In our previous project [2], we defined personal risk detection as the timely identification of a situation when someone is at imminent peril, such as a health crisis or a car accident. A risk-prone situation should produce sudden and significant deviations in user patterns, and the changes can be captured by a group of sensors, such as an accelerometer, gyroscope, and heart rate monitor, which are normally found in current wearable devices. Previous research findings were published in [2, 11] and presented at HPI Future SOC Lab. The present work rises with the aim of improving our previous results. In order to achieve it, the following three approaches were tested: 1) a visualization method in real-time of PRIDE users leveraged with a one-class classifier called Bagging-TPMiner, 2) the addition of frequency-domain features to the time-domain features embraced in the PRIDE dataset, and 3) improve the accuracy obtained by previous one-class classifiers through testing a cluster validation algorithm. We were able to report part of our results in [8], which have been recently submitted for publication. Although experiment results reported in this document are encouraging, due to the sheer amount of data, the results presented in this report are partial. In order to fulfil the experiments, we are submitting an extension at HPI Future SOC Lab for the period that ends on April 2017.
