With the surge in wearable device data in clinical trials, ensuring the authenticity and attribution of this data to the correct individual has become paramount. This research sheds light on the application of Machine Learning to detect fraudulent data generation from wrist-worn devices.
Key Insights:
Data Authentication Challenge: Clinical trials need data that can be reliably attributed to its source. However, wearable devices pose challenges since users could employ alternate means (like relatives or even washing machines) to generate data.
Machine Learning to the Rescue: Given the complexities and “black box” nature of Machine Learning, this study explored its application in verifying data authenticity. Deep Learning Image Analysis stood out as a promising method.
Objective: The goal was to determine the capability of Machine Learning in identifying fraudulent behavior when using wrist-worn activity devices in clinical trials.
Methodology: Using PhysioNet’s 3-dimensional accelerometer data from Shimmer 3 GSR+® devices, raw data from six adults was processed and converted into visual representations. These were then used with the Inception-v3 Machine Learning model, originally designed for image classification (e.g., dogs vs. cats), but adapted here to classify data to specific persons.
Impressive Results: The trained model exhibited high accuracy:
The training phase reached a prediction accuracy of over 90% within just a few cycles.
When tested on unseen data, the algorithm correctly predicted 97.9% of the instances. The accuracy ranged between 96% to 100% across the six datasets.
With these findings, Machine Learning proves to be a powerful tool in verifying the authenticity of wearable device data, bringing more trust and reliability to clinical trials.
