Brain–machine interfaces (BMIs) enable direct communication between the brain and machines [1,2]. Due to advancements in information and communication technology, BMIs have gained attention for their promising applications in medical, industrial, and household settings [3,4,5,6].

In a unidirectional BMI, the system consists of three components: devices used to record neural signals, components used to analyze the signals, and devices used to provide commands to operate the machine, as illustrated in Fig. 1. In a bidirectional BMI, additional components are required to provide feedback from the machine to the brain [7,8]. While both invasive and non-invasive methods have been developed for the acquisition of neural signals, the present review focuses on implantable neural probes. Thus, hereafter, we will use the terms “neural probes” and “invasive methods” interchangeably.

Electroencephalography (EEG) is widely utilized in non-invasive BMI systems due to its high temporal resolution, which makes it useful for mapping associations between EEG signals and cognitive function [9,10,11]. For example, NeuroSky, Inc. (San Jose, CA, USA) introduced an EEG-based brain–medicine interfacing headset for use in healthcare settings. In addition, Emotiv, Inc. (San Francisco, CA, United Sates) provides a 14-channel system for measuring EEG and other bio-signals for use in brain–computer interfacing games and neurofeedback treatment. Functional near-infrared spectroscopy (fNIRS) and magnetoencephalography (MEG) have also been used to develop non-invasive BMI systems [12,13,14,15]. However, non-invasive neural methods are limited in that neural signals from non-invasive probes are typically insufficient for complicated tasks that require a high degree of freedom, such as robot control [16,17,18,19]. For this reason, implantable neural probes are preferred for BMI systems that demand accurate controls and adjustments (e.g., neuroprosthetic devices).

Implantable neural probes are defined as devices implanted into the brain or other nervous tissues. Communication between neurons in the brain occurs via electrical and chemical signals. In most cases, electrical signals are the main source of information in BMI systems. In particular, single-unit activity (i.e., spikes) is regarded as most appropriate for extracting meaningful information, such as movement-related activity [20]. While non-invasive methods record neural activity through different media such as the dura matter, cerebral spinal fluid (CSF), and skull, implanted neural probes can record extracellular activity or local field potentials closer to neurons. To maximize the signal-to-noise ratio, the conductive material is often exposed at the end of the electrode, the shank of which is insulated with non-conductive material. Typically, single-wire electrodes [21] and glass micropipette electrodes are used in electrophysiological studies [22,23]. Recent advancements have enabled the development of implantable neural probes with the technical characteristics necessary for practical BMI applications, which include high spatiotemporal resolution and high signal-to-noise ratio. Biocompatibility, biochemical stability, and miniaturization are also important, as neural probes must be inserted into the brain. Indeed, implantable multi-array neural probes with these characteristics have been developed and applied for research, diagnosis, and treatment purposes. Representative neural probes in this category include the tetrode, Utah array, and Michigan probe.

In the present report, we review the features of these three implantable neural probes and their applications in BMI systems. Furthermore, we describe several novel technical approaches for improving the measurement of neural signals in BMI systems: (1) decreasing the invasiveness of implantable neural probes while maintaining performance, (2) using multi-modal probes to measure both electrical and optical signals from neurons, and (3) using flexible neural probes to enhance signal quality and biocompatibility. We further discuss (4) novel fabrication techniques and materials that can be utilized to improve neural probes. By exploring recent developments in implantable neural probe technology, we offer insight into the performance required to create practical, efficient, and accurate BMI systems.

BMIs have been successfully applied in the field of neuroprosthetics, enabling patients with paralysis to control robot arms using their thoughts. The underlying neural principle of this phenomenon involves population vectors, in which each neuron “votes” for the intended movement [20]. Increasing the number of neuronal signals increases the likelihood of movement in the population vector. Thus, probes including a high number of channels are suitable for BMIs. In this section, we discuss three popular implantable neural probes with high numbers of channels.

The tetrode is widely utilized to record extracellular electrical potentials in neural systems. As the prefix (tetr-) indicates, the tetrode consists of four electrodes at which neuronal signals are detected from slightly different spatial points originating from the same source. Such electrodes are typically constructed from a platinum–tungsten alloy and insulated with quartz coating. The metal end of the electrode is exposed to acquire electrical signals. The diameter of each electrode at the metal end is usually less than 30 µm, and may be up to100 µm at the coated end. The coating functions to minimize the interference of electrophysiological signals across electrodes. As illustrated in Fig. 2A, a single tetrode can measure the extracellular potentials of approximately 1,100 neurons within a 140-µm radius in the rat cortex [24,25,26]. One main advantage of this method over single-channel electrodes is that users can classify the extracellular potentials of adjacent neurons using a clustering approach [27,28,29]. The extracellular potentials of each neuron are detected by the four electrodes in the tetrode with different temporal points and waveforms because due to the difference in spatial distance between each electrode and the neuron. Spikes from multiple neurons can thus be separated in the post-processing stage. However, the tetrode cannot provide direct measurements of spatially multi-dimensional extracellular potential distributions unless multiple tetrodes are precisely implanted and arranged at regular intervals. Recently, several research groups have developed multi-tetrode arrays to compensate for these shortcomings [30,31].

