A framework for Interpretable deep learning in cross-subject detection of event-related potentials

• Published in: Engineering applications of artificial intelligence Vol. 139; p. 109642
• Main Authors: Jalilpour, Shayan, Müller-Putz, Gernot
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.01.2025
• Subjects: Brain-computer interface; Deformable convolutions; Electroencephalography; Error-related potential; Mixed depthwise convolutions; Perturbation-evoked potential; Brain-computer interface; Deformable convolutions; Mixed depthwise convolutions; Electroencephalography; Error-related potential; Perturbation-evoked potential
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2024.109642

Event-related potential-based Brain-Computer Interfaces are becoming widely popular due to their ability to send commands quickly with high accuracy. However, the stationary characteristics of electroencephalographic signals, coupled with their low signal-to-noise ratio, lead to variations in amplitude, time period, and latency in the patterns of event-related potentials across different trials, sessions, days and subjects. Conventional feature extraction and machine learning algorithms are not designed to handle these differences, requiring the development of methods that can address these variations. Here, we propose a novel lightweight deep neural network for event-related potential classification, consisting of three modules. In this model, we have a spatio-temporal module that learns local features simultaneously across channels and time points. Following this, there's a component extractor module comprising depthwise convolutions, inspired by mixed depthwise convolutions, to capture the event-related potential characteristics with different temporal durations. Lastly, an advanced temporal layer addresses event-related potential shape and scale variations using deformable convolutions. We conducted experiments on event-related potential detection in a subject-independent scenario using one error-related negativity potential dataset and three perturbation-evoked potential datasets. Comparisons were made with established methods including two conventional machine learning algorithms and three well-known deep learning architectures, demonstrating that our model outperformed them in terms of classification accuracy and parameter efficiency. In our analysis, we aimed to understand the model's performance using gradient-weighted class activation mapping and t-distributed stochastic neighbor embedding. These methods facilitated the visualization and interpretation of our model's effectiveness, providing insights into its relationship with the neuroscientific characteristics of event-related potentials.