I know you must be thinking that doesn't make sense: two different people are never going to have the same heart rate numbers anyway. Different resting baseline, different fitness levels, different everything. And yes, that part has nothing to do with wearables at all. But it's almost a competition at this point.

The real question here arises when you take the comparison of the same heart rate metric with different wearables but on the same person. The heart rate reading will be different – again because it will depend upon the type of sensors used in each case, the speed of sampling, and the software algorithms of the sensors.

After you learn how they work internally, you'll finally be able to understand why the numbers are different.

Table of Contents

• They're All Measuring the Same Thing, Just Not in the Same Place

• Why Your Heart Rate Isn’t Constant

• Sampling Speed Changes What the Device Actually Catches

• The Algorithm Matters as Much as the Sensor

  • Skin Tone and Wrist Size Add Another Layer

• Activity Type Throws Everything Off Differently for Each Brand

• What This Means If You're Comparing Numbers With a Friend

They're All Measuring the Same Thing, Just Not in the Same Place

All of the wearables worn on your wrist or finger use the same technique known as photoplethysmography (PPG). An LED shines light into your skin while a sensor measures how much light bounces back with each heartbeat. The technology remains the same among all manufacturers, but the difference here lies in the place of measurements.

Finger-based wearables measure arteries located closer to the surface of the skin than the ones on your wrist. This gives finger-based devices like the Oura Ring a real advantage.

This advantage becomes clear even in the sleep stage, since the movements of your fingers are smaller than the movements of your wrist during sleep. Research done to test the accuracy of the heart rate reading by the Oura Ring has been able to show high consistency with the ECG measurements.

This is one reason Oura tends to perform well in nocturnal heart rate verification. A comparative study conducted in 2025 in the journal Physiological Reports pitted 5 devices (the Oura Ring Gen 3, Oura Ring Gen 4, Whoop 4.0, Polar Grit X Pro, and Garmin Fenix 6) against a medical-grade ECG for 536 nights of sleep. The Oura Ring 3 and 4 correlated the most accurately with ECGs, WHOOP correlated moderately, and Polar had the lowest accuracy.

Why Your Heart Rate Isn’t Constant

One thing you need to know is that your heart rate isn’t constant even at rest.

It varies all the time with your breathing movement or any other stress response, even when you're sitting completely still. 68 bpm can become 72 bpm after just a few seconds and nothing out of the ordinary happened.

It's enough for two wearables to have slightly different readings just because they were measuring at different times. Before sensor location, sampling rate, or any other algorithm comes into play, what's being measured is already constantly changing.

Sampling Speed Changes What the Device Actually Catches

Position of the sensor is important, but frequency of readings is much more important than you might think. The PPG sensor from WHOOP 5.0 and WHOOP MG records reading 26 times per second. This is considered to be quite a good rate of recording from a continuous wearable.

Other smart devices, on the contrary, work every few seconds and not continuously. As a result, if your Heart Rate suddenly spikes for any reason, like standing up too quickly or being startled, the device will miss that spike and fill in the gap to replace a missing value.

The Apple Watch also has a different mechanism: it has a high hardware sampling rate, but that doesn't mean that its heart rate sensor works continuously. In reality, it works periodically depending on your activity, and only works continuously while working out and for several minutes after the exercise.

None of this is necessarily better in an objective sense. The continuous heart rate frequency sampling is fantastic for capturing all those short-term spikes, but results in more raw data that needs to be filtered and sorted out by the algorithm.

Garmin takes a middle ground here, depending on the model of the watch. In general, most Garmin devices take continuous samples at low frequencies all day long until they detect the start of your exercise and increase sampling frequency automatically.

And there's another trade-off here: the device saves battery when you're working at your desk, but this also means that the device is essentially betting on its own movement detection being accurate before it decides that your heart rate actually matters more right now.

The Algorithm Matters as Much as the Sensor

But here's the thing: the heart rate monitor only receives a raw signal of light intensity. And then something else must translate this information into a proper value. This translation is made using proprietary software for each brand separately. This is what results in the daily scores and graphs that we're all so obsessed with.

