Transfer Learning With Active Sampling for Rapid Training and Calibration in BCI-P300 Across Health States and...

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Machine learning and deep learning have advanced Brain-Computer Interface (BCI) performance.
• Wide-scale BCI applicability is hindered by individual health, hardware, and cultural variations in neural data.
• Current BCI studies often use uniform settings, limiting real-world generalizability.

Purpose of the Study:

• To enhance BCI classification accuracy using deep learning and transfer learning.
• To address challenges in adapting models to diverse, imbalanced datasets from varied sources (equipment, subjects, centers, populations).
• To propose a novel adaptive transfer learning method for P300 wave detection in highly heterogeneous BCI data.

Main Methods:

• Employed a convolutional neural network for P300 wave detection in BCIs.
• Introduced Active Sampling (AS), an adaptive transfer learning method based on Poison Sampling Disk (PDS).
• AS flexibly adjusts the transition from source to target domains for model adaptation.

Main Results:

• Achieved a 5.36% average improvement in classification accuracy with subject-adaptive fine-tuning (40%).
• Reduced classification accuracy standard deviation by 12.22% across two distinct, internationally replicated datasets.
• Outperformed existing methods in classification accuracy, computational time, and training efficiency.

Conclusions:

• The proposed Active Sampling (AS) method significantly improves BCI performance in heterogeneous settings.
• AS enhances model generalizability and reduces overfitting, crucial for real-world BCI applications.
• This approach offers a promising solution for effective model transfer and tuning in diverse BCI data scenarios.