The tetrode has been used as an implantable neural probe in BMI platforms, especially in small animals. Giszter et al. described a neurorobotic platform, which consisted of a tetrode-based neural recording module and a three-dimensional (3D) robotic module, for use with spinal or cortical prosthetics in small animals (i.e., frogs and rats) [32]. Song et al. developed a novel BMI system to identify movement-related information in the rat cortex during treadmill walking [33]. Using decoded neural activity acquired from six implanted tetrodes, specific regions related to proximal limb and trunk movements were identified in the rat motor cortex. In a follow-up study, Song and Giszter introduced a pelvis-attached robot that can be controlled by cortical neural signals using a multi-tetrode array implanted in the rat brain (Fig. 2B) [34]. Bender et al. investigated that activity of central complex neurons in the insect brain during walking using a tetrode-based system [35]. The authors reported a close association between movements and sensory responses, providing insight into the feasibility of such methods for use in larger animals with more developed brains. Additional groups have also reported advancements in post-processing techniques that enable sorting of spikes in practical BMI systems. For example, Oweiss introduced a novel spatiotemporal signal processing method to improve data compression and reduce latency in multi-tetrode BMI systems [36]. Kubo et al. further investigated the 3D distributions of neurons based on multi-site neural activity using the tetrode [37].

Advancements in semiconductor fabrication processes during the late 20^th century have led to the development of multi-channel arrays that perform better than single-channel electrodes in multiple respects. The Utah array is a commercially available intracortical electrode array consisting of up to 100 silicon needle-shaped electrodes, which are produced via microscale fabrication techniques such as thermomigration, a combination of mechanical and chemical micromachining, metal deposition, and encapsulation with a polymer made of imide monomers [38]. Due to the large number of electrodes, the Utah array has been mostly used in large animals, especially non-human primates (Fig. 3A). Velliste et al. investigated the ability of a BMI system to provide neuroprosthetic arm control via cortical motor activity in rhesus monkeys (Macaca mulatta) using Utah arrays [39]. In this study, the BMI system produced natural levels of multi-dimensional arm and hand movements. In addition, the same research group reported meaningful associations between visuomotor adaptation and neural activity in the primary motor cortex measured via the Utah arrays [40]. Capogrosso et al. developed a BMI platform to electrically stimulate the spinal cord according to decoded neural signals from the Utah array in the motor cortex of monkeys [41]. As described in Fig. 3B, the brain–spine interface enabled monkeys with spinal cord injuries to walk again based on commands from the motor cortex. Utah arrays have not only been used for recording, but also for stimulation purposes. For example, Tabot et al. induced tactile sensations in the hand via targeted neural stimulation with Utah arrays implanted in the somatosensory cortex, as illustrated in Fig. 3C [42,43]. This type of sensory feedback may aid in increasing the accuracy of various BMI devices and neuro-rehabilitation instruments. In addition, Suner et al. verified the reliability of chronic implementations of the Utah array in the brains of non-human primates [44].

The Utah array and its recording systems have been approved for clinical applications by the United States Food and Drug Administration (FDA). Several clinical trials of Utah array-based BMI systems involving human patients have been conducted. Simeral et al. reported that one patient with tetraplegia could control a computer cursor (including point-and-click functions) based on neural signals from the motor cortex [45]. Pandarinath et al. analyzed neural population dynamics during movement in two patients with amyotrophic lateral sclerosis (ALS) as they attempted to use their finger to move a computer cursor [46]. One year after implantation, the system still produced adequate signals for neural cursor control (Fig. 3D) [47] or virtual typing [48]. A more challenging task was performed by a human subject with Utah array implantation. Wodlinger et al. developed a BMI system for the control of an anthropomorphic robot arm and hand with 10 degrees of freedom [49]. Some studies have also indicated that BMI systems can bypass the spinal cord circuit to recover hand function in select patients. Bouton et al. introduced a platform for controlling a neuromuscular electrical stimulation sleeve using neural signals from an implanted Utah array, enabling the patient to perform accurate and continuous movements (e.g., grasp-pour-and-stir tasks) [50]. Ajiboye et al. developed an interfacing platform between intracortical neural signals recorded by two Utah arrays and functional electrical stimulation of peripheral muscles to restore arm and hand movements in patients with paralysis [51]. As experiments of tactile feedbacks has been successful in non-human primate, this type of research has begun to be applied to human subject. As tactile feedback experiments using Utah arrays have been successful in non-human primates, researchers have begun to apply such systems in human patients. For example, Flesher et al. stimulated the somatosensory cortex using Utah arrays to recover tactile sensation in human patients [52].