The task of this software is to determine whether the signal represents the heartbeat or some noise created by physical activity, skin contact, or ambient light.

A 2024 study published in Sensors concluded that the heart rate readings for the Oura Ring remained accurate compared to ECG when lower quality readings were included.

The case of heart rate variability was quite different. Heart rate variability measurements became less accurate when the device used lower-quality data, especially for people over the age of 45.

The heart rate algorithm used in WHOOP underwent a huge update in February 2026. The company used the training data collected at its own research facility for a wide range of skin tones, body types, and activity levels. Also, it used cloud-based processing for data analysis after the workout or a sleep session instead of edge processing.

This is a completely different strategy from having the watch do all the calculations on the device itself. Therefore, two devices that have identical hardware will give different figures from the same data depending on the interpretation of the software.

That’s also the reason why your figures might be altered by the brand itself even when it hasn't altered any of its hardware. With the release of the February 2026 update by WHOOP, there were some changes seen in heart rates and recovery scores for all 4.0 and 5.0 users, although none of the users’ activity had changed at all. The sensors remained the same, but the math behind them wasn’t.

The same happened to me with my step count on my Ultrahuman Ring AIR after a firmware update. If you've noticed your own numbers drift after an app update with no real change in your routine, this is usually why.

Skin Tone and Wrist Size Add Another Layer

There are more than just sensors, positioning, and algorithms involved, of course. There's also the human who wears the device, plus PPG sensors do not work the same for all people.

The amount of light absorbed by the skin is dependent upon the amount of melanin in the skin, meaning you'll get different results on different skin tones.

One of the most extensive analyses is offered by Koerber et al. (2023). They showed that this may lower the accuracy of reading in darker skin tones, not just for a particular brand, but for many brands.

PPG works by shining light into the skin and measuring the light reflected back. Since melanin absorbs more light, darker skin tones can reduce the amount of reflected light that reaches the sensor. This can lower the signal quality and, in some situations, reduce measurement accuracy.

The same goes if you have any tattoos on the area where the wearable sensor is placed. Dark or densely pigmented tattoo ink may absorb or scatter the emitted light, making it more difficult for the sensor to obtain a clean signal.

This doesn't necessarily prevent measurements altogether, but it can increase the likelihood of inaccurate or inconsistent readings.

In addition to this, the circumference of your wrist along with the device's tightness and fit will also have an impact on the amount of light that leaks out of the sensor rather than being reflected back into it. A watch that's too loose allows ambient light to enter and reduces the amount of reflected light captured by the sensor, while one that's too tight can alter blood flow beneath the sensor. Smaller wrists may also make it harder for the sensor to maintain consistent contact with the skin.

For these reasons, manufacturers generally recommend wearing the device snugly, but not tightly and positioning it slightly above the wrist bone during measurements.

Activity Type Throws Everything Off Differently for Each Brand

Resting heart rate during sleep is one thing, but your heart rate while working out is a totally different ball game for the sensors. This is where the difference becomes most evident between gadgets.

Motion artifacts are caused by movement, which means that your swinging arm interferes with the ability of the device to measure the reading the sensor is trying to capture. All companies have their algorithms to filter these movements and some handle certain movement patterns better than the others.

That's also the reason why chest strap monitors (that treat electrical signals directly from your heart rather than relying on light through skin) still outperform every wrist or finger-based wearable during high-intensity training.

What This Means If You're Comparing Numbers With a Friend

When you and your friend are wearing different brands of wearables and your numbers don’t line up, that’s perfectly normal and not an indicator that your device is faulty. All that really matters is consistency over time.

The resting heart rate trend on your Oura or Garmin should be giving you an accurate picture of your recovery and stress based on your personal baseline. Consistently improving your workouts and striving towards a better trend in HRV is better than a single measurement.

