Machine Learning-Based Classification of Marathi Words for Brain-Computer Interface Applications

This paper presents a study on the application of Machine Learning (ML) methods for recognizing Marathi words using Brain-Computer Interface (BCI) system. The research aims to address communication challenges faced by Marathi-speaking patients with neurodegenerative disorders. Electroencephalograph (EEG) signals collected using an 8-channel EEG device underwent preprocessing, including filtering with Butterworth-low-pass and Moving Average filters. Timedomain features (Hjorth parameters and AMC Feature Set) were then extracted from each channel. Classification tasks were performed by using the Support Vector Machine (SVM) and Naive Bayes machine learning algorithms. The results include metrics such as average training accuracy, average testing accuracy, and evaluation metrics to provide a comprehensive understanding of the model's performance. For two-word classification tasks, Naive Bayes with Hjorth parameters yielded accuracy rates of 95.11% and 95% for Marathi vocal and sub-vocal datasets, respectively. Meanwhile, English vocal data achieved an accuracy of 90%. Classification of three-words with Naive Bayes utilizing the AMC feature set achieved testing accuracies of 70% and 63.33% for Marathi and English datasets, respectively. For English vocal data SVM (linear) with Hjorth parameters achieved testing accuracy of 73%. This study sets the foundation for creating a Marathi language BCI system to improve communication between caregivers and Marathi-speaking individuals with neuro-degenerative disorders.