The previous two neural probes enable users to acquire neural activity from different cortical areas. However, they are limited in their ability to target deep neural structures in the axial direction [53,54]. In the late 1970s, the first prototype of Michigan probe, a multi-channel depth electrode prototype, was developed based on electron-beam lithographic techniques [55], which was before production of Utah array. Several preclinical studies involving small animals have demonstrated the safety of long-term implantation for the Michigan probe, suggesting that such probes can be utilized in BMI systems [56,57,58]. Electrodes in the Michigan probe (2 to 15 mm) are longer than those in the Utah array (0.5 to 1.5 mm for research). Thus, the Michigan probe may be more suitable when recording from deeper cortical structures.

Vetter et al. applied the Michigan probe to examine extracellular neural activity in the motor cortex of rats using a BMI system [56]. In a preclinical study on the control of emotions and memories, the Michigan probe was used to measure neural activity and provide electrical stimulation. Frost et al. investigated neural activity in the hindlimbs of rats with spinal cord injuries using a 16-channel Michigan probe and explored treatment methods based on neurostimulation [59]. Guggenmos et al. investigated the use of a neural interfacing system to recover neural function after brain injuries in small animals (i.e., rats) demonstrating that the missing brain functions can be restored using a brain-machine-brain interface as described in Fig. 4 [60]. Such findings suggest that neuroprosthetics can successfully applied in the treatment of various neural diseases. Michigan probes have also been validated in non-human primates [61], supporting the notion that such probes can be used to develop neuroprosthetic systems for use in humans.

In the previous session, we described the most widely used implantable neural probes in BMI applications. While the electrical characteristics, biological compatibility, and stability of these probes are sufficient, researchers have attempted to improve their impedance, flexibility, wireless communication, recording area, and accuracy to ultimately improve BMI systems. In this section, we present an overview of recent studies that have responded to these technical challenges for practical BMI applications.

While direct electrical measurements can be obtained using implantable neural probes as described in previous sections, optical measurement methods are also utilized to record neural activity at the level of single neurons and neuronal assemblies. Optical methods rely on indices such as the influx of calcium ions (Ca²⁺) and changes in voltage. Since changes in Ca²⁺ can represent neural activity, several optical indicators have been applied in studies ranging from in vitro cellular assays [101,102] to in vivo investigations in freely moving animals [103]: a bioluminescent Ca²⁺ probing protein (Aequorin) [104], chemical Ca²⁺ indicators (calcium green, fura-2 etc.) [105,106], genetically encoded Ca²⁺ indicators based on a single fluorophore [107] and Förster resonance energy transfer (FRET) [108]. In addition, voltage-sensitive optical dyes can be used to identify transient changes in action potentials in neurons [109,110,111]. In addition, voltage-sensitive optical dyes can be used to identify transient changes in action potentials in neurons. Indeed, researchers have developed optical implantable neural probes for use in freely behaving animals by combining fiberoptic and wireless communication technologies. For instance, Murayama et al. designed a miniaturized neuro-endoscopic periscope using graded-index (GRIN) lenses and micro-prism coupled fibers. This periscope was used to study dendritic Ca²⁺ changes in freely moving mice [112]. Ghosh et al. introduced a prototype of a miniaturized wide-field fluorescence microscope with which they could measure neural activity when the device was mounted on the head of freely moving animals [113]. Besides, optical neuro-stimulations were exercised since the discovery of optogenetics stimulation. Furthermore, optical devices can be used for stimulation purposes as well: In the case of Channelrhodopsin (ChR), trans-membrane pores are opened when ChR-integrated ion channels absorb blue light at a wavelength of 470 nm, causing ions to flow into the neurons [114,115,116]. Preclinical studies have confirmed that fiber-coupled illumination devices function well in freely moving animals. Aravanis et al. designed a fiber-coupled optical neural interface guided by ChR2-mCherry fluorescence images for use in rats [117]. Several research groups have also developed prototypes of wirelessly integrated LED-based implantable neural stimulation modules for use in freely moving animals [118,119,120].