My honest takeaway here is that heart rate tracking on consumer wearables has definitely improved, especially in terms of resting and steady state numbers. But there are still significant differences between brands and these are very much real. They come from design choices, not manufacturing defects.

The final number relies on three factors: the placement of the sensor, how often readings are taken, and how the softtware processes that information.

But will I stop arguing with my partner on how he stresses me out, which is why my heart rate drops later than his at night? Probably not.

This thing is quietly recording your heart rate and movements while you’re sleeping. So it's a valid question to ask whether the wireless signals or electromagnetic field (EMF) radiation it emits could interfere with sleep quality, disrupt hormones such as melatonin, affect circadian rhythms, or produce other biological effects over time.

The concerns are part of the wider discussion on EMF’s and wireless technology. Although all smartphones, Wi-Fi, and wearables emit RF waves, the amounts of energy used and how they work can be quite different. Understanding what your sleep tracker is actually doing helps put those concerns into context.

Table of Contents

• How Sleep Trackers Measure Sleep

• What's Actually Transmitting From Your Ring or Watch

• How That Stacks Up Against Safety Limits

• Where the Melatonin Research Gets Misapplied

• What the Latest Research Actually Says

  • The Pushback

• What This Means for How You Wear It at Night

• The Bigger Risk Isn't the Radio

How Sleep Trackers Measure Sleep

Before discussing the impact of signals and EMF, we should know what exactly happens when a device monitors your sleep.

Sleep trackers rarely actually measure your sleep but rather use sensors to track movements, heart rate, heart rate variability (HRV), breathing rate, temperature, and more.

The software analyzes data and distinguishes between states like awake, light sleep, deep sleep, and REM sleep. As opposed to professional polysomnography, sleep tracking devices don't measure brain waves and so can't actually observe sleep. They only make an educated guess.

This is critical to understand, because it's not the Bluetooth radio itself that measures your sleep, but rather sensors.

What's Actually Transmitting From Your Ring or Watch

The chip in your Oura, Ultrahuman, or Apple Watch communicates with your phone via Bluetooth Low Energy (BLE). BLE was designed with a power budget rather than a performance budget because, in this case, power is more important than range when you have the receiver just a few inches away, in your pocket or on your nightstand.

The transmit power, according to the Bluetooth specification, is capped at 100 milliwatts. But most chips used by consumers are well below the maximum power limit, often at 1-10 milliwatts.

By contrast, a cell phone during a voice call can transmit up to 250 to 2,000 milliwatts. You're not carrying a miniaturized cell tower, you're carrying an item that transmits in bursts of low-power signals.

How That Stacks Up Against Safety Limits

The SAR metric is used to quantify the amount of RF energy tissue absorbs. This metric is measured in watts per kilogram. The Federal Communications Commission (FCC) caps the SAR level to 1.6 W/kg averaged over 1 g of tissue. But according to International Commission on Non-Ionizing Radiation Protection (ICNIRP) standards, the average SAR should be maintained at 2 W/kg averaged over 10 g of tissue.

No wireless communication device will be certified and made available on the market unless it meets the criteria of the SAR metric.

In the United States, the SAR limit for wrist-worn devices is 4.0 W/kg. According to Apple's Apple Watch RF exposure data, the watch has a reported SAR value of approximately 0.17 W/kg, while Oura reports a SAR value of 0.0003 W/kg for the Oura Ring. Both are well below regulatory limits, illustrating just how little RF energy these wearables typically emit.

According to an engineering evaluation published in 2024 by Kim, Sharif and Nasim, SAR levels associated with commercial wearable technology operated at 2.4 GHz have been found to comply with the regulatory threshold and the safety guideline at the distance of skin contact.

Where the Melatonin Research Gets Misapplied

Melatonin is the natural hormone responsible for regulating your sleep-wake cycle, which is why it has been frequently mentioned in the context of the relationship between EMFs and sleep. Indeed, there have been numerous studies in the past indicating that exposure to specific forms of electromagnetic fields might have an impact on melatonin production, circadian rhythms, or oxidative stress.

