Deep learning models for electrocardiograms are susceptible to adversarial attack.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Device Security

Background:

• Electrocardiogram (ECG) analysis is crucial for diagnosing heart rhythm irregularities.
• Deep neural networks (DNNs) show promise in automating ECG interpretation, sometimes exceeding human accuracy.
• DNNs are vulnerable to adversarial examples, subtle data manipulations designed to cause misclassification.

Purpose of the Study:

• To develop a method for creating physiologically plausible adversarial examples for ECG signals.
• To assess the vulnerability of deep learning models for arrhythmia detection to these novel adversarial examples.
• To establish a generalizable technique for generating diverse adversarial ECG examples.

Main Methods:

• Developed a novel method to generate smoothed adversarial examples for ECG tracings, ensuring they are undetectable to human experts.
• Tested the vulnerability of a deep learning model trained for single-lead ECG arrhythmia detection.
• Created a technique to collate and perturb existing adversarial examples to generate new ones.

Main Results:

• Demonstrated that smoothed adversarial examples can fool a deep learning model for ECG arrhythmia detection.
• Showed that these adversarial examples are imperceptible to human evaluation.
• Successfully generated multiple new adversarial examples using the developed collating and perturbing technique.

Conclusions:

• Deep learning models for ECG analysis are susceptible to sophisticated adversarial attacks.
• Careful evaluation of AI diagnostic tools is necessary, especially in scenarios where data manipulation is possible.
• The findings underscore the need for robust security measures in AI-driven medical diagnostics to prevent misclassification.