Sensor Placement Is an Algorithm Decision

Mostafa Dhouib·September 3, 2026·18 min read

Sensor Placement Is an Algorithm Decision

The short answer. On a device that sits on the body, picks up a signal, and acts on it, where you physically put the sensor decides how strong the signal is when it arrives, and that decides whether you can track it with simple math or need to train an AI model to dig it out of the noise. Those are two different products with two different timelines, 6 months versus 18. The decision is usually made in month 2 for comfort, by someone who does not work on signals, and nobody realizes it was made at all.

This is for you if you are building a device that sits on the body, picks up a signal, and does something with it in real time, or if you are funding one and want to understand what actually makes these products ship or die. I am going deep and not slowing down for hype.

There is one decision that is the difference between shipping in 6 months or 18 plus. Same product either way, three times the runway, and enough of a difference to sink startups that cannot fund the longer timeline. I found it from pattern recognition, watching similar teams, same idea, close budgets, good engineers, where some shipped fast and some dragged for years or failed. It was never that one team was smarter or better funded. It came down to one decision, made very early, usually by accident, often by the wrong person.

The decision nobody in the room knows they are making

Here is what makes it sneaky. It is not a decision your software person makes, or your hardware person, or your firmware person. These are all good people, each excellent at their own piece, and each owns a lane, the AI model, the circuit board, the loop code. This decision does not live in anybody's box. Walk up to any of them and ask who decided it, and every one will say "not me, not my area," and all of them are telling the truth. Nobody was lazy, nobody messed up, which is exactly why it is so hard to catch.

The person who actually made the call was whoever designed the shape of the device, the physical housing you wear. They figured out it should sit here so it is comfortable, looks good, does not fall off after 20 minutes. A comfort decision, made for fair reasons, usually very early, long before anyone is thinking about the signal. And that decision, where the device physically sits, quietly decided how hard the entire signal problem was going to be. Nobody in the room knew it, because that is not what they were thinking about. The gap between the person who placed the sensor and the person who has to make sense of it is the whole point of this article.

The timing trap: month 2 to month 9

In month 2 the physical design gets locked, because you had to order parts. Locking the shape locks where the sensor sits. Around month 4 you finally hire or assign the signal and software people, and someone looks at the actual signal for the first time. Around month 9, that person says "why is this so hard to work with?" The most important decision about your whole signal got made in month 2, by a person who does not work on signals, who had no idea they were making it. This is not a rare disaster. It is the default path.

Why these devices are hard: the loop and the cracks

Boil any of these devices down and it is a loop that repeats: sense (pick up the signal), decide (figure out what is happening right now), act (do something back to the body). Three things pile up at once. The loop runs through a living person, who is not a steady, predictable machine and changes minute to minute. The device's own action changes the very signal it is trying to read, like taking your own pulse while running. And the whole thing runs on a coin-cell battery and a shape you picked eighteen months ago and cannot swap now, like buying the engine and the gas tank before you finished designing the car. These devices almost never fail because one of those problems was too hard. They fail in the cracks between two of them, the place nobody is looking.

What you lose in the first inch, you cannot get back

Start with sensing, because everything after works from whatever this step hands over. The signal off your skin is unbelievably tiny, a few millionths of a volt. Meanwhile there is a constant electrical hum all around you, from the power in the walls, the wiring, the lights, and your body acts like an antenna and picks it up. So your sensor is not seeing your tiny signal, it is seeing your signal riding on top of a hum that can be a thousand times bigger. Picture trying to hear a whisper at a loud concert. The whisper is the signal.

There is a spec for how well your circuit handles this, called common mode rejection, which in plain terms is how good the circuit is at throwing away the noise sitting on both sensor wires equally while keeping the real difference you care about. And here is the rule the whole article hangs on: whatever your sensor and that first little circuit throw away is gone, and you cannot get it back later with software. Think of a photo. If it comes out blurry when you snap it, no amount of editing makes it sharp, because the detail was never captured. Same here. If the first stage loses your signal in the noise, the smartest AI on the planet cannot rescue it, because there is nothing left to work with. It just reads static and pretends it is signal.

The two facts that multiply

Fact one. Your signal comes from a specific place inside the body and is strongest right there. To reach the skin where your sensor sits, it travels out through tissue, spreading and getting weaker, like a flashlight under a thick blanket, bright right above the bulb and dimmer the further out you go. There is one spot where the signal is as strong and clean as it will ever be, right on top of the source, and everywhere else it is weaker. You do not get to choose where that spot is. Human anatomy sets it, and it is fixed.

Fact two. Where you actually put the sensor is a comfort decision, where it fits nicely, is comfortable to wear, looks right, stays put. And 99% of the time it is not chosen for the signal, and it is chosen before the signal people even show up.