But the results from these studies have been quite mixed. Some research has found a notable effect, while others have found no significant effect at all. Most importantly, most studies that are constantly being cited refer to extremely low-frequency (ELF) fields generated by power lines, electrical wiring, and household electricity in general and not the Bluetooth device radiofrequency signals.

That research is real, but it often concerns extremely low-frequency (ELF) fields, those at 50-60 Hz found in power lines and electrical wiring, rather than the 2.4 GHz radiofrequency used by your Bluetooth ring. These are different parts of the electromagnetic spectrum with different interaction mechanisms.

Citing ELF melatonin studies to explain RF wearable exposure is a bit like citing research on UV exposure to explain what your microwave does. Related field, wrong frequency range.

What the Latest Research Actually Says

The strongest evidence we have today comes from a series of systematic reviews commissioned by the World Health Organization.

There were several reviews published between 2024 and 2025 which aimed to assess whether RF-EMF (Radiofrequency Electromagnetic Fields) exposure was connected with outcomes such as sleep disorders, headaches, and nonspecific symptoms.

No evidence suggesting a cause-and-effect relationship between RF-EMF exposure below current safety thresholds and sleep disorders was reported by either experimental or observational research.

This certainly doesn't solve the problem because of the uncertainty of the evidence. And this is quite low due to challenges in estimating actual RF exposures experienced by people. We're all surrounded by various signals emitted by our smartphones, Wi-Fi routers, laptops, cellular towers, and so on.

The conclusion is relatively simple, though: there's currently no evidence that would show that RF exposures using Bluetooth disrupts people’s sleep. That being said, researchers are still actively exploring the issue.

The Pushback

Not everyone agrees with these findings. Some researchers claim that WHO review fails to give enough weight to certain studies and that research findings on the subject are still lacking.

This criticism targets the whole body of work on RF-EMF radiation since most of them are based on mobile phones rather than wearable technology.

The debate is ongoing, but present research doesmn't show any disruptions caused by wearables with Bluetooth functionality.

What This Means for How You Wear It at Night

If the thought of wearing a sleep-tracking wearable next to your body for a full eight hours leaves you feeling uncomfortable, the quickest solution won’t be getting rid of it. Many wearables offer a low-power mode, airplane mode, or similar settings that disable Bluetooth communication while allowing the device to continue collecting data through onboard sensors such as the accelerometer and optical heart rate sensor.

For people concerned about EMF exposure, this setting reduces wireless transmissions during the night while still preserving most sleep-tracking functionality on the device. The reduction in radiofrequency emissions is typically small in absolute terms because Bluetooth Low Energy already transmits at very low power and only intermittently. Still, you may prefer minimizing any unnecessary wireless activity while you sleep.

I’m not saying that disabling Bluetooth improves sleep quality or health outcomes. But if doing so helps you feel more comfortable or less worried about wearing a device overnight and its EMF, it can be a practical compromise that allows you to continue tracking your sleep without the added concern.

The Bigger Risk Isn't the Radio

The bigger sleep-tracking problem probably isn't EMF at all. Neurologists are seeing more patients who walk in fixated on hitting a target number of REM minutes when using a wearable device that measures sleep stages based on movement and heart rate.

Unlike a laboratory polysomnography test, where sleep stages are measured directly from brain activity, the measurement provided by the wearable device is only an inference based on movement and heart rate. The term used for such obsession is 'orthosomnia', which is better described as the downside of wearing a sleep tracker rather than anything related to EMFs.

If you're going to worry about something at 2 a.m., the accuracy of the sleep data is probably a better place to focus than the Bluetooth chip. Sleep trackers estimate sleep stages rather than measuring them directly, and that limitation can sometimes create more anxiety than the radio signals themselves.

For women who are in their reproductive years, chances are your wearable technology has misread your luteal phase symptoms as either a result of being overtrained or even sick. This is because the technology likely detected a symptom that it doesn’t actually understand.

