Decorrelative network architecture for robust electrocardiogram classification.

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning Security
• Signal Processing

Background:

• Trustworthy AI is crucial for patient-critical medical applications.
• Current ensemble methods for uncertainty estimation are vulnerable to adversarial attacks.
• Shared vulnerabilities among ensemble models limit their robustness.

Purpose of the Study:

• To develop a robust ensemble approach for AI in medical tasks.
• To enhance AI's ability to recognize uncertainty and resist adversarial attacks.
• To improve the reliability of AI in patient-critical applications.

Main Methods:

• Proposed an ensemble method combining feature decorrelation and Fourier partitioning.
• Trained networks to learn diverse features, reducing susceptibility to perturbations.
• Tested against white-box adversarial attacks on electrocardiogram (ECG) data.
• Adapted adversarial training and DVERGE for ensemble comparison.

Main Results:

• The decorrelation and Fourier partitioning ensemble maintained performance on unperturbed data.
• Demonstrated superior uncertainty estimation against projected gradient descent and smooth adversarial attacks.
• The approach showed robustness against adversarial attacks of varying magnitudes.
• No expensive adversarial sample optimization was required during training.

Conclusions:

• The proposed ensemble method enhances AI robustness and uncertainty estimation in medical tasks.
• Feature decorrelation and Fourier partitioning effectively reduce vulnerability to adversarial attacks.
• This approach offers a more reliable and efficient way to build trustworthy AI for healthcare.
• The methods are applicable to other domains requiring robust AI models.