A. EEG, ECoG, fMRI, fNIRS: Strengths and Weaknesses by Measurement Method

There isn't just one way to look inside the brain. Astronomers use visible light, infrared, and radio telescopes to see different faces of the universe. Neuroscientists do the same, peering through several kinds of 'windows' to observe the brain. Some windows are fast but blurry. Others are sharp but slow. Some peer in from outside the skull; others go directly inside. To understand BCI, you first need to know the characteristics of these windows.

EEG, or electroencephalography, places electrodes on the scalp to measure the brain's electrical activity. It picks up voltage differences generated when millions of neurons fire at the same time. EEG's strength is temporal resolution. Electrical changes in the brain happen on the scale of milliseconds, and EEG can capture these rapid shifts in real time. The equipment is relatively cheap, portable, and requires no surgery. From gaming headsets to clinical-grade devices, EEG is the most widely used BCI tool.

EEG, however, has a critical weakness. Its spatial resolution is poor. Electrical signals from the brain weaken sharply as they pass through the skull and scalp, and signals from different regions blur together. Scientists compare it to listening to the crowd from outside a football stadium. When the fans roar 'GOAL!' all at once, you know a goal was scored. But you can never make out a conversation between two specific spectators. The concrete wall blocks it. In the same way, EEG can detect broad shifts in brain state, but it struggles to decode fine-grained commands like 'bend the right index finger.' Muscle signals from blinking or jaw clenching create noise powerful enough to drown out the brain waves entirely.

ECoG, or electrocorticography, emerged to overcome that limitation. Surgeons open the skull and place a grid of electrodes directly on or beneath the membrane covering the brain. If EEG is listening from outside the stadium, ECoG is watching the game from a VIP seat inside. With the skull barrier removed, signals come through far more clearly. Spatial resolution climbs to the millimeter scale, and high-frequency bands become measurable, making it easier to decode specific information like finger movements or the intention to speak. A study published in 2024 reported recording stable high-gamma signals from an ALS patient via ECoG over 36 months, maintaining an average accuracy of 91 percent.

The price of ECoG is clear. It requires a craniotomy. Infection is a risk, and it places a surgical burden on the patient. It has mostly been used for pre-surgical epilepsy evaluation and research purposes, and it remains a high barrier to entry for the general public.

fMRI, or functional magnetic resonance imaging, looks not at electricity but at blood flow. When a particular region of the brain becomes active, oxygen-rich blood rushes to that spot. fMRI uses magnetic fields to capture this change as a three-dimensional image.

fMRI's strength is spatial resolution. It can map deep structures of the brain with millimeter-level precision. For pinpointing where complex functions like emotion or memory occur in the brain, it stands in a class of its own.

From a BCI perspective, though, fMRI carries a fatal drawback: time lag. Neurons fire electrically on the millisecond scale, but it takes several seconds for blood to rush to that location. After a user thinks 'go right,' two to three seconds pass before fMRI can detect it. That makes it unsuitable for BCIs that need real-time control. The equipment is enormous, filling an entire room, and costs billions of won. The subject must lie perfectly still inside a narrow, noisy tunnel.

fNIRS, or functional near-infrared spectroscopy, is often called a portable version of fMRI. It takes advantage of the fact that near-infrared light can pass through the skull. A band worn on the head emits light and analyzes the reflected signal to measure oxygen concentration in cerebral blood flow. fNIRS is far cheaper and lighter than fMRI. You can take measurements while sitting, standing, or even walking. Because it is well suited to monitoring brain activity in everyday life, it has recently been put to active use in detecting driver drowsiness, measuring student focus, and stroke rehabilitation therapy.

fNIRS, too, has its limits. The depth that light can reach is restricted to the cortical surface, and spatial resolution stays at the centimeter level. Because of the delay in the hemodynamic response, it is unsuitable for rapid control, just like fMRI. There is also a practical problem: acquiring a clean signal is difficult when the subject has thick hair.

No perfect tool exists. EEG is fast but blurry. fMRI is precise but slow. ECoG is sharp but requires surgery. Since 2024, researchers have been trying hybrid approaches that combine multiple methods to overcome these limitations. Studies using EEG and fNIRS simultaneously, capturing both fast electrical signals and blood flow changes, are on the rise. One 2024 study reported that an integrated EEG-fNIRS system achieved a classification accuracy of 95.86 percent on a motor imagery task. That figure is markedly higher than EEG alone.

What researchers face when designing a BCI comes down to a matter of choices. The balance between precision and convenience. The balance between invasive and non-invasive. What do you want to measure? How quickly must the system respond? What risk is the patient willing to accept? The answers to these questions determine the tool.

