Sabotage Detection Using DL Models on EEG Data From a Cognitive-Motor Integration Task.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Concussion rehabilitation requires objective clinical tools to accurately assess patient effort.
• Cognitive-motor integration (CMI) tasks are valuable but susceptible to performance variability.
• Detecting willful under-performance (sabotage) is crucial for reliable assessment.

Purpose of the Study:

• To develop and validate a deep learning (DL) method for distinguishing between best effort and sabotage using electroencephalogram (EEG) signals.
• To analyze cortical activity differences during CMI tasks under varying effort conditions.
• To assess the efficacy of a CNN-LSTM model with self-attention for EEG-based performance classification.

Main Methods:

• Acquired four-channel EEG signals during CMI tasks using a wearable headband.
• Participants completed both sabotage and no-sabotage conditions in a randomized order.
• Employed a multi-channel convolutional neural network with long short-term memory (CNN-LSTM) model with self-attention, transforming time-series EEG data into 2D scalogram representations for classification.

Main Results:

• The DL model achieved high discrimination with 98.71% intra-subject accuracy and low false-positive rates.
• Average intra-subject accuracy was 92.8%, and average inter-subject accuracy was 86.15%.
• Scalogram-based analysis outperformed raw time-series analysis for performance classification.

Conclusions:

• The proposed DL method effectively detects differences in cortical activity between genuine and sabotaged effort during CMI tasks.
• This technique shows significant potential for enhancing the objectivity and reliability of concussion assessment and rehabilitation.
• Applications extend to baseline testing, injury status evaluation, recovery tracking, and industrial performance monitoring.