Electrical and optical signals can be measured without interference from one another because their frequency domains are different. For this reason, several prototypes of multi-modal neural probes have been developed to measure both electrical and optical activity in preclinical studies [121, 122]. LeChasseur et al. developed a micro-probe, which consists of a dual-core optical fiber and a 50-µm electrical wire probe for recording both electrical and optical neural activities, as illustrated in Fig. 6A [123]. In a comparative study, the authors recorded optical and electrical neural signals from the same neuron in rats, observing that the two modalities were complementary and highly correlated with one another. Anikeeva et al. developed an optetrode, which is composed of a 200-µm single-channel optical fiber and a tetrode, to investigate optical stimulation and neural responses (recorded by the tetrode), as described in Fig. 6B [124]. The optetrode detected multi-unit neural activity and transient changes when optical stimulation was provided at different frequencies. Voigts et al. investigated an ultralight weight neural probe (i.e., FlexDrive) consisting of a single-core optical fiber and 16, 32, or 64-channel electrodes for studies of multi-dimensional neural responses to optogenetic stimulation [125]. Kwon and colleagues combined a transparent ECoG electrode array and an optical illumination device for optogenetic stimulation [126]. In another preliminary in vivo study, the authors proposed a prototype of a multi-modal neural probe termed the Opto-µECoG array, which exhibits sufficient biocompatibility. Several researchers have also introduced variations of the Utaharray combined with miniaturized optical components to acquire high-resolution electrical neural signals following optical neurostimulation [127,128]. Similarly, previous studies have reported the success of BMI systems based on optogenetic neuromodulation and electrical neural recordings in small animals [129,130,131,132,133,134,135]. For example, Liu et al. developed a compact optogenetics system using graphene electrodes to reduce artifacts [136]. We expect that multi-modal BMI systems will evolve toward more specific applications when the safety and stability of both probes have been demonstrated.

Based on previous research, we expect the following benefits to be associated with the use of multi-modal neural probes that combine optical and electrophysiological modalities in BMI systems. When using neural probes to record both optical and electrical neural signals, cross-validation of fidelity between the two signals can increase signal reliability. In particular, since the noise sources of the two signals are likely to be different, one signal can be used when another signal is perturbed. In the context of BMI applications, multi-modal probes may be useful when the patient is in an electromagnetically noisy environment. Such probes may also be advantageous for advanced BMI systems that require precise stimulation and minimal interference between signals. However, few optogenetic genes have been identified and cleared for use in humans due to safety concerns [137,138]. Genetic engineering remains quite challenging due to its unknown long-term impact on humans, although it is expected that these technologies will be utilized in various BMI systems once issues regarding safety and stability have been addressed.

In order for successful implantation in the human brain, neural probes must exhibit sufficient biocompatibility, safety, stability, and electrical performance. However, electrodes composed of classic materials such as iridium oxide and platinum are limited in their ability to meet these conditions. To overcome these limitations, advanced materials with special functions and properties including metals, inorganic materials, and polymers have been applied to the development of neural probes.

Carbon nanotubes (CNTs) are cylindrically structured allotropes of carbon that are stronger and have better elasticity and electrical properties than more traditional materials [139,140]. As such, various studies have aimed to develop neural probes using CNTs. Wang et al. designed a microelectrode array composed of CNTs, reporting significant improvements in the charge injection limit when CNT electrodes were used to stimulate cultured hippocampal neurons [141]. Keefer et al. developed a CNT coating to enhance charge transfer in microelectrode arrays [142]. In this study, the authors obtained stronger neural activity from CNT-coated electrodes in the primate visual cortex, with an improvement factor of 7.4 dB (Fig. 7A). Guitchounts et al. developed 16-channel carbon fiber neural probes with which they could acquire long-term chronic neural signals due to the strength and stability of the carbon fiber [143]. In addition, previous studies have demonstrated the successful integration of a soft CNT fiber microelectrode array in long-term bidirectional neural interfaces. Such findings suggest that this system can be used to develop high-performance and biocompatible BMI devices and neuroprosthetic instruments [144]. Graphene, a two-dimensional carbon allotrope, can also be utilized to improve the performance of neural probes due to its mechanical stiffness, flexibility, and superior electrical properties [145]. For instance, Chen et al. designed flexible graphene microprobes that can be applied to record electrical signals associated with neural or cardiac activity [146]. Graphene-based flexible electrodes have also been fabricated to acquire ECoG signals by improving contact with the brain. Additional studies have indicated that graphene can be used to investigate changes in neural activity due to optogenetic stimulation due to its high transparency [147]. Similarly, Kuzum et al. developed a graphene electrode array with high flexibility and transparency: They acquired electrical neural signals and fluorescence images indicating changes in calcium flux using the same electrode [148]. Furthermore, graphene-based electrodes can be integrated into multi-functional neural probes. For instance, Liu et al. introduced neural probes that consist of graphene-oxide and gold-oxide electrodes, as illustrated in Fig. 7B [149]. This novel probe can be used to obtain measurements of both electrophysiological neural signals and electrochemical information via cyclic voltammetry. Indeed, this type of electrode can be used to monitor neural activity and various electrochemical changes after the induction of brain damage due to photo-thrombosis. Apollo et al. also investigated the application of needle-type flexible neural probes made from graphene oxide for use in bidirectional neural interfaces [150].