Let’s get into how this is actually happening by going from sensors to algorithms and finally to where the accuracy gap actually lives.

Table of Contents

• What the Menstrual Cycle Actually Does to Your Biometrics

  • Resting Heart Rate and HRV

  • Heart Rate Variability (HRV)

  • Skin Temperature

• How Wearables Measure These Signals

  • PPG Sensors and What They Actually Capture

  • Temperature Sensors: Continuous vs. Spot Measurement

• How the Algorithms Work

  • Calendar-Based vs. Physiology-Based Detection

  • How Machine Learning Classifies Cycle Phases

• Why the Accuracy Gap Exists

• What Cycle-Aware Algorithms Look Like in Practice

• Wrapping Up

What the Menstrual Cycle Actually Does to Your Biometrics

Before jumping into the sensors and algorithms, here's what they're actually detecting. The menstrual cycle isn't the noise within wearable data, but an active component that alters the physiology upon which any recovery or health algorithm relies.

There are three signals that tell the story.

Resting Heart Rate

Multiple studies using continuous wearable monitoring have confirmed that resting heart rate increases 2-7 bpm from the follicular phase to the luteal phase. One prospective study of 91 women observed that resting heart rate was 3.8 bpm higher in the mid-luteal phase compared to the period of menstruation.

Heart Rate Variability (HRV)

On the other hand, HRV changes in the opposite direction. In particular,  a meta-analysis of more than 1,000 participants showed the reduction of vagally mediated HRV from follicular to luteal phases of the menstrual cycle.

For example, one study reported that SDNN decreased from 154 ms in the follicular phase to 136 ms in the luteal phase, which represents a decrease of 12%. Progesterone is responsible for such effects. Specifically, it triggers the renin-angiotensin system (RAS), increases the total blood volume, raises HRj, and reduces parasympathetic influence. On the other hand, estrogen decreases HR (negative chronotropic effect) and leads to greater HRV.

So during the mid-luteal phase, you already have an increased RHR but a reduced HRV. To a recovery algorithm that does not know where you are within your menstrual cycle, this combination signifies stress, sickness or overtraining.

Skin Temperature

The temperature shift has been most thoroughly studied out of the three. Postovulatory rise of basal body temperature by 0.3–0.7°C due to progesterone’s effect has been known for over 100 years and constitutes the basis of traditional fertility awareness methods.

My Oura Ring data also shows that skin temperature usually increases during the luteal phase. It also tends to drop briefly just prior to ovulation due to an abrupt drop in body temperature related to estrogen.

The key point here is that signals change in the same direction at the same time, every cycle, predictably. When an algorithm treats these indicators separately, it's structurally wrong.

How Wearables Measure These Signals

PPG Sensors and What They Actually Capture

Heart rate and HRV measurements from wearables are done by Photoplethysmography (PPG). This sensor emits LED light, generally green for heart rate and red & infrared for SpO2, to shine on your skin. Light gets absorbed differently by blood depending on its volume, so as your heart beats and blood flows in capillaries, light reflected from your skin will be different for each heartbeat. Variation in light reflected is known as the PPG waveform.

Based on PPG waveform data, wearables calculate beat-to-beat intervals. While calculating the heart rate is relatively easy as it simply counts peaks per minute, HRV needs precise timing since it measures the variation in milliseconds between consecutive heartbeats. That’s where signal quality starts to matter a lot.

Placement of sensors on your skin also plays a vital role in this. Generally, finger devices such as smart rings like Oura and Ultrahuman give cleaner PPG signals compared to wrist-worn devices such as your Apple Watch, Garmin, or WHOOP. The finger has higher density capillaries, resulting in larger pulse amplitude and lower motion artifacts.

Wristwear makes up for this problem with more sophisticated signal processing techniques. But there's always a price to pay for that. For instance, Oura Ring 4 provides users with an 18-path multilayered wavelength PPG sensor with adaptive sensor configurations.

