1. Introduction

The Brain-Computer Interface (BCI) technology, particularly non-invasive technologies, has taken center stage in innovation due to the fact that it can transform human-computer interaction (HCI). It incorporates three main steps to link human brain signals to direct computer commands: acquiring neural signals from the brain, decoding the signals to obtain simple intentions, and analyzing these intentions to translate them into different computer output forms. Unlike invasive BCIs that require surgical implants, non-invasive BCIs employ external sensors, such as EEG or fNIRS, to record brain activity painlessly and without the need for surgeries. Such an approach makes BCIs more accessible by providing an alternative means for people with physical disabilities to interact with the world and also new ways for the general population to communicate with technology.The advancement of non-invasive brain-computer interfaces (BCIs) demonstrates significant potential to unveil novel opportunities in human-computer interaction (HCI) by improving user experience and introducing innovative paradigms of computer interaction. This essay explores the technological advancements in non-invasive BCI systems that shape HCI, focusing on key components such as signal acquisition, neural decoding models, and machine learning algorithm contributions to optimizing the performance of BCI. The purpose of this research is to examine how advancements in non-invasive BCI systems enhance HCI, while also exploring the theoretical foundations of the models and techniques that contribute to the accuracy and efficiency of BCI-driven interactions.

2. Signal Acquisition

2.1. EEG: The Most Widely Used Technique

Electroencephalography (EEG) has become the most widely used technique for non-invasive brain signal acquisition as it has high temporal resolutions and is supported by multiple sophisticated analytical methods [1]. Electroencephalography (EEG) employs electrodes affixed to the scalp to assess and document the brain's electrical activity produced by neuronal functions. These electrodes are capable of capturing voltage variations stemming from ionic movements within neurons, which can subsequently be analyzed to produce visual representations of brainwave patterns.

EEG demonstrates multiple advantages that promote efficient communications with electrical devices. Firstly, it offers exceptional temporal resolution on the order of milliseconds, which makes it ideal for capturing rapid brain activity such as motor commands, cognitive processing, and neural responses to stimuli. Such rapid capturing can be especially useful in the medical field. For instance, the detection of epileptic seizures requires accurate capturing and timing of electrical discharges, which can be effectively done by EEG [2]. The capability to monitor cerebral activity in near real-time through EEG facilitates advancements in Human-Computer Interaction, where prompt feedback frequently improves the efficacy of communication. Secondly, existing mathematical algorithms for decoding and analyzing EEG data are already mature, hence EEG can be applied to not only clinical settings, but also other areas such as autonomous control of robotics, rehabilitation, neuroimaging entertainment, neural science research, etc. [3].

2.2. Recent Advancement: Dry Electrodes

Despite EEG’s benefits and wide usage, the traditional method of EEG signal acquisition with gel-based electrodes comes with inconveniences. With conventional EEG systems, a conductive gel is generally required to be applied between the electrode and the scalp to ensure proper signal transmission. However, this conductive gel makes preparation tasks time-consuming and can cause discomfort for the user, especially during long time use. Frequent reapplication, skin irritation, and the messiness of the gel can deter individuals from using EEG systems for long periods of time or in non-clinical settings. These practical barriers reduce the usability of EEG in BCIs, where ease of use and comfort are crucial for promoting prolonged engagement and interaction with devices.

As a result, recent advancements in EEG technology have led to the development of dry electrodes, electrodes that do not need conductive gel. Instead, these electrodes are designed to make direct contact with the skin, so skin preparation or gel application is no longer needed. Dry electrodes typically work by using materials with high surface area or advanced conductive coatings to ensure low impedance between the electrode and the scalp. Liao et al. conducted a study that illustrates how a dry electrode technology, produced through an injection molding manufacturing process, can attain signal quality on par with conventional gel electrodes, thereby presenting a feasible alternative for practical applications. The research highlighted that dry electrodes not only reduced the discomfort associated with gel-based systems but also increased the efficiency of EEG signal acquisition, enabling quicker setup times and improving the overall user experience [4]. These electrodes are proven to be useful for applications that require quick setup and comfort, such as in mobile or consumer-grade BCIs, neurofeedback systems, and cognitive monitoring applications. This development also opens up opportunities for EEG systems to be used continuously over extended periods, such as in sleep studies, cognitive training, or real-time brain-computer interaction scenarios. The user-friendly nature and low maintenance demands of dry electrodes render them particularly suitable for consumer-oriented applications, enabling users to utilize them comfortably in domestic, professional, or mobile environments without compromising data integrity.

