COMPARATIVE ANALYSIS OF SPIKING NEURAL NETWORK AND ARTIFICIAL NEURAL NETWORK MODELS IN PREDICTING FINGERTIP KINEMATICS USING INTRACRANIAL BRAIN SIGNALS

• Main Author: Morenzo Muten, Eraraya
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86167
• Institution: Institut Teknologi Bandung

According to the World Health Organization (WHO), each year, 250,000 to 500,000 people worldwide experience spinal cord injury (SCI). Among them, more than half suffer from tetraplegia, a condition in which patients lose the ability to move all four limbs, including arms and legs. Without adequate recovery, tetraplegic patients have a lower life expectancy and face higher medical costs compared to other types of paralysis. Based on various independent surveys conducted in multiple countries, the recovery of upper limb motor functions, especially hand functions, is considered the top priority for tetraplegic patients. The ability to use their hands again is a key factor determining their quality of life. The recovery of hand motor functions can be achieved through various methods, ranging from non-operative rehabilitation with motor therapy and reconstructive surgery to prosthetics controlled by a brain-machine interface (BMI). With technological advancements, intracranial BMI solutions, or those placed within the skull, have garnered increasing attention from researchers due to their ability to be personalized and to record neural activity with higher resolution and quality. In this context, predicting hand movements based on brain signals can be accomplished through neural pattern recognition systems or decoders based on artificial neural network (ANN) models. ANN-based decoders have been shown to achieve higher prediction accuracy compared to conventional decoders such as Wiener filters and Kalman filters. However, the main challenges of ANN-based decoders lie in their size, complexity, and power consumption, which exceed the limitations acceptable for brain implants. As an alternative, decoders based on spiking neural network (SNN) models, often referred to as the third generation of neural networks, offer a more efficient solution regarding computational resource usage without compromising accuracy. SNN- based decoders leverage neuromorphic algorithms, which have been applied in various medical fields such as image processing, epilepsy seizure detection, and computer cursor movement control. These decoders are not only more energy- iv efficient but also capable of operating in real-time, making them an ideal choice for critical medical applications. This study proposes an optimization scheme and training for SNN-based decoders, specifically the Spiking Multilayer Perceptron (SMLP) and Spiking Long Short- Term Memory (SLSTM), to predict fingertip velocity based on spike activity from intracranial brain signals in rhesus monkeys (Macaca mulatta). These SNN-based decoders were then compared to ANN-based decoders that were retrained based on previous research. The comparison was conducted using two different input processing methods: the binning method, commonly used for ANN-based decoders, and the streaming method, which is more suitable for SNN-based decoders. The comparison involved various metrics, including accuracy measured by Pearson correlation coefficient (CC) and coefficient of determination (R2), memory footprint size in kilobytes (kB), complexity in terms of the number of synaptic operations, and prediction time. This study showed that the combination of the streaming input processing method with the SMLP decoder yielded the highest accuracy compared to other decoder variations using real-time input schemes. This decoder demonstrated a 16.86% improvement in CC and a 33.33% increase in R2 compared to the Binning-MLP decoder. Additionally, the SMLP decoder was found to be 68.83 times smaller in size and required 71.84 times fewer synaptic operations. The prediction time of the Streaming-SMLP decoder is lower than the sampling frequency of the fingertip movement sensor, allowing for real-time application. Communication between units in the Streaming-SMLP decoder involved only binary activations, which requires minimal computational resources. Implementing the Streaming-SMLP decoder on neuromorphic hardware could highlight the advantages of SNN-based decoders. In conclusion, this study underscores the significant potential of SNN-based decoders as a more efficient solution for predicting hand movements in tetraplegic patients, offering new hope in more advanced and personalized motor recovery technologies. Furthermore, this research provides a comparative analysis between SNN-based and ANN-based decoders, which could serve as a critical foundation for further development of both decoder types. This analysis offers deep insights into the strengths and weaknesses of each approach, as well as opportunities to integrate their advantages in future clinical applications.