Brain-machine interfaces (BMIs) offer transformative potential for restoring lost motor functions and treating neurological disorders such as epilepsy and Parkinson's disease. BMIs typically record electrical signals from the brain using neural probes with an increasing number of recording sites (hundreds or more). These signals are then processed to generate control commands for external effectors like prosthetic limbs. Most existing BMIs rely on silicon-based CMOS technology and the von Neumann architecture, which separates memory and computation units. This architecture necessitates the conversion of analog neural signals into digital signals, followed by compression and digital processing using ASICs. While this approach has yielded successful demonstrations, it faces challenges related to power consumption, processing delays, and scalability, especially as the density of recording sites increases exponentially. The conventional approach also differs fundamentally from the brain's analog and continuous information processing, leading to potential information loss and reduced accuracy due to conversion and compression. This research seeks to overcome these limitations by exploring bio-inspired solutions using memristors, whose inherent properties mimic biological synapses and neurons.

Prior research extensively explores the use of CMOS-based systems for BMI signal processing, showcasing advancements in signal acquisition, compression, and decoding. However, limitations in power consumption and scalability remain prominent challenges for high-density neural recordings. Various studies have demonstrated the use of application-specific integrated circuits (ASICs) for efficient neural signal processing, yet the power and delay associated with these systems are still significant hurdles. Meanwhile, the field of bio-inspired electronics has seen increasing interest, with studies investigating memristors as potential components in neuromorphic computing systems. Memristors' inherent non-volatility and ability to perform in-memory computing make them promising candidates for reducing power consumption and improving the speed of computations. Several research groups have demonstrated the use of memristor crossbar arrays for various computing tasks, including analog signal processing; however, their application to BMI signal analysis remains largely unexplored.

This study proposes a memristor-based neural signal analysis system for next-generation BMIs. The system uses memristor arrays to implement both signal preprocessing (using a long-tap FIR filter bank) and decoding (using a perceptron neural network). The researchers employed a TiN/HfO_x/TaO_y/TiN memristor array in a 1T1R cell structure. The memristors exhibited excellent bidirectional analog switching behavior and linear current-voltage (I-V) characteristics, crucial for accurate analog signal processing. The Bonn Epilepsy Dataset, containing local field potential (LFP) signals recorded from individuals in normal, interictal, and ictal states, served as the dataset. The FIR filter bank, consisting of four band-pass filters (delta, theta, alpha, beta bands), was implemented in one memristor array, with filter coefficients mapped onto the array as device conductance values. The filtered signals were then used to extract biomarkers (amplitude and energy in each frequency band), which were fed into a single-layer perceptron neural network implemented in another memristor array. The weights of this neural network were trained offline and mapped onto the memristor array. The system's performance was evaluated by comparing the results of the memristor-based system against software simulations. Power efficiency was compared against state-of-the-art CMOS systems by estimating the power consumption of both systems.

The memristor-based system achieved a high accuracy of approximately 93.46% in identifying epilepsy-related brain states (normal, interictal, ictal) from the LFP signals. The accuracy obtained with the memristor array-based system was comparable to the software simulations, demonstrating the system's ability to effectively process analog neural signals. Importantly, the memristor-based system showed a remarkable improvement in power efficiency, achieving a power efficiency of 1.4 µW/class compared to an estimated 551.0 µW/class for a comparable CMOS-based system. This represents approximately a 400x improvement in power efficiency. The study carefully analyzed the error between software-calculated and memristor array-filtered results, attributing small discrepancies to non-ideal device characteristics. The system's robustness to noise was also demonstrated, as the high accuracy was achieved despite inherent noise in the dataset and non-ideal characteristics in the memristor array. The researchers' analysis demonstrated that the memristor array-based filter bank retained sufficient information for accurate identification of brain states. The conductance map of the memristor arrays used for both filter bank implementation and neural network implementation were also provided.

The results demonstrate the feasibility and advantages of using memristors for high-efficiency neural signal analysis in BMIs. The high accuracy achieved in identifying epilepsy-related brain states, coupled with the significant power efficiency improvement, addresses crucial limitations of conventional CMOS-based BMI systems. The memristor-based approach's bio-plausibility and potential for in-situ analog processing align well with the brain's inherent processing mechanisms. The demonstrated 400x power efficiency improvement is particularly significant for the development of implantable BMIs, where power consumption is a major constraint. The research successfully bridges the gap between biological and artificial neural networks, paving the way for more energy-efficient and scalable BMIs.

This study presented a memristor-based neural signal analysis system that exhibits high accuracy and significantly improved power efficiency for BMI applications. The system's successful implementation using memristor arrays for filtering and identification of epilepsy-related neural signals shows great promise for next-generation BMIs. Future work will focus on integrating this memristor-based system with state-of-the-art neural probes to create a fully implantable BMI capable of processing high-density neural recordings. Further device optimization and advanced training strategies could further enhance the system's performance and robustness.

While the study demonstrates a significant improvement in power efficiency, the achieved accuracy in identifying brain states using the memristor-based system was slightly lower than that of software simulations. This minor decrease in accuracy is likely attributable to the non-ideal characteristics of the memristor devices used. The study also focused on a specific application (epilepsy detection), and further research is needed to determine the system's generalizability to other BMI applications and different types of neural signals. Finally, although the impact of noise was addressed, further investigation might explore the system's resilience under even more challenging noisy conditions.