An effective ECG signal classification method based on a minimalistic memristive reservoir computing system.

Area of Science:

• Electronics
• Artificial Intelligence
• Signal Processing

Background:

• Reservoir computing (RC) excels at time series processing due to its simplified training, but hardware implementation of high-dimensional reservoir layers remains challenging.
• Memristors offer unique nonlinear and memory properties, suitable for mapping input data into high-dimensional feature spaces required by RC.
• Existing memristor-based RC systems utilize volatile memristors, posing limitations for practical applications.

Purpose of the Study:

• To design and implement a novel non-volatile memristor-based reservoir layer for constructing an RC system.
• To leverage memristor characteristics for nonlinear mapping of input signals into a high-dimensional feature space.
• To evaluate the performance of the proposed RC system in a real-world application, such as electrocardiogram (ECG) signal classification.

Main Methods:

• A non-volatile memristor-based reservoir layer was designed, utilizing the voltages across two memristors to compute reservoir states.
• A one-dimensional voltage input signal was nonlinearly mapped to a two-dimensional space, simplifying data analysis and enhancing feature separability.
• The proposed RC system was experimentally tested on an electrocardiogram (ECG) signal classification task.

Main Results:

• The non-volatile memristor-based reservoir layer successfully mapped 1D input signals to a 2D space, enhancing feature separability for RC requirements.
• The proposed RC system achieved high classification accuracies of 98.3% for shifted QRS complexes and 100% for unshifted QRS complexes in ECG signals.
• The experimental results validate the effectiveness of the non-volatile memristor-based RC system for complex time series data processing.

Conclusions:

• The developed non-volatile memristor-based reservoir layer offers a simplified and effective approach for high-dimensional feature mapping in RC systems.
• This novel design addresses the hardware implementation challenges of traditional RC systems, particularly for time series analysis.
• The high accuracy achieved in ECG signal classification demonstrates the potential of this memristor-based RC system for practical biomedical applications.