OBJECTIVES

Continuous patient authentication during a clinical trial or in a healthcare setting may be advantageous over discrete authentication at the beginning of an action i.e. exercise. The objective of this project was to determine if single lead ECG data from a wearable device with dry electrodes can provide data to positively authenticate a user, similar to finger-printing.

METHODS :

Unfortunately publicly available ECG data was not eligible for this analysis and a new data base with ECG data from 33 healthy volunteers with a total recording time of about 244 hours was setup. Based on activity data recorded in parallel we discarded ECG data that was not collected while patients were at rest in a sitting position. Data from 18 volunteers with a minimum of 10.000 recorded heartbeats was then included in the training of the machine learning algorithm. In a second step this algorithm was then presented with data ECG data from various persons to test the accuracy of its authentication prediction.

RESULTS :

We were able to positively authenticate users with a Balanced Accuracy Rate of 89.8%. Only 2 out of 576 tests yielded false positives, means the system did not detect a change in who wore the device and we achieved a perfect 100% True Negative Rate for the remaining 574 tests, at the expense of a lower True Positive Rate.

CONCLUSIONS :

With an increase in remote data collection in clinical trials and healthcare it is imperative that we develop systems to detect potential fraudulent behaviour. This project shows that a single lead ECG with additional Inertial Measurement Units (i.e. actigraphy), has potential to continuously authenticate users. More work needs to be done to improve the performance of the algorithm itself and to investigate further parameters that influence the training accuracy.