Temperature Sensors: Continuous vs. Spot Measurement

Temperature sensors incorporated in current wearables measure skin temperature and not core body temperature. These sensors, called thermistors, are capable of detecting temperature fluctuations in terms of changes in electrical resistance.

While there's a relationship between skin temperature and core body temperature, the two aren't the same. Skin temperature responds to factors such as room temperature, weather conditions, and temperature variation caused by changes in blood flow around the skin surface.

Even so, continuous overnight monitoring of skin temperatures may provide better information compared to traditional basal body temperature (BBT). With the fertility awareness technique, temperature is always measured at the same time each morning, right before getting out of bed. Missing a measurement or a bad night of sleep may negatively impact results.

Wearables take a different approach. By collecting temperature data throughout the night, they can identify longer-term trends and reduce the impact of short-term fluctuations.

Some devices, such as the Apple Watch Series 8 and later, Fitbit Sense, and Oura Ring, have temperature sensors. Most smart rings track temperature changes from an individual’s baseline, not the absolute temperature itself. It makes identifying temperature increases, which happen after ovulation, easier.

How the Algorithms Work

Calendar-Based vs. Physiology-Based Detection

Perhaps the most basic way of detecting the menstrual cycle is through a calendar model. The user inputs the first day of their period, the app calculates the average cycle length, and predicts the fertile window forward from there.

Apps like Clue, Flo, and older versions of Apple’s period tracker use this as their foundation. It’s a simple algorithm that needs no sensor data at all.

The problem with calendar algorithms is accuracy. These types of methods operate on regular cycles, but these aren't common in many women. For ovulation detection, for example, studies reveal that there's an average error of 3.44 days for calendar methods alone.

Also, calendar methods predict menstrual phases based only on dates entered by the user, whereas physiology-based approaches analyze sensor data such as temperature, heart rate and HRV to detect ovulation and cycle-related changes. For example, Oura uses heart rate and temperature to detect ovulation with an average error of 1.26 days.

How Machine Learning Classifies Cycle Phases

Machine learning algorithms don't use a single metric to determine where you are within your menstrual cycle. Rather, they examine patterns in several physiological indicators taken from wearables, such as skin temperature, heart rate, heart rate variability (HRV), and in some cases, electrodermal activity (EDA).

Over time, machine learning algorithms figure out which cycle stages correspond to which physiological patterns. For example:

• The luteal stage is characterized by an increase in skin temperature and changes in cardiovascular metrics.

• Ovulation causes changes in patterns in terms of temperature and heart rate.

• The menstrual phase can show its own distinct combination of physiological changes.

• The follicular phase is generally the most difficult one to recognize since its biometric signatures aren't clearly defined and tend to coincide with those from other phases.

A 2025 study found that machine learning algorithms can effectively determine the menstrual, ovulatory, and luteal phases. The accuracy of the results decreased when the follicular phase was added to the list of phases.

Modern cycle tracking apps have become complex because of this reason and they no longer depend solely on temperature. It becomes easier for a device to identify the phases of the menstrual cycle with every additional physiological signal that it captures.

Other technologies like the Vivoo FlowPad are also emerging that attempt to collect menstrual health data directly rather than inferring it from wearable sensors.

Why the Accuracy Gap Exists

The issue with wearables comes down to the fact that many of the metrics related to menstrual cycle phases aren’t exclusive to the menstrual cycle.

Take, for instance, the metrics such as a high resting heart rate, reduced HRV, and increased skin temperature. These could be observed during the luteal phase, but can also occur thanks to a range of other factors, including illness, lack of sleep, stress, consumption of alcohol, or even jet lag.

Yet another hurdle with menstrual tracking involves individual differences since some women might have significant changes during their menstrual cycles when it comes to temperature and HRV, whereas others will have minimal changes in those metrics.

This is why most menstrual tracking algorithms require individual baselines instead of population baselines, meaning that the more data is collected from a woman regarding her menstrual cycles, the better it gets at identifying her personal patterns.

