PNS230 USING MACHINE LEARNING FOR SIGNAL DETECTION IN REALWORLD DATA FROM WRISTWORN WEARABLE DEVICES TO IDENTIFY FRAUDULENT BEHAVIOUR

Abstract

Publicly available machine learning algorithms can be used to identify fraud in real world data from wearable devices. Data from adult volunteers was used to train a Machine Learning Algorithm to detect a change in the raw data from wrist-worn accelerometer as it occurs when the same device is worn by different patients. We analysed data sets from healthy adults who wore devices on their wrist during a 3-5 day period. Publicly available Machine Learning libraries were used to develop and train an algorithm to detect unique patterns in raw data from wrist-worn accelerometers. These patterns were unique for each patient. Only a very small sample was needed to train the algorithm and to detect patterns. The algorithm allowed us to show differences in the data between different patients when wearing the same device. This identified reliably a switch between patients wearing the same device as it would occur during fraudulent behaviour in a clinical trial. With the growing utilisation of wearable devices in clinical research and healthcare and the lack of any access control, we are now in a position to identify potential fraud early in the process. It is possible to improve data quality significantly during and after a clinical trial, if the raw data from the accelerometer is available.