Due to the development of efficient techniques for fabricating and manufacturing nanoscale structures and particles, nanotechnology has quickly become relevant to biological and biomedical applications Indeed, nanoscale structures and particles have been applied to improve neural probes. Park et al. developed a nanoporous platinum electrode that can be used to record neural signals [151]. In this study, the authors reported that the nanoporous electrodes were associated with improvements in electrical properties, impedance, and charge injection limits when compared with commercially available alternatives such as platinized platinum and iridium oxide electrodes. Other studies have demonstrated the enhanced biocompatibility of gallium phosphide nanowire electrodes [152]. These nanowire electrodes were tested in the rat primary somatosensory cortex. Abidian and colleagues introduced nanostructured polymer-coated conducting electrodes with enhanced electrical properties, improved mechanical adhesion, and better neuronal attachment capabilities [153]. Piret et al. developed boron-doped diamond electrodes with three dimensionally fabricated nanostructures using multiple semiconductor fabrication techniques [154]. Relative to conventional boron-doped diamond electrodes, the nanostructured electrodes offer higher sensitivity to neural activity while maintaining stability and biocompatibility, suggesting that these electrodes can be applied in the development of highly sensitive BMI systems and rehabilitation instruments.

In this report, we have reviewed the application of current multi-channel neural probes (i.e., the tetrode, Utah array, and Michigan probe) to BMI systems, as well as developments regarding next-generation neural probes. When comparing the electrodes currently in use, it is important to choose the proper neural probe based on the function and purpose of the BMI application. For instance, for measurement and stimulation of pyramidal neurons in deep sulcus areas—which are known to directly control fine muscles such as those in the hand—a neural probe with a longer shank is required. We also discussed several developing neural probes, which aim to improve upon the electrical properties, biocompatibility, and reliability of current devices using multi-modal approaches. If high-density electrodes with improved biocompatibility can be developed, flexible ECoG systems can be more widely applied in both clinical and preclinical BMI studies. Although stent electrodes are currently at the proof-of-concept stage, their minimal invasiveness suggests that they can replace many other methods for acquiring neural signals in the future. Safety concerns represent the greatest challenge for BMI systems based on optogenetic stimulation. The application of new materials, high-resolution integrated technologies, and nanotechnology will ultimately contribute to the development of high-performance neural probes and BMI systems.

While the long-term fidelity of neural recordings is critical for success in BMI systems, several subsequent stages of processing are critical for ensuring the efficiency of these systems (Fig. 1). Decoding algorithms that allow for accurate analysis of acquired neural signals and encoding techniques that transfer outward information to the brain are also important components of a BMI system [155,156,157,158]. High-speed computing and wireless signal processing are also critical for the development of high-performance BMI systems that can be applied in real-world settings.

Perhaps the most important issue facing developers is the biocompatibility and mechanical suitability of implantable neural probes. Advancements in material science and mechanical techniques will aid in developing strategies for minimizing brain scarring while maintaining adequate electrode contact [159,160]. Assessments of signal quality and side effects should be performed in both in vitro and in vivo systems (e.g., non-human primates) to determine the practical applications of novel BMI systems prior to use in humans. For these reasons, it is also necessary to choose appropriate packaging techniques to enhance the biocompatibility of implantable neural probes.

Mass production and security concerns should also be addressed in the development of neural probes for commercial and real-world use. Because the electrode is inserted into the brain, proper mass manufacturing technologies are required, without sacrificing the quality of the probe from production to packaging. Thus, systematic inspection methods should also develop. Several security experts have cautioned that information leakage may occur when using BMI systems [161, 162]. Therefore, establishing technologies to prevent information leakage is another issue that must be addressed when considering the practical applications of BMI systems. In conjunction with technological advancements, consensus regarding social and ethical concerns will lead to widespread utilization of implantable neural probes and BMI systems.