3. Neural Decoding Models

3.1. Preprocessing Techniques

Raw EEG signals obtained from the brain are inherently noisy and complex, since various components contributes to the data. Other than brain activity related to user intentions, EEG signals are also composed of artifacts introduced by external sources, such as eye movements, muscle contractions, and environmental noise. Therefore, it is crucial to employ preprocessing techniques to clean and filter the raw data before these neural signals can be further decoded. This essay will discuss several preprocessing techniques improving the quality of EEG signals, such as ICA, band-pass filtering, and baseline correction.

In a systematic evaluation of diverse preprocessing methodologies, Coelli et al. sought to identify the most efficient techniques for the purification of EEG signals prior to their application in neural decoding. Their study compared several noise reduction techniques and established that Independent Component Analysis (ICA) is still among the most efficient techniques for removing artifacts like eye blinks and muscle contractions. ICA is a blind source separation technique that decomposes a multivariate signal into statistically independent components. It is particularly effective at identifying and isolating artifacts by separating the neural activity from the noise. The authors highlighted ICA's flexibility in handling various types of artifacts and its widespread application in EEG-based BCI systems.

Other than ICA, band-pass filtering is another essential method that excludes irrelevant frequencies, focusing the analysis on the frequency bands that correspond to only the relevant brain activity. Coelli et al. emphasized the importance of choosing the correct frequency range, as each frequency band conveys distinct forms of cognitive and motor information. For example, filtering for alpha (8–13 Hz) and beta (13–30 Hz) bands is particularly useful in motor imagery or relaxation tasks, as these are the bands typically associated with brain activity during these states. Their findings suggest that band-pass filtering, combined with ICA, produces higher signal quality and better performance in decoding models [5].

Baseline correction is another preprocessing technique that is crucial for enhancing signal quality. This technique involves adjusting the EEG signals to a common reference point, typically the average signal across all channels, to remove any systematic variations that might occur during the recording session. This normalization process helps ensure that the recorded signals reflect the brain's true activity rather than being distorted by extraneous factors.

3.2. Decoding Models

After preprocessing raw neural signals, various decoding models can be applied based on the scenario and the type of data to translate signals into simple intentions. This essay will look at two frequently used models: Linear Discriminant Analysis and Kalman Filters.

EEG signal categorisation using Linear Discriminant Analysis (LDA) is popular. LDA is a supervised learning technique that finds a linear combination of features to distinguish data classes. LDA is commonly used in BCIs to classify brain states or intentions, such as picturing movement or focussing on a task. LDA presupposes that class data are normally distributed and feature covariance is the same. The method then projects the data onto a lower-dimensional space where the classes are most distinct to maximise separation. LDA is useful in BCI applications involving motor imagery or simple mental instructions because it can efficiently isolate brain activity associated with each intention based on frequency or spatial factors.

On the other hand, Kalman Filters are particularly useful in regression-based decoding tasks, where continuous outputs such as cursor movement or control of robotic limbs are required. Kalman Filters are recursive, state-space models that predict the current state of a system based on previous states and noisy observations. This method uses a two-step process: the prediction step, which estimates the next state of the system based on prior information, and the update step, which refines this estimate by incorporating the most recent measurements. Kalman Filters are especially effective in dynamic BCI systems where the user’s intentions evolve continuously over time, as they are able to incorporate both previous and current data to provide accurate predictions of the user’s movements. It significantly improved the real-time control of devices by reducing latency and improving the accuracy of movement predictions, even in the presence of noisy EEG signals [6].

4. Machine Learning Algorithms in Brain-Computer Interfaces

As AI has become the focal point of technology, Machine learning (ML) is also revolutionizing the field of BCI, acting as a powerful tool for classifying, predicting, and personalising neural signals. ML techniques have been shown exceptionally in translating large amounts of neural signals into meaningful information that can be used for complex intent detection, movement prediction, and adaptive BCI systems. The techniques that can be applied to EEG data include supervised / unsupervised learning, convolutional neural networks (CNN), recurrent neural networks (RNN), transformers, etc. They enable BCIs to learn complex hierarchical patterns from raw data and further improve the ability for computers to accurately understand and obtain human’s needs.

