Impact of dataset size and long-term ECoG-based BCI usage on deep learning decoders performance

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCI) require large datasets for optimal performance, but data acquisition is costly and time-consuming.
• The non-stationarity of neuronal signals poses challenges for consistent BCI decoding over time.
• Understanding the impact of long-term recordings on BCI performance is crucial for clinical applications.

Purpose of the Study:

• To investigate the effect of long-term recordings on motor imagery decoding in BCIs.
• To compare the dataset size requirements of deep learning (DL) models versus traditional models.
• To explore patient adaptation and its influence on decoding performance in long-term BCI studies.

Main Methods:

• Evaluated multilinear and two deep learning models using ECoG data from a tetraplegic patient in a long-term BCI clinical trial (NCT02550522).
• Conducted computational experiments by increasing and translating training dataset sizes to assess model performance.
• Utilized UMAP embeddings and local intrinsic dimensionality for data visualization and quality evaluation.

Main Results:

• Deep learning decoders achieved higher decoding performance than the multilinear model, with similar dataset size requirements.
• High decoding accuracy was achieved with smaller datasets recorded later in the experiment.
• Evidence suggests improvement in motor imagery patterns and patient adaptation over the course of the long-term study.

Conclusions:

• Deep learning decoding is a promising approach for BCIs, effective even with real-world dataset sizes.
• Patient-decoder co-adaptation is a critical factor for improving performance in long-term clinical BCI applications.