fNIRS Headband: Building an Optical Brain Activity Sensor from Scratch

Functional near-infrared spectroscopy, fNIRS, measures brain activity by shining light into tissue and reading how much of it scatters back. It is a well established technique in research labs, usually running on equipment that costs tens of thousands of pounds. We wanted to know whether a working version could be built as an internal project, from first principles, using our own analogue design and embedded systems experience.

The Challenge

An fNIRS sensor is fundamentally an analogue design problem before it is anything else. The signal you are trying to measure, light scattered back through skull and tissue, is extremely small, and it sits underneath a much larger amount of noise from ambient light, motion, and the electronics themselves. Getting a usable reading means designing a precision analogue front end capable of pulling a genuinely faint signal out of that noise, then processing it in real time on embedded hardware small enough to sit comfortably on someone's head.

We took this on because we wanted to prove, from first principles, that we could take a technique normally locked inside expensive lab equipment and build a working, wearable version of it ourselves, using our own analogue design and embedded systems experience rather than an off-the-shelf reference design.

Building an optical sensing wearable prototype around this kind of signal also means the mechanical design and the electrical design can't be treated as separate problems. Motion artefact, the noise introduced simply by a sensor shifting slightly against the skin, is one of the most common sources of unreliable data in wearable biosensing, and no amount of clever filtering in software fully compensates for a sensor that isn't held consistently in place. That meant the headband's physical fit had to be engineered with the same rigour as the analogue front end itself, not treated as an industrial design afterthought once the electronics were finalised.

Our Approach

We started with the analogue front end, since it was the part of the design with the least room for error. Photodiode selection, transimpedance amplifier design, and filtering all had to work together to reject noise without also filtering out the signal we actually wanted. Each sensor node was built and tested individually before being mounted into the array shown above, so any faulty channel could be caught and corrected before final assembly rather than after.

Once that stage was validated on the bench, we built the embedded processing layer to sample, filter, and log the data in real time, and packaged the whole system into a headband form factor that could be worn comfortably for extended sessions.

Before handing the device over for client testing, we validated it on ourselves first, recording our own brain activity during different tasks and comparing the results against what published fNIRS research would predict for similar activity. Testing on a known, controllable subject before it reached the client gave us a genuine benchmark for whether the analogue design was actually working as intended, rather than finding out for the first time during the client's own evaluation.

Outcome

The finished prototype produced clean, usable data. One of our own recordings, shown below, tracked brain activity while switching between thirty seconds of mental arithmetic and thirty seconds of rest. The signal rises clearly during the arithmetic periods and falls back during rest, exactly the pattern published fNIRS research would predict for that kind of task switch, which gave us real confidence the analogue front end was doing its job correctly rather than just producing plausible looking noise.

Recorded signal alternating between 30 seconds of mental arithmetic and 30 seconds of rest.

We published this recording alongside the project rather than keeping it internal. That result led directly to further interest in the work: a client is now developing a miniaturised second version of the same underlying sensing approach for their own product.

This project is a useful example of what precision analogue design and embedded systems work looks like when there is nowhere to hide behind a simplified spec. If the front end design is wrong, there is no signal to recover afterward. Getting it right the first time is exactly the kind of technical risk we help clients identify and resolve before they commit to building something similar themselves.

What This Demonstrates

Optical biosensing hardware sits at the harder end of analogue design, the useful signal is measured in microvolts, and separating it from noise generated by ambient light, motion, and the sensing electronics themselves requires a level of front end precision that most general embedded projects never need to touch. Building this fNIRS wearable prototype from first principles, rather than adapting an existing reference design, meant every decision in the signal chain, photodiode selection, transimpedance gain, filter topology, had to be justified on its own merits rather than copied from a datasheet example.

What this proves in practice is that we can take a technique normally confined to expensive lab equipment and translate it into a compact, wearable brain activity measurement device, then validate that it actually works with real recorded data rather than a plausible-looking simulation. That combination, precision analogue front end design plus embedded signal processing plus genuine end-to-end validation, is the same skill set behind any sensor-driven product where the signal you care about is small, noisy, and easy to get wrong.

It's also worth being direct about why we chose to run this as a self-funded internal project rather than waiting for a client brief to justify it. Some of the most technically demanding problems in hardware, precision analogue sensing being a clear example, are exactly the kind of work that's hard to sell speculatively but easy to demonstrate once it exists. Building it ourselves, testing it on ourselves, and publishing the results honestly, drift, noise floor, and all, gave us a genuine credibility asset rather than a marketing claim, which is precisely why a client is now building on the same underlying approach.