The more, the better? Evaluating the role of EEG preprocessing for deep learning applications. DOI Creative Commons
Federico Del Pup, Andrea Zanola, Louis Fabrice Tshimanga

et al.

IEEE Transactions on Neural Systems and Rehabilitation Engineering, Journal Year: 2025, Volume and Issue: 33, P. 1061 - 1070

Published: Jan. 1, 2025

The last decade has witnessed a notable surge in deep learning applications for electroencephalography (EEG) data analysis, showing promising improvements over conventional statistical techniques. However, models can underperform if trained with bad processed data. Preprocessing is crucial EEG yet there no consensus on the optimal strategies scenarios, leading to uncertainty about extent of preprocessing required results. This study first thoroughly investigate effects applications, drafting guidelines future research. It evaluates varying levels, from raw and minimally filtered complex pipelines automated artifact removal algorithms. Six classification tasks (eye blinking, motor imagery, Parkinson's, Alzheimer's disease, sleep deprivation, episode psychosis) four established architectures were considered evaluation. analysis 4800 revealed differences between at intra-task level each model inter-task largest model. Models consistently performed poorly, always ranking average scores. In addition, seem benefit more minimal without handling methods. These findings suggest that artifacts may affect performance generalizability neural networks.

Language: Английский

The more, the better? Evaluating the role of EEG preprocessing for deep learning applications. DOI Creative Commons
Federico Del Pup, Andrea Zanola, Louis Fabrice Tshimanga

et al.

IEEE Transactions on Neural Systems and Rehabilitation Engineering, Journal Year: 2025, Volume and Issue: 33, P. 1061 - 1070

Published: Jan. 1, 2025

The last decade has witnessed a notable surge in deep learning applications for electroencephalography (EEG) data analysis, showing promising improvements over conventional statistical techniques. However, models can underperform if trained with bad processed data. Preprocessing is crucial EEG yet there no consensus on the optimal strategies scenarios, leading to uncertainty about extent of preprocessing required results. This study first thoroughly investigate effects applications, drafting guidelines future research. It evaluates varying levels, from raw and minimally filtered complex pipelines automated artifact removal algorithms. Six classification tasks (eye blinking, motor imagery, Parkinson's, Alzheimer's disease, sleep deprivation, episode psychosis) four established architectures were considered evaluation. analysis 4800 revealed differences between at intra-task level each model inter-task largest model. Models consistently performed poorly, always ranking average scores. In addition, seem benefit more minimal without handling methods. These findings suggest that artifacts may affect performance generalizability neural networks.

Language: Английский

Citations

1