What Cycle-Aware Algorithms Look Like in Practice

Until 2025, most wearables considered tracking cycles and recovery as two separate concepts. Oura became the first big company to connect the two.

Its updated algorithm accounts for increased resting heart rate, decreased HRV, and increased body temperature, all common during the luteal phase. Instead of automatically lowering readiness scores, it checks whether those changes are a normal part of the menstrual cycle.

This reduced the number of falsely low recovery scores during the second half of the menstrual cycle. In 2026, Oura went further with a dedicated AI model focused on cycles, fertility, pregnancy, and menopause.

WHOOP chose a different route through its metric called cardiovascular amplitude that measures heart rate and HRV variability throughout the whole cycle. Rather than focusing on individual phases, it looks at the overall physiological impact of hormonal changes.

Natural Cycles became the first fertility app that obtained FDA approval for contraceptive use, collecting users' body temperature data with the help of their wearables’ sensors like the Apple Watch, Oura Ring, Garmin, or its own dedicated NC Band.

Garmin, Fitbit, and Samsung track menstrual cycles, but those insights remain largely separate from their recovery and readiness metrics.

Wrapping Up

This boils down to the mismatch between measurements taken by wearables and what recovery algorithms were designed to handle.

PPG sensors and temperature sensors allow wearables to detect changes that happen across the menstrual cycle and they work well enough. Multi-parameter machine learning allows for reliable classification of the cycle phases, particularly those happening during ovulation.

But problems arise because many recovery algorithms have been trained on data biased towards male samples, where hormonal cycle variations are considered to be noise. These recovery algorithms lack the means to differentiate between luteal phase physiology and initial phases of an illness. Sensors won’t solve this problem, but algorithmic design will.

From the perspective of developing health apps using wearable device APIs, we already have access to health metrics that incorporate information about the current stage of the cycle. Oura provides it in specific endpoints, Apple integrates with HealthKit’s HKCategoryTypeIdentifier, and WHOOP ties it into its recovery model.

The problem here is that data can be accessed on these platforms via different APIs, data models, and integration techniques. While Oura, Apple HealthKit, and WHOOP may expose similar health metrics, there can still be differences in the sampling frequency, preprocessing methods, and metric definitions, making it hard to create algorithms that would work consistently across platforms.

This lack of standardization also contributes to the training data problem. Data collected by Oura, Apple Watch, and WHOOP can't always be combined easily since each platform stores and works with data differently. As a result, researchers and developers have to do additional work preparing and normalizing data before it can be used to train models.

There are sensors and the models have been improving, but the APIs are fragmented and the lack of training data is real. That’s where the work is.

But then I put another smartwatch on my wrist in an attempt to see how different those results were. Both watches were on the same wrist, at the same time, counting the same walk. One said 8,400 steps, while the other said 6,900.

Same wrist, 1,500 steps apart.

So naturally, I had a small crisis about everything I thought I knew.

The strange thing is that nobody really tells you this when you buy a fitness tracker. The packaging doesn’t read “lab accurate, not life accurate” or anything remotely close to this idea. The app never mentions the fact that two people who own wearables from the same company might actually have their steps counted differently.

But when you start looking into how wearables calculate your steps, things make more sense than you think.

Table of Contents

• Inside the MEMS Accelerometer

• How a Step Becomes a Number

• Why Wrist-Based Tracking Is Hard

• Why Slow Walking Confuses Wearables

• False Steps Are Real

• Why Lab Accuracy Doesn’t Match Real Life

• Do Some Brands Perform Better?

• How the Person Wearing the Device Affects Accuracy and What You Can Actually Do to Improve It

• Final Thoughts

Inside the MEMS Accelerometer

Each modern fitness tracker and smartwatch includes a MEMS accelerometer (Micro-Electro-Mechanical System). The MEMS consists of a tiny silicon chip with microscopic moving parts inside.