Line those up. There is the loudest spot, fixed by the body, and there is where you are actually listening, chosen for comfort, and they are usually not the same place. Your microphone is comfortable, but it is not sitting next to the speaker. Multiply the two, and the gap between them is how much of your signal you lose before you even start. Engineers call it attenuation. That number decides what kind of software you need, and it does not just cost you a little signal quality, it decides whether you build one kind of product or a completely different one. It is the single most important decision on the whole project, and on most teams nobody makes it. It just falls out of two reasonable choices that nobody connected. And it is a number you can estimate on a whiteboard in an afternoon, not something you should be surprised by in month 9.

The fork: a tracker or a model, and two different companies

That attenuation number sits at a crossroads and decides whether you need a learned AI model at all.

Road one, the tracker. You did not lose much, the signal is strong at your sensor. Very often you can track it directly with straightforward math. Tracking means clapping along to a song whose beats you already know, you do not learn anything, you follow the rhythm that is there. No data to train, works today, and it does not break on new people because it is not guessing based on who it has seen, it is following the actual signal in front of it. There is a nasty trap inside real-time tracking that deserves its own article, but the big picture holds: if the signal gives you a rhythm to follow, you have a working product today, and you can add the fancy AI later as a bonus.

Road two, the model. You lost a lot, the signal is weak and buried, and you might genuinely need an AI model to dig it back out. Instead of clapping to a known beat, you have to record thousands of songs and train a system to even recognize what a beat is. And here is what nobody says out loud: every AI piece you drop into your device is not a feature, it is a chain of things you now owe. The model needs data, which has to come from real people wearing the device, which needs work and hardware to collect on, which needs time to train, and then you inherit drift, where the model learns from one group of people and slowly gets worse as it meets new ones, or acts strangely when the signal blurs, someone had a rough night, changed location, added noise. At the end you are holding a system you cannot fully explain, a black box, that you now have to defend to an investor asking a hard question you do not have a clean answer for. Count the AI pieces. Each one is delay, technical risk, and one more thing standing between you and shipping.

Put the two side by side and it is not a technical choice, it is a choice about what kind of company you become. The tracker company ships now with no data, the hard part is signal math which is known and solvable, the risk is small and predictable, they need one good signal engineer, and when investors poke, the product makes sense. The model company cannot ship before the data is ready, data nobody can promise, the hard part is collecting data and training, there is no clear finish line, the risk is drift which can drag on forever without the right expertise, they need a whole data and AI team, and when investors poke, they have a black box they cannot fully explain at the center of everything. Same idea on the whiteboard, completely different burn rate, timeline, and risk. And most teams pick between them by accident, in month 2, because of where the device fits comfortably.

Road two also hides a circular trap. The model needs data, the data comes from people wearing the device, wearing it needs the device built, and it is not built because you are still designing it, and part of what you are designing is the AI that needs the data. A snake eating its own tail, nobody planned it, and it drags an 8-month launch to 18.

You do not have an AI problem, you have a placement problem

So let me say it straight. Most teams in this situation think they have an AI problem. You do not. You have a placement problem, a decision made in month 2 whose consequences turned into something that looks like it needs an AI solution, because nobody caught the problem at the right time. It is like going to the doctor with a broken bone and getting treated for the bruise. Wrong diagnosis, wrong treatment, and you burn time and money while the real thing never gets fixed.

Treat it as an AI problem and you hire AI people, buy expensive computers, burn capital training a model, grind for a year or more, raise more rounds, and your burn rate goes astronomical, trying to pull out a signal that was never really there. See it for what it is, catch it early, and the fix is simple, move the sensor a couple of centimeters, or add a second sensor. Days of work, not a year-long AI project. Both look identical on the surface, the cost is wildly different, and the only thing that decides which one you pay is whether you correctly name the problem.

You cannot hire a gap

Why does nobody catch it? The decision lives in the gap between the layers. The hardware and design team owns where the sensor sits. The signal and AI team owns the software. And the decision, that where you place the sensor decides what software you need, sits in the empty space between them, belonging to neither. Each team made a call completely correct for their spot, the design team made it comfortable, the signal team built the best software for the signal they were handed. Both did good work, and the product still got decided wrong, because the decision was standing in the hallway between two offices where nobody was standing. They cannot even see it. The AI team cannot see that placement made the signal weak, so from where they sit the signal is just hard, and reaching for AI is the logical move. The design team cannot see that their comfort call shrank the software options to almost nothing, so from where they sit it was a comfort call. To see the problem, you have to hold the physics of the signal and the constraints of the physical design in your head at the same moment.

Which brings me to the one line to remember: you cannot hire a gap. You can hire depth, the best hardware person, the best AI person, and you should. But depth does not fill gaps, because a gap is not inside anybody's area, it is the empty space between areas, and stacking more expertise inside each area does nothing for the space between them. The decision that decides your whole timeline lives in that gap, so somebody has to stand in it on purpose. That person does not need to be smarter than your specialists, that is not possible, each spent decades in their arena. They need enough understanding of each to stand in the hallway between the offices and ask the questions that fall through the cracks. It is a different job than any deep expert has, and if nobody has it, the decision gets made by accident, and the accident is usually wrong.