B. Microelectrode Arrays and Flexible Electrodes: The Biocompatibility Challenge

One day in 2012, in John Donoghue's lab at Brown University, a woman drank coffee with a robotic arm. Cathy Hutchinson, age 58. A stroke had left her quadriplegic fifteen years earlier. Implanted in her brain was a small chip called the Utah Array. It was a silicon plate smaller than a fingernail, bristling with 96 tiny needles. Those needles read neuron signals from her motor cortex, a computer interpreted them, and commands went to the robotic arm. Using thought alone, she picked up a cup and brought it to her lips. It was the first time in fourteen years.

That scene showed the world what invasive BCI could do. But even amid the celebration, researchers felt uneasy. How long would the Utah Array keep working? Would the brain truly accept these metal needles?

The core of an invasive BCI is the electrode. Instead of listening faintly from outside the scalp, the invasive approach goes close to the brain to listen clearly. Microelectrode arrays like the Utah Array are inserted directly into the cortex and record the firing of individual neurons. They can pull signals from hundreds of channels at once, and those signals are incomparably clearer than what EEG can offer.

The problem is that the brain does not welcome the intruder. Brain tissue is extremely soft, like tofu or pudding. Traditional electrodes, by contrast, are made of silicon or metal and are relatively rigid. This mechanical mismatch causes serious trouble. The brain does not sit still. With every heartbeat, blood flow makes it expand and contract slightly; every time you turn your head, it sloshes inside the skull. A stiff needle lodged in soft, moving brain tissue inflicts continuous damage on the surrounding area. Like a blade slicing through jelly.

This triggers the brain's immune system. Microglia and astrocytes swarm around the electrode, the intruder, and set off an inflammatory response. Eventually they encase the electrode in a tough scar tissue. This scar acts as an electrical insulator, gradually preventing the electrode from picking up neuron signals. That is the single biggest factor limiting the Utah Array's lifespan. A few months after implantation, signal quality begins to deteriorate. After a few years, a significant number of channels lose function.

To solve this problem, researchers turned their attention to flexible electrodes. What if you made electrodes from materials whose mechanical properties resemble brain tissue? The electrode threads developed by Neuralink are a prime example. Gold electrodes are deposited onto a polyimide film one-twentieth the thickness of a human hair, allowing them to ripple with the brain's movements. In theory, this can dramatically reduce the damage and immune response caused by mechanical mismatch.

But flexible electrodes don't end with 'softer is better.' If they're too flexible, inserting them into the brain becomes difficult. You need a certain stiffness to pierce the brain's protective membrane, but a flexible electrode just bends over.

Various engineering approaches have been tried to solve this. Neuralink developed a surgical robot based on the sewing-machine principle. A rigid tungsten needle grips the flexible electrode thread and pushes it into the brain, then the needle alone withdraws. The robot uses computer vision to detect blood vessels on the brain surface in real time and steers the electrode around them.

Another approach is dissolvable coatings. The flexible electrode is coated with sugar or a biodegradable polymer to make it rigid during insertion; once inside, bodily fluids dissolve the coating and the electrode becomes flexible. Charles Lieber's group at Harvard demonstrated a technique for injecting mesh-shaped electrodes into the brain through a syringe. The mesh unfolds among brain tissue, naturally intertwining with neurons, and was reported to minimize immune response.

A study published in Nature Communications in 2024 introduced a method for inserting ultra-flexible electrodes through blood vessels. Without a craniotomy, the electrode is pushed up through the jugular vein to a blood vessel near the brain, then guided through the vessel wall to settle into brain tissue. In experiments on sheep, this method succeeded in recording signals at the single-neuron level. By 2025, a 1,024-channel ultra-thin electrode array that could be inserted through only a small slit in the skull, without a full craniotomy, was validated in a pig model and in a human surgical setting.

Advances in materials science deserve attention as well. Flexible electrodes based on nanoporous graphene achieved low impedance and high charge injection capacity even with electrode diameters as small as 25 micrometers. Biocompatibility was maintained after chronic implantation for 12 weeks in rodent brains and 8 weeks in peripheral nerves. Research is also underway on coating electrode surfaces with conductive polymers or hydrogels to soften the interface with brain tissue, or applying anti-inflammatory drug coatings to suppress the initial immune response.

The ultimate goal is a chronic BCI that works for years, even decades, without performance degradation. That requires more than mechanical flexibility. The electrode material must resist corrosion from bodily fluids, and thermal management is essential to keep heat from wireless transmission from damaging brain tissue. As of 2025, however, the maximum lifespan of most chronically implanted flexible electrodes remains around one year. For the human brain and a machine to coexist for decades, there is still a long road ahead.

