RESUMO
Deep learning models have demonstrated remarkable performance in the classification of motor imagery BCI systems. However, these models exhibit sensitivity to challenging trials, often called hard trials, leading to performance degradation. In this paper, we address this issue by proposing two novel methods for identifying and mitigating the impact of hard trials on model performance. The first method leverages model prediction scores to discern hard trials. The second approach employs a quantitative explainable artificial intelligence (XAI) approach, enabling a more transparent and interpretable means of hard trial identification. The identified hard trials are removed from the entire motor imagery training and validation dataset, and the deep learning model is further re-trained using the dataset without hard trials. To evaluate the efficacy of these proposed methods, experiments were conducted on the Open BMI dataset. The results for hold-out analysis show that the proposed quantitative XAI- based hard trial removal method has statistically improved the average classification accuracy of the baseline deep CNN model from 63.77% to 68.70%, withp-value =7.66e-11for the subject-specific MI classification. Additionally, analyzing the scalp map representing the average relevance scores of correctly classified trials compared to a misclassified trial provides a deeper insight into identifying hard trials. The results indicate that the proposed quantitaive-based XAI approach outperformes the prediction-score-based approach in hard trial identification.
Assuntos
Algoritmos , Interfaces Cérebro-Computador , Aprendizado Profundo , Eletroencefalografia , Humanos , Eletroencefalografia/métodos , Imaginação , Inteligência Artificial , Redes Neurais de ComputaçãoRESUMO
Recently, deep learning and convolutional neural networks (CNNs) have reported several promising results in the classification of Motor Imagery (MI) using Electroencephalography (EEG). With the gaining popularity of CNN-based BCI, the challenges in deploying it in a real-world mobile and embedded device with limited computational and memory resources need to be explored. Towards this objective, we investigate the impact of the magnitude-based weight pruning technique to reduce the number of parameters of the pre-trained CNN-based classifier while maintaining its performance. We evaluated the proposed method on an open-source Korea University dataset which consists of 54 healthy subjects' EEG, recorded while performing right-and left-hand MI. Experimental results demonstrate that the subject-independent model can be maximumly pruned to 90% sparsity, with a compression ratio of 4.77× while retaining classification accuracy at 84.44% with minimal loss of 0.02% when compared to the baseline model's performance. Therefore, the proposed method can be used to design more compact deep CNN- based BCIs without compromising on their performance.