Machine Learning for Continuous User Authentication via ECG Data in Clinical Trials
Harnessing the power of wearable devices for data capture during clinical trials and healthcare applications is becoming a mainstay. However, the looming threat of fraudulent behavior necessitates rigorous verification mechanisms to ensure data authenticity.
Main Highlights:
The Issue at Hand: With incentives like reduced premiums for active lifestyles, there’s an increased risk of fraudulent behavior. Both healthcare providers and clinical trial sponsors need reliable means to authenticate the origin of the data. Regulatory bodies also emphasize the requirement for data that can be directly attributed to specific patients.
Previous Efforts: Prior studies had leveraged Machine Learning, analyzing data points like PPG and Actigraphy, to link patients with their data. However, these methods authenticated data post-collection, sometimes leading to the exclusion of ambiguous data sources.
Current Study’s Approach: This study introduces a real-time, continuous authentication method analyzing single lead ECG data from chest straps throughout the observational period, similar to how fingerprints are used for identification.
Objectives: The goal was to test if continuous authentication using single lead ECG data can effectively verify a user, thus preventing scenarios where another person might be tempted to produce activity data for someone else.
Methodology:
A database was created using ECG data from 33 healthy volunteers, accumulating 244 hours of recordings.
Data gathered while patients were at rest was considered, with data from 18 volunteers (with 10,000+ recorded heartbeats each) used for algorithm training.
The Quadratic Discriminant Analysis (QDA) classifier was employed due to its lower False Positive Rate, prioritizing intruder detection over sheer classification accuracy.
QRS detection identified individual heartbeats, while the Discrete Cosine Transform minimized the feature set for training.
The QDA classifier was trained to differentiate between the user’s heartbeat and the general populace, using accelerometer data to discard heartbeats during excessive movement.
Online verification involved matching heartbeats against a pre-set threshold to confirm or deny user authenticity.
Results: The algorithm was effective, particularly in identifying unauthorized users. However, it sacrificed accuracy for some enrolled subjects, prioritizing the avoidance of false positives. Excluding heartbeat data during active phases improved accuracy. A discrepancy was observed in performance across the subjects, indicating that a more comprehensive training dataset might be necessary for consistent positive identification.
In conclusion, ECG data shows promise for continuous user authentication in clinical trials. While current results are promising, there’s scope for refining the approach for better universal applicability.
