EEG, neuro-analytics, and brain-computer interfaces: inside Parul University’s hands-on BCI workshop!

EEG, or electroencephalography, measures the brain’s electrical activity using sensors placed on the scalp. The computer picks up that activity and converts it into a command that lets a person move a cursor, a wheelchair, or a robotic arm without moving a muscle. This is a brain-computer interface, or BCI. The first one is a measure, and the second one is a control system on top of the first one.

The appeal is that it is a non-invasive method of measurement. Much of the clinical neuroscience is still based on subjective behavioural evaluation, self-administered questionnaires, or, at the extreme, invasive surgery to gain insight into what is going on in somebody’s brain. Computational neuro-analytics is another way: it observes complex activities in the brain and the internal mental processes without entering the skull and prior to any surgical procedure.

Parul University in October 2026, along with ITIE Knowledge Solutions and ADInstruments, hosted a five-day national hands-on program at Lakshya 2047 Centre for Future Skills, Vadodara, focusing on the same subject from the start of signal acquisition up until the application of AI and BCI. The following is a working guide to the field, which is based on what that program taught, and the sequence that the program has followed is the sequence that anyone has to understand to do this work; that is, where the signal comes from, how it is cleaned, and how it becomes an intelligent system.

The first fact for a clean account of EEG is that the signal is small. The level reaching the scalp is typically 10 to 100 microvolts (millionths of a volt) and therefore must be amplified many times by a computer to be read, and is easily lost in regular electrical noise in a room. It is also not what people would think. Scalp EEG isn’t recording when single neurons are firing. These are too short and not synchronised to pass through the skull. The signals recorded by the electrodes reflect the combined activity of many vertically aligned pyramidal cells in the cortex, which take a relatively long time to postsynaptically process signals. In order for anything to show at the surface, a patch of about 10 square centimeters of cortex needs to fire in synchrony, as the skull and cerebrospinal fluid are insulating. A network of excitatory and inhibitory neurons, driven by thalamocortical loops, creates rhythmic oscillations at characteristic frequencies, which are related to mental states in a way that is close enough to make them the language of the entire system.

The simplest proof that any of this is real is a phenomenon called alpha blocking, which the workshop used as its first live exercise. The EEG headbands were worn on the participants, with their alpha power rising when they closed their eyes, because their visual lobe is idle when there is no visual input, and falling the moment they opened their eyes, when the visual lobe started to process. The fact that anyone can perform a sanity check and, if it moves when closing their eyes, then it is real brain activity, not ambient hum, is a testament to this.

Most of the engineering involved in EEG is really about rejecting everything that is not the signal, as it is so small. Competing electricity is generated by the body itself (from eye blink to muscle tension to the heartbeat), as well as by the environment (the fifty-hertz hum of mains power). The hardware answer is the differential amplifier. It reads the difference between an active electrode and a reference and discards whatever the two have in common, which is mostly noise. The figure of merit here is the Common-Mode Rejection Ratio, and a usable instrumentation amplifier needs to exceed a hundred decibels of it. There is a catch that the workshop’s instrumentation sessions, led by Dr. Ramesh Kumar, stressed hard: if the impedance at the two electrodes is even slightly mismatched, common-mode mains noise converts into a false differential signal that no amplifier can then remove. Hence, the care taken at the scalp.

That care is visible in how a clinical cap goes on. The positions are given according to the international 10-20 system, based on the fixed anatomical landmarks in the brain, namely the nasion at the bridge of the nose and the inion at the base of the skull. The scalp is divided and gently abraded to remove dead skin and oils, conductive gel is applied next to each of the silver-chloride electrodes, and then the impedance between the electrodes is checked channel by channel to ensure it remains below the clinical level (five kilo-ohms), at which point the signal is considered reliable. The analogue signal is sampled and converted to a digital signal with a resolution of 24 bits to capture microvolt detail even when a large artifact spikes through the signal and sampled at a rate exceeding the Nyquist frequency to avoid aliasing.

Even a well-acquired recording is still a mixture of brain activity and artifacts, and it has to be separated before any analysis. Two ideas did most of the work here. The first is the type of filter you’ll pick, and it has a clear trade-off. Non-causal, zero-phase filters treat the entire recording as one and do not cause any phase distortion; however, they diffuse energy back into the past, which can cause a cognitive event to be shown on the graph a little before it actually occurred. A causal filter only uses past and current samples, and it preserves the true order of events, but there is a minor lag time. The first one is okay in the offline learning case, but in the case of a real-time BCI, or live neurofeedback, where future samples simply do not yet exist, causal filters are required, and latency is a design constraint.

Dr. Alpana Pandey‘s sessions on the second idea, which is a method of blind source separation called Independent Component Analysis, started here. Extended Infomax algorithms attempt to decompose the mixed scalp signal into statistically independent components (spatial and spectral signatures). An eye blink appears as a sharp low-frequency component with a dipole configuration that is symmetrical over the frontal electrodes; lateral eye movements as well as the cardiac pulse have their recognisable fingerprints. If identified, those components can be subtracted and the signal reconstructed, leaving no trace of the artifact on the signal but not compromising the underlying neural rhythms.

