Single-subject auditory ERP-BCI performance enhancement in ALS via an AI coding assistant prompt

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Auditory event-related potential (ERP) brain-computer interfaces (BCIs) are crucial for communication in individuals with amyotrophic lateral sclerosis (ALS).
• Optimizing individual BCI pipelines is challenging, time-consuming, and limits performance, particularly in increasing selection speed while maintaining accuracy.

Purpose of the Study:

• To investigate if an AI coding assistant could optimize an auditory ERP-BCI for a single ALS patient.
• To assess if AI-driven optimization can enhance selection speed and classification accuracy.

Main Methods:

• A three-class auditory ERP-BCI was optimized using an AI coding assistant (Claude Code) for a single ALS patient.
• The AI generated and evaluated 23 optimization scripts over 24 hours with minimal human oversight, creating an AI-Designed ERP classifier (AIDE).
• The AIDE classifier was evaluated on 189 EEG trials using five cross-validation strategies.

Main Results:

• The AI-Designed ERP classifier (AIDE) achieved 85.03% mean cross-validation accuracy with a 17s selection time, doubling the information transfer rate (ITR) to 2.92 bits/min.
• AIDE prevented accuracy degradation when selection time was halved, unlike baseline models.
• Online testing showed AIDE achieved 66.7% accuracy compared to 50.0% for the baseline model.

Conclusions:

• Single-subject BCI performance can be significantly improved using a single AI prompt.
• This approach offers an efficient pathway for individualized BCI optimization in clinical and research settings.
• AI-driven optimization holds promise for enhancing communication support for individuals with severe neurological conditions like ALS.