Application of Machine Learning in Electroencephalogram and Bio-Electricity Signal Processing

1. Introduction

Since the seminal discovery of alpha waves by Hans Berger [1], electroencephalography (EEG) has arisen as a valuable tool both in clinical practice and in the investigation of neural function. In the field of cognitive neuroscience, EEG and event-related potentials (ERPs) have provided ample data for researchers to interpret the temporal course of cortical activation. Although EEG/ERPs lack the high spatial resolution of functional magnetic resonance imaging (fMRI), the high temporal resolution [2,3] of EEG is unmatched, and this technique remains a useful and affordable instrument to assess the temporal course of cognitive processing. In clinical practice, researchers have invested considerable effort into clarifying the correspondence between sleep stages and the characteristics of ongoing EEG patterns. The knowledge obtained from this line of research has been integrated into clinical practice [4], and polysomnography [5] is now considered the “gold standard” technique for the diagnosis of sleep disorders.

In many cases, particularly in basic research, EEG signals are measured at multiple locations scattered throughout the scalp surface. Thus, EEG is spatiotemporal in nature, and its multidimensionality makes it challenging to exploit the information contained in EEG to the full extent. In the early days of EEG research, the interpretation of EEG patterns relied solely on the qualitative observation of “waveforms”. It soon became common practice to apply signal processing algorithms to EEG data during preprocessing and analysis [3]. However, although some notable efforts have been made in the field of data-driven EEG analysis [6,7], many researchers have only investigated a small number of hand-picked features, such as spectral power in a predefined frequency band, peak amplitude, and latency in ERPs.

Researchers might miss a vast amount of useful information contained in the spatiotemporal pattern of EEG by relying solely on the traditional approaches of focusing on hand-picked EEG/ERP features. Another caveat of this approach is that each indicator of local neural activity is usually analyzed independently. The widely accepted view of neural function claims that some information is represented in a distributed manner in the brain [8,9] and that the ensemble activation of multiple neural populations may underlie an identical cognitive or motor function. To fully capture such a distributed pattern of neural activation, multivariate analysis [10] of “holistic” EEG patterns is more suitable.

Until recently, the development of custom-made pipelines for data-driven EEG analyses was beyond the reach of many cognitive neuroscientists and researchers in clinical medicine. However, this situation has changed following the advent of a suite of accessible and convenient libraries for machine learning. The application of machine learning has been the norm in the fields of brain–computer interfaces (BCIs) and brain–machine interfaces (BMIs). Researchers in these interdisciplinary fields have searched for algorithms to efficiently exploit the information embedded in EEG data with sufficient processing speed. In many of these studies, the model was trained using supervised learning to achieve a predefined goal based on instructions provided in the form of measured EEG data [11]. Dimensionality reduction in EEG data is often utilized together with supervised learning to enable the efficient handling of data and feature extraction with limited computational resources [12]. Machine learning and multivariate analyses are gaining popularity in other basic and applied neuroscience fields. The application of a convolutional neural network (CNN) to EEG spectrograms is also a notable example [13,14]. Currently, these lines of research primarily focus on the practical use of EEG, such as the development of assistive technology for clinical diagnosis and novel educational interventions. It is also possible to simultaneously gain novel insights into neural and physiological functions and how information is represented in EEG patterns by scrutinizing and interpreting the resultant model using explainable AI tools.

Bioelectric signals other than EEG have been utilized as valuable sources of information in various fields of research and clinical practice. At first glance, the signal pattern is simple compared to EEG, with a relatively small amount of useful information contained in peripheral signals, such as electrocardiograms, galvanic skin conductance responses, and electromyography. However, close attention should be paid to recent findings obtained by the application of machine learning to peripheral biological signals [15,16,17], which indicate that seemingly simple signals can sometimes contain far richer information than expected. For example, a recent study [15] showed that a participant’s emotional state could be inferred based on the waveform of a single-pulse wave peak. Moreover, peripheral bioelectric signals are generated through a complex web of interactions between central nervous system activation and the functioning of peripheral effectors and could thus potentially contain rich information on the function of both the periphery and central nervous systems. Basic methods to exploit information from peripheral signals have been used for many years and are generally considered well established [18,19,20]. However, considering the potential richness and complexity of the information in peripheral signals, the application of machine learning has opened up the possibility of extracting further information from the patterns of peripheral signals.

2. An Overview of Published Articles

Epileptic seizures are characterized by sudden outbursts of abnormal electrical activity in the brain. Along with the assessment of sleep stages, the automatic detection of epileptic seizures is among the most well-studied topics in the application of machine learning to EEG data. The foremost goal of this line of research is to improve the classification performance of normal and abnormal EEG patterns. Improvement in classification performance also leads to the detection of EEG features specifically linked to epileptic seizures, which, in turn, provides researchers with a better understanding of this neurological condition. Mera-Gaona et al. [21] proposed a novel ensemble feature selection method to capture a subset of relevant EEG features by aggregating the results of multiple feature selection algorithms. They succeeded in identifying a robust and stable subset of relevant features, achieving high classification performance based on the selected features. The proposed method thus has significant utility in the early detection of seizures in clinical situations. Moreover, a deeper look into the set of selected features could be a valuable source of information for clarifying the etiology of epileptic seizures.

The incorporation of novel features could potentially improve the predictive performance of machine learning models based on peripheral bioelectric signals. There are a number of features that researchers and clinicians focus on in the assessment of ECG, including heart rate variability indices in both time and frequency domains, and indices of ECG waveform, such as the amplitude of and intervals between PQRS peaks. By incorporating recently crafted topographical data analysis (TDA) features, which capture the nonlinear dynamic properties of ECG, Ling et al. [22] further showed that the performance of ECG-based ventricular tachyarrhythmia (VA) was improved compared to previous methods. Their study illustrated the point mentioned above that hitherto neglected information useful in tackling real-life problems may be embedded in peripheral bioelectrical signals.

Machine learning has the potential to provide novel information on biological signals other than EEG and peripheral bioelectric signals. Several studies have further reported recent attempts to apply machine learning to the analysis and utilization of bio-optical signals, with measurement techniques ranging from pulse wave measurement to functional near-infrared spectroscopy.

3. Conclusions

EEG and bioelectric signals have a long history of application in basic research and clinical practice, and a wealth of knowledge about the informative features embedded in these signals already exists. However, the application of machine learning to biological signals has the potential to add novel insights into the information embedded in the multidimensional data of bioelectric and bio-optical signals, resulting in a wider application of biological measurements in real-world settings.