increasing elderly population, mainly due to expensive healthcare and a shortage of nurses for home-care. Fall detection systems have evolved over the past few years, from a button pendant to three newer types of fall detection systems - wearable sensors, ambient sensors, and camera-based sensors. Wearable sensors are regarded as the most popular, as it provides both indoor and outdoor monitoring and is the least expensive among the newer fall detection systems. Detection of a fall, using wearable sensors, started off at first by using a threshold method, where the features extracted from the wearable sensor data are compared to a pre-defined value. The problem with this approach is that the pre-defined value only works on a small set of people with certain user characteristics. It was also difficult to set a value that can distinguish between everyday activities and fall activities. To solve this problem, supervised machine learning algorithms were incorporated - these obtained higher accuracies when compared to the threshold method. Supervised machine learning algorithms achieved high accuracy during laboratory experiments. In a practical scenarios, the performance of these fall detections were low, due to the supervised machine learning algorithms making use of simulated fall data which is performed on a soft mattress which does not represent a real fall event (which is usually spontaneous). Since it is difficult to obtain real fall data, a lot of studies make use of simulation data. Using artificial fall as training data can result in over-fitting, which causes poor decisions. Both threshold and supervised classifiers cannot provide a user-specific solution for each individual user. Since supervised machine learning algorithms require everyday activities and fall activities to classify, as well as limited fall data (which creates an imbalance i.t.o classification), it is hard for these algorithms to classify accurately. Another problem is that these systems are limited to a certain number of activities that a user can perform, and it does not work for everybody. User-specific personalization can be provided using unsupervised machine learning algorithms, resulting in the following advantages: a) more activities can be included in the classifier, and b) the fall detection system can address the inter-individual differences. In this research, the effects of personalization models using user movements are analysed (in terms of accuracy). A low-cost smartphone accelerometer sensor was used in the system. The study was divided into two parts: a simulation phase and an experimental phase. The simulation phase made use of a public dataset known as the tFall dataset. The type of input data to be used, which machine learning algorithm to use and the different types of personalization models, were investigated. For the type of input data, the following were considered: raw accelerometer values, statistical features extracted from the accelerometer, principal component analysis on the statistical features extracted, or the statistical features selected from the variance-threshold feature selection method. Both supervised and unsupervised machine learning algorithms were implemented to determine the best algorithm. The following unsupervised machine learning algorithms were implemented: nearest neighbour, oneclass support vector machine, angle based outlier detection, and isolation forest. Angle based outlier detection and isolation forest were not implemented in any fall detection systems before. For the supervised machine learning algorithm, the two most popular machine learning algorithms were selected: support vector machine learning algorithm and k-NN. The following models were tested: a) Model 1 ð€€€ the classifier itself; b) Model 2 the non-fall activity is retrained whenever the classifier correctly detects a non-fall activity; c) Model 3 the false positive is retrained when the classifier detects a non-fall activity as a fall activity, and d) Model 4 combining Model 2 and Model 3. The unsupervised machine learning algorithm is applied to all the models, whereas supervised machine learning algorithm is only applied to Model 1. During the simulation phase, the following evaluation parameters were used: sensitivity, specificity, geometric mean and F1-measure. During the experimental phase, the best input data set (raw accelerometer values), model (Model 4) and classifier (angle based outlier detection) were implemented on an Android smartphone, to demonstrate how accurately the fall detection can classify. From experimental results, it was shown that personalization models using user movements can improve the overall performance of the system, achieving a sensitivity of 90.48% and specificity of 92.31%.