4.1. Popular ML Techniques in BCI

Supervised learning techniques, with common examples being support vector machines (SVM), linear discriminant analysis (LDA), and random forests, are applied to BCI systems for classifying brain activity into specific categories such as movement intentions and cognitive states. These algorithms rely on labeled training data, where the neural signals are paired with their corresponding intentions or actions. After training on the labeled dataset, the model can predict the user’s intentions based on new, unseen EEG data.

On the other hand, unsupervised learning techniques, such as k-means clustering or principal component analysis (PCA), are useful for extracting underlying patterns from unlabeled data, which can be very helpful in exploratory phases or when there is a lack of labeled data. These techniques can be employed to identify new brainwave patterns without manual labeling.

Both machine learning and deep learning methods like CNN have showed promise in improving BCI accuracy and efficiency in recent years. Convolutional neural networks (CNNs) use convolutional filters in both temporal and spectral domains to discover key brain electrical activity patterns without prior knowledge of feature relevance. They are therefore ideal for analysing image-like electroencephalogram (EEG) signals [7].

In addition to CNNs, recurrent neural networks (RNNs) have been explored for BCI applications. RNNs, which are particularly well-suited to sequential data processing, can learn temporal dependencies in EEG signals over time. This benefits tasks such as continuous motor movement prediction or cognitive state monitoring, where past brain activity influences future behavior. RNNs can capture such temporal structures, and hence, they are well suited to real-time BCIs in which ongoing updates to the user intention are required.

Transformers, a more recent development in deep learning, have also begun to show promise in the realm of EEG-based BCIs. Unlike RNNs, where data processing is sequential, transformers can concurrently process various parts of the input data, making it more efficient and better for long-range dependencies. This feature makes transformers particularly suitable for high-dimensional EEG data, where there could be many channels with complex interactions over time.

4.2. Example Applications of ML in BCI

This section of the essay will demonstrate a few successful examples of ML techniques and algorithms applied to BCI, enhancing human computer interactions.

Vernon Lawhern et al. created a compact CNN architecture for EEG-based BCIs that classifies motor imagery EEG data better. This CNN structure learns features from raw EEG data, unlike typical classification algorithms that need considerable feature extraction and operator tweaking. Their research describes it as a lightweight CNN with few convolutional layers that preserves EEG spatial and temporal patterns. This tiny architecture allows real-time deployment with lower computing costs, which is critical for wearable or mobile BCIs. The model outperforms LDA and SVM in classification precision and reduces feature engineering, making BCI systems more automated and scalable [8].

Additionally, Jahanikia et al. demonstrate in their research the use of ML paradigms for detecting mental states related to inner speech, enabling users to control external devices via thought alone. The study applies a range of supervised learning techniques, including Support Vector Machines (SVM) and Random Forests, to classify neural signals associated with speech production. This project offers a potential method for individuals with speech impairments to communicate using only brain activity. The ability to decode inner speech commands can lead to more intuitive, non-invasive communication tools, advancing BCIs toward more seamless user experiences [9].

Acı et al. used machine learning to discriminate mental focus states in users using a passive EEG-based BCI system, demonstrating the impact of ML on BCI systems. They classified attention states from EEG data using supervised learning approaches like Linear Discriminant Analysis (LDA) and K-Nearest Neighbours (KNN) and unsupervised methods like Principal Component Analysis (PCA). Their research shows that the system can accurately categorise attention-related cognitive states, making it realistic for passive brain-computer interface (BCI) systems to measure and respond to user interaction. Such technology could be used in education to assess attention levels in real time or in healthcare to monitor cognitive states. This work shows how ML may improve human-computer interactions [10].

5. Conclusion

This essay explored key aspects of non-invasive BCI, including signal acquisition, neural decoding methods, and machine learning algorithms. The development of these aspects greatly shaped user experiences in BCI and shows significant promise in advancing human-computer interaction. Examining the contents covered in this essay, there are also areas for improvement. While the essay discusses signal acquisition, decoding, and machine learning applications, there is no further comparison of different algorithms and their specific uses. More detailed experimental data or case studies could be included in future research to make the analysis stronger. Further, it did not dive deeply into examples of how BCIs are practically implemented in real-world HCI applications. Future research could provide in-depth case studies or application examples, such as BCIs in assistive mobility tasks for individuals with disabilities. Looking ahead, the future of non-invasive BCIs is to not only enhance HCI but also open up entirely new dimensions of human-computer interaction, with the potential to significantly impact fields such as healthcare, gaming, smart home automation, and personalized learning.