Innovative Approach to Chronic Back Pain Classification

Biosignals play an important role in understanding how our bodies work without needing invasive tests. Two types of biosignals that are particularly useful are Surface Electromyography ([SEMG](/en/keywords/surface-electromyography--kkglv5d)) and data from Inertial Measurement Units (IMUs). sEMG looks at the electrical signals from our muscles, while IMUs measure how our body moves. Together, they give a detailed view of how we move and can help in managing conditions like chronic pain.

Chronic lower back pain (CLBP) affects many people around the world, lowering their quality of life and leading to economic costs. Proper detection and monitoring of CLBP are essential for creating effective treatment plans. However, analyzing the biosignal data to achieve this goal can be tricky. One of the main issues is that signals can be mixed or contaminated with noise from movement and other body processes. Additionally, differences between individuals can lead to variations in how these signals appear, making it harder to analyze them. The data from sEMG is complex, as it can include multiple channels and change over time. IMU data also has its own set of challenges, like sensor drift and errors in movement tracking.

Traditionally, researchers have used manual methods to extract features from these complex signals and apply filters to clean them up. Although this can work, it often requires specialized knowledge and may not effectively capture the complexities of the signals. Recently, machine learning methods, especially deep learning, have shown promise. These methods can automatically learn features from raw data, but they usually need large amounts of high-quality data and can require considerable computational power.

Spiking Neural Networks (SNNs) present a new opportunity for analyzing biosignal data. SNNs are inspired by how our brains work, as they process information through discrete spikes, like how neurons communicate in the brain. This feature of SNNs makes them particularly relevant for analyzing the complex signals from sEMG and IMUs. Furthermore, SNNs can be run on specialized hardware that uses much less energy, making them suitable for portable devices like wearables.

Nevertheless, one of the challenges with using SNNs is converting continuous signals into binary spikes, which is the format SNNs need to work. Various methods exist to achieve this conversion, but many struggle to maintain the key timing details of the original signals and can be sensitive to noise.

#Purpose of the Study

This study introduces a new approach to classify CLBP using the EmoPain database, which includes biosignal data from both healthy individuals and those suffering from CLBP. This research focuses on developing a method that effectively converts continuous sEMG and IMU signals into spike trains while preserving important timing information.

The main contributions of this work are:

1. Spike Threshold Adaptive Learning (STAL): This is a learnable encoder that converts continuous sEMG and IMU signals into spike trains while maintaining the important timing details of the signals.

2. Ensemble of Spiking Recurrent Neural Network (SRNN) Classifiers: This method combines multiple SRNN models that analyze the spike trains, enhancing accuracy and reliability in classifying individuals into healthy or CLBP categories.

By merging the biological inspiration of SNNs with the proven effectiveness of ensemble learning strategies, this approach provides new opportunities for developing better rehabilitation techniques for those with CLBP.

#Literature Review

The potential of SNNs in analyzing biosignals has been recognized due to their ability to mimic biological processes and consume less power than traditional deep learning methods. However, their application in detecting chronic pain and understanding pain-related behaviors remains largely unexplored.

A key aspect of SNN operation is converting continuous biosignals into discrete spikes, a process called spike encoding. This can be challenging for data obtained from EMG and IMU sensors, which can change rapidly and be affected by noise. Various methods for encoding spikes exist, each with its own strengths and weaknesses.

Recent studies have focused on creating learnable encoders that adapt to specific signals for analyzing biosignals with SNNs. However, many of these methods struggle with maintaining high temporal resolution and noise resistance, which are crucial for real-world applications in CLBP management.

#Proposed Method

The proposed method for classifying CLBP using biosignals employs a two-stage approach. First, the STAL encoder processes the biosignals (sEMG, Joint Angle, Joint Energy) and converts them into spike trains. Next, independent SRNNs analyze these spike trains within a multi-stream classification framework. This framework combines the predictions from each SRNN to improve classification results.

The STAL encoder consists of two main modules: feature extraction and feature-to-spike conversion. The feature extraction module captures important aspects of the input signals, while the feature-to-spike conversion module translates these features into spike activity.

To ensure the output from the STAL encoder is suitable for the SRNN, adjustments are made to align its dimensions. The encoder also includes a specially designed loss function that helps it learn efficiently while managing sparsity in the generated spike trains.

The SRNN takes the encoded spike trains and processes them through layers of recurrent neurons, allowing it to capture the complex timing patterns in the spikes. This ensures that the classification takes into account the temporal dynamics of the input data.

#Experimental Setup and Dataset

The EmoPain dataset serves as the source of biosignal data, which was collected from participants during different exercises. This dataset includes recordings from both healthy individuals and those experiencing CLBP, providing a rich source of information for analysis. Participants performed various physical tasks, allowing for the collection of both sEMG and IMU data.

Data preparation involved normalizing the signals to ensure consistency, as well as handling variations in recording lengths. The final data input for the models was organized into manageable segments to capture both short-term and longer-term patterns.

#Results

The performance of the proposed method was evaluated using various classification metrics, including accuracy and F1 score. The results indicated that the STAL-SRNN approach provided competitive performance compared to existing models. Specifically, it achieved favorable results in identifying individuals with CLBP and distinguishing between healthy and affected subjects effectively.

An analysis of the encoding methods confirmed that STAL outperformed traditional spike encoding techniques, showcasing its ability to maintain critical information during the conversion process. Additionally, an ablation study highlighted the importance of the feature extraction blocks, emphasizing that their presence improved the model's performance significantly.

#Conclusion

This study marks a significant advancement in the application of SNNs for classifying chronic lower back pain using biosignals. The introduction of the Spike Threshold Adaptive Learning encoder allows for better performance in converting continuous signals into meaningful spike trains. Additionally, using an ensemble of Spiking Recurrent Neural Network classifiers enriches the overall classification process by leveraging the strengths of each data modality.

The results suggest that this approach can effectively manage the challenges presented by biosignal analysis, leading to improved classification outcomes and potentially better pain management strategies. Future work may expand on these findings by applying the framework to other health conditions or exploring its use in real-time monitoring applications.