The fix: four steps, in month one

The best part is you can do this before you lock the design.

1. Figure out where your signal is strongest on the body. You can usually look it up or measure it quickly.
2. Figure out where your device actually puts the sensor.
3. Estimate how far apart those two spots are.
4. Make the decision on purpose, out loud, with both your hardware and signal people in the same room, which road you are taking. The sweet spot between where the signal is strongest and where it is comfortable decides which path you build your company on.

Step three is the whole thing, so do it properly

The third step is the one that decides the company, and "estimate how far apart those two spots are" deserves more than a line.

You are not trying to produce a precise number. You are trying to produce an honest one, early, with a stated uncertainty, and a week of measurement beats a year of assuming.

Measure at the intended location and at the strongest location, on the same person, in the same session, with the same front end. Comparing a measurement from your device against a number from a paper is comparing two different setups and will mislead you in whichever direction is most convenient. One instrument, one body, two positions, back to back.

Measure on more than one body. Anatomy varies, and the amount of signal lost between two positions varies with it. A gap measured on the one colleague who happens to be available is a sample of one, and the variation between people is frequently larger than the effect you are trying to characterize.

Measure under motion, not at rest. The device will be worn by someone moving, and the signal you can extract while still is not the signal you will have in the field. If the placement is only viable when the wearer holds still, you have not measured the product.

Record what you lose, not only how much. A signal that is attenuated uniformly is a different engineering problem from one whose informative component has disappeared into the noise while the overall amplitude looks acceptable. The first you can amplify. The second you cannot, and the difference decides whether more gain helps or whether you need a model.

Then write the number down where the hardware decision is being made, with a date and the conditions it was taken under. The most common failure here is not measuring badly, it is measuring correctly in month two and having nobody remember it in month nine.

The rule: if you are only losing a little signal, follow it, skip the AI, build the simple tracker, ship now with no data, and keep the fancy AI in your back pocket for later. If you are losing a lot, you may need to commit to the model, but make that call with your eyes open, and plan the data collection from month one as the big, long, complicated thing it actually is. And the option most people forget is on the table: push back on where the sensor goes. Moving it a little, even at some cost to comfort, may save you an entire year-long AI project and an astronomical training bill. The only reason nobody moves it is that by the time they realize they should, the design is locked and it is too late.

The honest limits

This assumes your signal has a clear pattern you can follow. Sometimes it does not, and then the model is the only possible and correct tool, so go build it. And on the treatment side of these devices, there is solid evidence, shown over years, that nudging the body with the right small signal at the right moment can boost a natural rhythm. The bigger claim, that boosting that rhythm makes people think or remember better, does not hold up cleanly, some solid studies boosted the rhythm and it made no real difference. If you build something like this, claim the proven physical effect and be up front that the bigger benefit is still open. That honesty is exactly what separates a claim that survives a smart investor's questions from one that falls apart the second they push.

If you forget everything else: use the simple physics before you reach for AI, watch the gaps between your teams, and make the placement call early and on purpose. Run the check yourself and most of the time you will catch it. Where it gets genuinely hard is the judgment at the edges, when the design is already locked, or the signal sits right on the line between the simple version and the model, with real money riding on getting it right. That is the part I do for people. If that is where you are stuck, you know where to find me.

FAQ

How do I measure how much signal I lose from sensor placement? Measure at the intended location and at the strongest location on the same person, in the same session, with the same front end, so you are not comparing your device against a number from a paper. Repeat across several bodies, measure under motion rather than at rest, and record what you lose rather than only how much, since uniform attenuation and a lost informative component are different problems.

Why does sensor placement matter so much on a wearable? Because the signal is strongest at one anatomical spot and weakens as it travels to the skin, and whatever the first circuit loses cannot be recovered in software. Where you place the sensor sets how much signal arrives, which determines whether you can track it with simple math or need a trained model.

Do I need an AI model for my wearable's signal? It depends on how much signal you lose between the strongest spot and where the sensor sits. If the signal is strong, a deterministic tracker works today with no data. If it is weak and buried, you may need a model, which means a data-collection project, training, and drift, so make that decision on purpose, not by accident.

Why does my wearable project keep slipping? Usually because a comfort-driven placement decision made in month 2 quietly turned a solvable signal problem into a weak signal that now seems to need an AI model. The model brings a data-collection project nobody planned, which stretches an 8-month launch toward 18.

Can software fix a weak biosignal? Only up to a point. If the signal was captured well, software can do a lot. If the first stage lost it in the noise, no model can restore detail that was never captured, the same way editing cannot un-blur a photo. Fix the capture first, at the sensor.