An Efficient and Private ECG Classification System Using Split and Semi-Supervised Learning

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Electrocardiography (ECG) is crucial for diagnosing cardiovascular diseases.
• Deep learning models show promise for ECG classification but require centralized data.
• Real-world ECG data is often decentralized and sparsely labeled, posing privacy and efficiency challenges.

Purpose of the Study:

• To develop a privacy-preserving and efficient ECG classification system.
• To address the limitations of centralized data and limited labels in deep learning for ECG analysis.
• To enhance system efficiency by reducing communication and client-side computation.

Main Methods:

• Analysis of deep learning models, identifying Temporal Convolutional Networks (TCN) as most efficient.
• Development of a modified split-learning (SL) system based on TCN.
• Implementation of semi-supervised learning to leverage unlabeled data.
• Testing on an Internet of Things (IoT) setup.

Main Results:

• The modified SL system reduced communication overhead by 71.7% and client computations by 46.5% compared to basic SL with TCN.
• Semi-supervised learning improved classification performance by 9.1%-15.7% with only 10% labeled data.
• Satisfactory classification accuracy was achieved in a private and energy-efficient manner on the IoT test setup.

Conclusions:

• The proposed system offers an efficient and private solution for ECG classification.
• The integration of modified split-learning and semi-supervised learning effectively handles decentralized and limited labeled data.
• The system is well-suited for green AI applications, particularly in resource-constrained IoT environments.