Biocompatibility is not just a materials problem. Damage at the moment of insertion, repeated micro-motion injuries, cumulative immune response, electrochemical degradation, and the impossibility of long-term maintenance all pile up to destroy the signal. The real benchmark is not 'it works at first,' but 'it still works at the same quality six months, one year, and several years later.' Reconciling the brain, a soft and sensitive tissue, with a machine, a hard and foreign object. That is the most fundamental barrier invasive BCI must overcome.

C. Signal Bandwidth and Accuracy: What Makes a BCI 'Good Enough'

In 2021, a man at Stanford University wrote text using nothing but his thoughts. He was a quadriplegic patient who imagined writing letters by hand, while electrodes implanted in his brain read signals from his motor cortex. A computer interpreted those signals and displayed characters on a screen. Ninety characters per minute. That far exceeded the average smartphone typing speed of 40 to 50 characters per minute. The research team published the results in Nature, showing that BCI could become a practical communication tool.

BCI always looks impressive in demo videos. Moving a robotic arm with thought alone, steering a wheelchair, playing games. But once a user relies on it every day, the spectacle stops mattering. What works in a lab and what holds up in daily life are entirely different problems. The conditions for a "usable BCI" are defined not by feelings but by numbers. Bandwidth, accuracy, latency. These three are what count.

Bandwidth refers to how much information can be transmitted at once. Elon Musk defined the core problem of BCI as "the bandwidth bottleneck." When we tap a smartphone screen with our fingers, we input dozens of bits per second. Early EEG-based BCIs managed only a few bits per minute. Users could select "yes" or "no," or nudge a cursor at a painfully slow pace. That level of frustration made real-world use impractical.

In the BCI field, speed and accuracy are evaluated together through a concept called information transfer rate. In classification tasks where the system must identify a user's intention from several options, the correct response rate, number of choices, and number of attempts are combined and converted into bits per minute. As of 2024, noninvasive EEG-based typing systems remain at roughly 5 to 10 bits per minute. Invasive ECoG-based systems, by contrast, have reached tens to hundreds of bits per minute. Stanford's handwriting-imagination study recording 90 characters per minute stands as a symbolic case illustrating this bandwidth gap.

Accuracy is "the probability that when I think A, the computer recognizes it as A." Ninety percent accuracy sounds excellent, but if one out of every ten attempts produces a typo or a wrong click, users quickly grow exhausted. The cost of undoing an incorrect command is high. And accuracy is not a simple number. It includes the shape of the errors. In a cursor-control BCI with 90 percent accuracy, if the errors take the form of "occasionally lurching in the opposite direction," users must stay tense at all times. If the errors instead mean "slightly less movement than intended," correction is possible.

Brain signals change over time. Morning and evening brain states differ, and when you are tired or your mood shifts, the same thought produces a different signal pattern. Even microscopic movement of an electrode alters signal characteristics. This is called nonstationarity. Older BCIs required a 30-minute calibration session before each day's use. That severely undermined practicality. A "usable BCI" must either maintain performance over long periods after a single training session, or its algorithm must adapt on its own while the user operates the system. In the case of Neuralink's first patient, Noland Arbaugh, some electrodes pulled away from the brain after implantation and signals weakened, yet the team

adjusted the decoding algorithm and reportedly restored accuracy.

Latency is central to the user experience. The machine must respond the instant you think, or you won't feel it as part of your own body. If a cursor moves one second after you think, you cannot accept that tool as an extension of yourself. Research indicates that the delay from thought to machine response must stay within 100 to 200 milliseconds for users to feel a sense of agency, the feeling that "I did that." Achieving this requires the entire pipeline, reading brain signals, transmitting them wirelessly, and having an external computer interpret and issue commands, to execute very quickly. Recent work has introduced edge computing, where the implanted chip itself performs initial data compression or spike detection, reducing transmission delay.

Depending on the measurement method, these conditions face different structural limits. Noninvasive EEG and fNIRS are safe and accessible, but their low signal-to-noise ratio makes high-speed, high-precision control difficult. ECoG involves surgical burden but offers signal quality favorable for high-performance decoding. Fully invasive microelectrode arrays have the potential to capture the highest-quality signals, but long-term stability and biocompatibility remain obstacles.

The conditions for a "usable BCI" are defined differently in clinical and consumer markets. For patients with severe paralysis, the key question is whether they can express their own intentions at all, even if slowly and somewhat inconveniently. In consumer BCIs,

speed and accuracy alone are not enough; user experience factors like comfort, battery life, and design become more important adoption criteria. A "usable BCI" means an interface so fast, accurate, and attuned to the user's mind that the person forgets they are operating a machine at all. Current technology has not yet reached that level.

