Human nonverbal communication is challenging to translate for machine understanding. Improving AI's comprehension of the user's mental state during decision-making, especially in ambiguous situations, is crucial. This necessitates a system that directly integrates human brain signals into the AI's decision-making process. While Brain-Machine Interfaces (BMIs) exist, they often involve complex setups and lack seamless integration into real-world applications. Current EEG technology suffers limitations, including the need for multiple electrodes, bulky headsets, wired connections, limited long-term use, and testing primarily with monitors rather than autonomous machines. The ideal HMI should feature bidirectional interaction, allowing the AI to understand human emotions and modify its behavior accordingly. This research aims to address these limitations by introducing BACLoS, a wireless, user-friendly system that integrates lightweight, comfortable EEG sensors with AI algorithms to provide continuous real-time feedback based on the user's ErP signals. This approach avoids the conscious effort required by methods using hands, voice, or facial expressions and is vital in the rapidly evolving field of AI-powered machines.

The authors reviewed existing literature on bio-integrated wearable technologies, BMIs, and AI-human interaction. They highlight the limitations of current BMI systems, focusing on issues such as long-term measurement capabilities, electrode placement and comfort, wireless communication, and the integration with autonomous machines. Previous studies often used limited conditions, such as monitoring displays instead of interacting with autonomous systems. The review underscores the need for a practical and convenient BMI system that addresses these challenges.

The researchers developed a wireless earbud-like EEG device (eEEG) with tattoo-like electrodes and connectors. The eEEG is lightweight (8.24g) and designed for comfortable, long-term use, minimizing motion artifacts. The tattoo-like electrodes and connectors have low bending stiffness, enabling conformal contact with the skin and reducing impedance. The eEEG wirelessly transmits data to an AI system. This AI system uses deep learning to classify EEG signals, specifically focusing on identifying ErP patterns indicative of unexpected AI actions. Two main AI systems are implemented: an Emergency Interrupting System (EIS) that immediately stops or modifies AI actions upon detection of ErP, and a User-Customized Reinforcement System (UCRS) that uses ErP feedback as a reward/punishment signal to reinforce desired AI behaviors. The eEEG's performance was compared against a commercial EEG device (cEEG) to highlight the reduction in motion artifacts. The classification performance of deep learning models (DNNs and LSTMs) was compared with traditional machine learning models (LR, LDA, k-NN, RF, SVM). The BACLoS was tested with an autonomous driving RC car, a maze solver, and an assistive interface to demonstrate its functionality and effectiveness. The study involved multiple experiments with human subjects, ensuring ethical considerations through IRB approval and informed consent. The P300 ERP was measured using auditory oddball tasks to validate the system's ability to detect relevant brain signals.

The tattoo-like electrodes and connectors significantly reduced motion artifacts compared to a commercial EEG system. The eEEG device achieved low electrode-skin impedance without using electrolytes and maintained signal quality even during activities such as walking, cycling, and driving. Deep learning (LSTM) accurately classified ErP patterns in EEG signals with high accuracy (83.81%), outperforming traditional machine learning methods. The BACLoS successfully implemented both EIS and UCRS. EIS reduced braking time and distance compared to manual control in an RC car experiment. UCRS enabled the RC car to learn and adapt to user preferences in a maze-like environment. BACLoS also improved the efficiency of an autonomous maze solver by integrating user feedback, reducing completion time. The system demonstrated a capability for detecting P300 signals even during movement, showcasing its robustness and potential for various applications.

The BACLoS successfully integrates wearable EEG technology with AI for seamless and intuitive human-in-the-loop machine learning. The study's results demonstrate the feasibility of using ErP feedback to enhance autonomous machine decision-making. This approach significantly improves the interaction between humans and AI-based machines, paving the way for safer and more user-friendly systems. The improved performance over conventional methods, especially in dynamic conditions, opens doors for applications beyond those tested in this study. The successful real-time integration of deep learning classification directly into the AI decision-making process shows promising results for the future of human-AI collaboration.

This research successfully developed and demonstrated the BACLoS, a novel brain-AI closed-loop system using wearable EEG technology. The system's ability to detect ErP signals and use them to refine AI decisions was validated through various experiments. Future research could focus on improving electrode materials, implementing more sophisticated AI algorithms, and conducting larger-scale studies with a wider range of AI-powered machines, such as self-driving cars in realistic road conditions. The BACLoS shows promise for enhancing human-machine interaction and creating more personalized and adaptable AI systems.

The study used a small number of participants, limiting the generalizability of findings. The experiments were conducted in controlled environments, and further investigation in real-world scenarios is needed. While deep learning models demonstrated high accuracy, the robustness to noisy environments could be further optimized. The specific AI applications tested (RC car, maze solver, assistive interface) may not fully represent the complexities of all potential use cases. Finally, the long-term stability and reliability of the tattoo-like electrodes and the eEEG device require additional testing.