Physiological signals, such as electrocardiograms (ECGs) and electroencephalograms (EEGs), offer valuable insights into human health and cognition. Analyzing these signals for anomaly detection is essential for diagnosing conditions like arrhythmias and epileptic seizures. Traditional physiological signal processing systems rely on analog-to-digital conversion (ADC), digital processing, and memory, leading to high power consumption and latency due to the von Neumann bottleneck. Neuromorphic computing, inspired by the human brain's parallel and event-driven processing, presents a promising alternative. Memristors, with their unique electrical characteristics resembling biological neurons and synapses, offer a compelling platform for building energy-efficient neuromorphic systems. While some CMOS-based neuromorphic systems exist, they often suffer from area and energy inefficiencies. This research aims to address these limitations by developing a complete memristor-based neuromorphic physiological signal processing system, focusing on efficient spike encoding and biologically plausible neural network architectures, specifically incorporating ALIF neurons for enhanced computational capabilities. The system's performance is evaluated through arrhythmia classification and epileptic seizure detection tasks, demonstrating its potential for next-generation human-machine interfaces.

Existing literature highlights the growing interest in using deep learning and physiological signals for healthcare applications. Studies have demonstrated the efficacy of various machine learning techniques for heartbeat classification and epileptic seizure detection. However, these methods often rely on resource-intensive digital computing, limiting their applicability in portable and low-power devices. Previous neuromorphic approaches using CMOS technology have shown promise but struggle with size and energy efficiency due to complex circuitry and bulky capacitors. Research into memristor-based neuromorphic systems has showcased the potential for compact and energy-efficient implementations of LIF neurons, but implementing ALIF neurons and integrating them into a comprehensive physiological signal processing system remains largely unexplored. This work directly addresses this gap by developing a fully functional system incorporating both LIF and ALIF neurons for improved performance.

This study presents a VO2 memristor-based neuromorphic physiological signal processing system integrating an asynchronous spike encoder and an LSNN-based decision system. The VO2 memristors, fabricated using pulsed-laser deposition and electron beam lithography, exhibit volatile and symmetric threshold switching characteristics. The asynchronous spike encoder converts analog physiological signals into sparse spike trains representing signal changes beyond a fixed threshold. The encoder design is inspired by LC-ADCs and delta modulators, minimizing circuit complexity and power consumption. LIF and ALIF neurons were designed using the VO2 memristors, with the ALIF neuron incorporating an adaptive control circuit for adjusting membrane leakage current. An improved SPICE model for the VO2 memristor was developed for accurate circuit simulations, accounting for the symmetric threshold switching behavior. The LSNN architecture comprises an input layer, a hidden recurrent layer (using both LIF and ALIF neurons), a low-pass filter, and an output classification layer. Backpropagation through time (BPTT) with a surrogate gradient was employed to train the LSNN. The system's performance was evaluated on the MIT-BIH arrhythmia database for heartbeat classification (4 classes) and the CHB-MIT scalp EEG database for epileptic seizure detection (2 classes). For epileptic seizure detection, a post-processing step (moving average and thresholding) was applied to enhance performance, particularly specificity, due to the imbalanced nature of the dataset. Performance metrics included accuracy, sensitivity, specificity, and G-mean.

The proposed system demonstrated high accuracy in both arrhythmia classification and epileptic seizure detection. In arrhythmia classification, the system achieved a maximum accuracy of 95.83% using a 3 × 100 × 4 LSNN (3 input nodes, 100 hidden neurons, 4 output nodes). A comparative analysis revealed that incorporating ALIF neurons significantly improved accuracy compared to using only LIF neurons. The ALIF neurons’ adaptive property, enhancing the LSNN's temporal processing capabilities, played a key role in this performance. Epileptic seizure detection achieved an impressive 99.79% accuracy after post-processing, with 100% sensitivity and 99.89% specificity, utilizing a 37 × 40 × 2 LSNN (37 input nodes, 40 hidden neurons, 2 output nodes). Again, the inclusion of ALIF neurons significantly improved performance compared to systems using only LIF neurons. This system also required significantly fewer weights compared to state-of-the-art methods. The VO2 memristor-based encoder and neurons showed significant area advantages over existing CMOS or memristor implementations. The compact LIF and ALIF neurons achieved areas of ~41.3 μm² and ~53.4 μm², respectively, while the encoder reached an area of ~2231 μm², considerably smaller than existing encoders.

The high accuracy achieved by this memristor-based neuromorphic system in both arrhythmia classification and epileptic seizure detection showcases its potential for real-world applications. The use of ALIF neurons significantly improved performance, demonstrating the benefit of incorporating biologically plausible adaptive mechanisms in neuromorphic architectures. The system's compact design and low power consumption are particularly advantageous for wearable and implantable medical devices. The system's success underscores the potential of memristors for building efficient and compact neuromorphic hardware. Future work could explore the integration of non-volatile memory with the VO2 memristor-based system, further improving energy efficiency and enabling larger-scale networks for more complex physiological signal analysis. Optimizing the system’s parameters, such as the ALIF neuron's adaptive time constant, could further improve performance and robustness.

This research introduces a novel VO2 memristor-based neuromorphic physiological signal processing system for next-generation human-machine interfaces. The system demonstrates high accuracy and efficiency in arrhythmia classification and epileptic seizure detection tasks. The integration of ALIF neurons substantially improves performance. The compact design and efficient use of memristors pave the way for low-power, wearable, and implantable medical devices capable of real-time physiological signal analysis. Future studies should focus on optimizing the system's architecture and integrating it with non-volatile memory for larger-scale applications.

While this study demonstrates excellent results, limitations exist. The dataset size, especially for the epileptic seizure detection task, could be expanded to further improve the model’s generalizability. The post-processing step used in the epileptic seizure detection, while enhancing performance, introduces an additional step that might not be directly applicable in all real-time scenarios. Further research could explore alternative post-processing methods or integrate this step more directly into the network architecture. The current system uses a specific type of memristor. Investigating other memristor types or materials could broaden its applicability and potentially lead to further performance improvements.