Only once the data is clean does machine learning enter, and the workshop was built around a point worth repeating: mastery of raw signal acquisition is the prerequisite for using AI here, not an afterthought. A noisy input will defeat even a strong model, so no amount of deep learning rescues a bad recording.

The analysis usually starts with a stable estimate of the signal’s frequency content. A raw Fourier transform of a non-stationary EEG signal is noisy and leaks energy across frequencies, so the standard move is Welch’s method, which splits the recording into overlapping windowed segments and averages them to produce a far more stable power spectrum. From there, feature extraction for movement-related tasks typically uses Common Spatial Patterns, a spatial filter that maximises the difference between two mental states, such as imagining left-hand versus right-hand movement, which is why it is the standard choice for motor imagery. Classification can then be done with classical models such as Support Vector Machines and Linear Discriminant Analysis, or with purpose-built deep networks such as EEGNet, a convolutional architecture that learns temporal filters and spatial patterns directly from the raw epochs.

One warning from these sessions deserves to travel beyond the room, because it is where a great many published brain-decoding results quietly go wrong. If any preprocessing step that learns from the data, such as scaling, is fitted before the data is split into training and test sets, the model has effectively seen the test set, and accuracy inflates toward an unreal ninety-nine percent. The honest standard is leave-one-subject-out cross-validation, which trains on some people and tests on a person the model has never seen, and is the only way to show that a brain-decoding result generalises rather than memorises.

With a clean, classified signal, the interface can drive something. The workshop demonstrated three of the paradigms the field relies on, and they differ in how the brain is made to produce a readable signal in the first place.

The vivid demonstrations were the motor-imagery ones: a thought-controlled wheelchair, a brain-controlled robotic arm, a remote-controlled car, and, on the final day, a RoboCar that mapped live attention and relaxation scores to the commands to move forward and backwards, respectively, but where the two were inverted. The participants also created and played neural games on this basis—for example, the Calm Ball game, in which the ball could only rise when relaxation exceeded a certain limit, or the Laser Focus game, in which the ball would ascend only if concentration remained constant. The games aren’t gimmicks; they’re the very same acquire-clean-classify-threshold pipeline with a screen on. Which is why the workshop was using the games to teach the coding behind the pipeline.

Dr. Iype Cherian connected the laboratory signal to the operating theater, where the intraoperative monitoring system and invasive BCI operate in a setting where electromagnetic artifacts are aplenty and where effective algorithms for suppression are crucial for patient safety, not just a convenience. The final day turned to where the field is heading and what it owes the people it measures. The transportation of lightweight deep-learning models onto headbands and edge devices, eliminating the need for large lab computers, is bringing with it new questions about continuous brain monitoring: privacy of neural data, informed consent to record someone’s brain state, and whether a decoding model is true or just accurate for those on whom it was trained. In a field as fast-moving as this, these are engineering requirements, not add-ons.

The program was designed in a way that it had people from applied sciences, computer engineering, biotechnology, and healthcare, because computational neuro-analytics is not something that belongs to one department; it’s something that is shared: a clinician, one studying a disorder; an engineer, one building an interface; a data scientist, one training a classifier; and all using the same pipeline. The Provost, Dr. K. N. Madhusoodanan, who is an advocate for personalised medicine, has argued that engineers with an in-depth understanding of biology and biologists with computational expertise is needed to enable such medicine, and Dr. Swapnil Parikh, Dean of the Faculty of Engineering and Technology, has made a case for interdisciplinary research across healthcare, engineering, and applied artificial intelligence. Parul University is running a five-day nationwide program of this magnitude with hands-on hardware and live clinical input instead of slideshows, which is a statement about the future-skills work that it plans to offer. The vision is simple – They’re not just software tricks to turn brain signals into intelligent systems. It requires biology, hardware physics, mathematics, and honest statistics in the same room, and that’s just what the program was designed to do. If you aspire to grow your career story at the intersection of health and technology, pursue Parul University’s B.Tech in Biomedical Engineering and get to work globally!

SOURCE AND CONTEXT. Based on the report of the five-day national STTP/FDP, From Brain Signals to Intelligent Systems, held at the Lakshya 2047 Center for Future Skills, Parul University, 12 to 16 October 2026, with ITIE Knowledge Solutions and ADInstruments.

• Dr. Sanjeev Kubakaddi Lead resource person; biological foundations, acquisition and the AI/ML pipeline
• Dr. Iype Cherian Neurosurgeon; clinical talk on intraoperative neuromonitoring and invasive BCI
• Dr. Ramesh Kumar Biomedical instrumentation and hardware physics sessions
• Dr. Alpana Pandey Advanced signal processing and Independent Component Analysis
• Mr. J. K. L. Prasad Advanced BCI paradigms on the final day
• Dr. K. N. Madhusoodanan Provost, Parul University; inaugural address
• Dr. Swapnil Parikh Dean, Faculty of Engineering and Technology, Parul University
• ADInstruments Scientific instrumentation collaborator
• Parul University Host; Lakshya 2047 Center for Future Skills, Vadodara