D. Deep Learning Decoding: How Machine Learning Interprets Brain Waves

In 2016, a research team at Meta (then Facebook) launched an ambitious project. A brain-to-text interface targeting 100 words per minute. The goal was to let people write with thought alone, without moving their fingers. The problem was that the electrical signals collected from the brain looked like meaningless noise. Crackling waveforms. How do you extract the intention "hello" from that? The answer lay in artificial intelligence.

The core of brain signal decoding is pattern recognition. You find the signal patterns that appear in the brain when a user has a specific intention, then classify incoming signals by which intention they match. In the past, humans defined the rules manually. They would specify that "a decrease in alpha waves between 8 and 12 hertz indicates an intention to move," extract only that feature, and feed it into a classifier. Traditional machine learning methods like linear discriminant analysis and support vector machines were used.

This approach had clear limits. Brain signals are so complex that handcrafted features fail to capture much of the information they contain. Brain structure and signal patterns differ from person to person, and even the same person's signals vary with their state. Deep learning transformed this process. The key idea: stop having humans pick the features, and let the model learn directly from the data.

Convolutional neural networks, a structure that proved powerful in image processing, were applied to brainwave data as well. EEG signals are treated as two-dimensional data spanning time and channel (electrode position), and the network learns complex spatial and temporal patterns on its own. It picks up on subtle neural firing patterns that no human taught it to find. EEGNet, published in 2018, became a landmark model: a compact convolutional neural network architecture optimized for EEG signal processing that achieved high classification accuracy even with limited data.

Recurrent neural networks and long short-term memory networks excel at processing data that unfolds over time. They are suited for grasping context: "Given the signal that appeared earlier, this current signal probably means that." They were used to interpret continuous signals where previous states influence the current one, such as the trajectory of a reaching movement or the process of speaking a sentence.

The most talked-about development is the transformer architecture. The transformer, the core technology behind large language models like ChatGPT, is now being applied to BCI. The transformer's attention mechanism assigns weights to determine which parts of the data matter, capturing overall context. It treats brain signals as if they were a language. When brainwave signals are sliced into segments like words and fed into a transformer, the model reads the surrounding context and interprets which intention that brainwave pattern corresponds to.

Studies published in 2024 showed transformer-based decoders significantly outperforming prior methods. In one study, a convolutional transformer network achieved 82.52 percent accuracy in subject-specific motor imagery classification and 58.64 percent in cross-subject evaluation. In inner speech recognition research,

a spectro-temporal transformer demonstrated markedly higher accuracy than conventional EEGNet-based approaches. The reason is that transformers can simultaneously learn long-range temporal dependencies and frequency-domain dependencies.

In invasive BCI, the impact of deep learning is even more dramatic. Signals from ECoG or microelectrodes are far clearer than EEG. The better the signal, the stronger deep learning becomes, because the model can learn intention-relevant patterns more directly. Recent work combining high-density ECoG with deep learning has begun targeting high-speed communication at the level of dozens of words per minute. Decoders are evolving beyond simple classification to incorporate language models, sequence decoding, and error correction.

The latest trend is foundation models. The idea is to build a large-scale AI trained specifically on brainwave data from tens of thousands of people. Just as a person who already knows English can quickly pick up new medical terminology, such a large model can understand a new user's brain signals with very little data. This is called transfer learning. It is expected to become a key to mass adoption by drastically reducing BCI's chronic problem of long training times.

Deep learning has its own challenges, though. The first is the black box problem. It is difficult to explain why a deep learning model made a particular decision. In medicine, comprehensible reasoning matters. Research is underway to apply explainable AI techniques to brainwave decoding, analyzing which time windows, which channels, and which frequency features the model focuses on.

There is also the session-to-session drift problem. Yesterday's brain is not today's brain. EEG electrode positions and impedance shift, and invasive electrodes see signal distributions change due to long-term biological responses. A model may find that patterns learned yesterday no longer work today. Individual differences are another issue. Skull thickness, cortical structure, and physiological noise vary from person to person. A universal model remains a dream; in practice, personalization is usually necessary.

Deep learning does not solve BCI in one stroke. But as signals improve, as data accumulates, and as operations grow more refined, it serves as an amplifier that dramatically boosts performance. Better hardware provides cleaner signals; better software interprets those signals more accurately. A virtuous cycle. Deep learning is, in the end, a translator that converts noisy electrical signals into meaning and communication. It stands between the brain and the machine, interpreting each one's language for the other.