Processamento de biossinais com métodos de aprendizagem de máquina em framework de código aberto

Abstract

Signal processing plays a fundamental role in the biomedical field, enabling the analysis and interpretation of biosignals for various applications. This project aims to utilize and validate a biomedical applications framework for the detection of manual gestures through biosignals, with the objective of assessing its functionality in problems related to this field. This framework performs the entire biosignal processing through the configuration of a single JavaScript Object Notation (JSON) file, in a flexible and easy-to-understand manner, facilitating its use by professionals in the healthcare and biomedical fields. Initially, publicly available Electroencephalography (EEG) data from the Grasp-and-Lift EEG Detection dataset, hosted on the Kaggle data science competition platform, were used. These data were classified into six different gestures representing actions performed by the right hand when grasping and releasing different objects (movement initiation, touch, force onset, lift, replace on the support, and release). Subsequently, using the Cyton board from OpenBCI, Surface Electromyography (sEMG) signals were collected from the forearm of volunteers during the execution of manual gestures, with the vowels of the Brazilian Sign Language (Libras) chosen as representative gestures. Both EEG and sEMG signals were filtered and segmented using windowing techniques. In the sEMG signals, the extracted features included Mean Absolute Value (MAV), Slope Sign Change (SSC), Waveform Length (WL), and Zero Crossing (ZC), while the EEG signals underwent feature extraction using the Common Spatial Patterns (CSP) method. Additionally, EEG signals were smoothed through convolution. Gesture classification was performed using Linear Discriminant Analysis (LDA) for both signal types, with all these processes integrated into the framework. The classification of EEG data achieved an average Area Under Curve (AUC) of 0.689, a result close to that obtained by similar works that used the same dataset. For gesture classification via sEMG, the average accuracy was 0.658, with some individuals achieving averages above 0.800, while others scored below 0.600, highlighting significant variability among participants. The study validates the effectiveness of the tool in biosignal processing and analysis, emphasizing its potential for applications in areas such as motor rehabilitation, brain-computer interfaces, and assistive devices.