Body movements cause these microscopic components to move by an extremely small margin, which is captured by the sensor as a change in the electric signal. Most wearable devices have 3-axis or triaxial accelerometers. This means that they measure motions in three directions at once:

• up/down

• left/right

• forward/backward

These signals are captured continuously at about 50 times per second.

When you walk, your body produces a recognizable motion pattern, such as hip movement downward, movement of torso up and down and arm swinging rhythmically. What's most important is that your body bounces up and down with each step taken.

This vertical bounce is considered one of the clearest indications when someone walks and that is why step counters are so dependent upon it.

The accelerometer sends out three streams of information regarding movement. Many algorithms combine them into a single magnitude signal using the Euclidean norm:

‖a‖ = √(x² + y² + z²)

This gives the device a rotation-independent way to measure total acceleration.

The Role of Gyroscopes

Higher-end wearables also have a gyroscope that detects any rotations. The accelerometer and gyroscope together make up the Inertial Measurement Unit (IMU). This makes sure that the device can differentiate between you walking or just moving your wrist around.

How a Step Becomes a Number

The sensor itself can't calculate anything. All it does is produce raw motion data. The real work happens when that signal is interpreted through an algorithm, which is where things begin to differ dramatically.

All companies have their own proprietary algorithms. The algorithm used by Garmin is not the same algorithm used by Apple. And Apple’s is not the same as Samsung’s. Some of the most common approaches include:

• Peak detection: The algorithm detects repeating peaks in the acceleration signal that correspond to walking activity in humans. Since most people walk at a step rate in the range of 100 to 130 steps per minute, the device is designed to detect periodic activity in this range.

• Zero-crossing detection: Instead of looking for peaks, this approach involves determining how often the signal crosses a midpoint value.

• Autocorrelation: This method involves searching for repeating patterns in the motion signal over time.

• Frequency-domain analysis: Some algorithms search for the dominant frequency of walking.

• Machine learning models: Modern wearable devices have adopted machine learning approaches for walking pattern recognition based on training on large datasets of labeled walking behavior. These systems can differentiate between walking, running, typing, driving and random arm motion most of the time.

But every algorithm carries its own trade-offs. If the sensitivity of the algorithm is set too high, then the smartwatch may overlook slow or subtle motions. On the other hand, if the sensitivity is set too low, the watch would generate false steps when washing dishes or riding on bumpy roads.

There's no perfect setting, but every company has chosen to adopt its preferred compromise.

Why Wrist-Based Tracking Is Hard

The primary problem with modern wearable devices is their position on the body.

The waist-mounted pedometer is located close to the center of mass, which allows for the detection of very clear and strong signals. This is why traditional clip-on pedometers often perform better than you might expect.

On the contrary, smartwatches operate with signals detected from arm movements, and your wrist is noisy. The natural walk accompanied by a natural swinging of the arms provides a very strong correlation between leg movements and arm movements, but any real-life scenario may disturb this reading. If you’re pushing a stroller, holding shopping bags, a mobile phone, or walking with hands deep in your pockets, the correlation weakens and it becomes impossible to detect a specific signal related to leg movement.

Beyond this, even the side of the body where you wear the smartwatch plays a role in data interpretation. For example, the dominant arm tends to produce a stronger acceleration signal, but it also produces additional signals due to its active work in performing different routine actions.

This is one of the reasons why you would’ve noticed that most wearable devices require specifying which wrist you are strapping the device to during setup.

Why Slow Walking Confuses Wearables

One of the most surprising things about this research is that it’s far easier to accurately detect regular walking than it is to detect slow walking.

Wearables do well enough in detecting walking speed at normal paces. But as your speed drops, the acceleration and rhythm of the motion signals start to become smaller. This makes it difficult for algorithms trained primarily on healthy adults walking at normal treadmill speeds to identify slow walking correctly. The same studies may also show that wrist-worn trackers underestimate the number of steps in slow walking.

But why does this matter? The older we get, the slower and more restricted our gait becomes, which means that the motion signals detected by our wearable devices will be weaker. Then there are patients with neurological conditions like Parkinson’s disease or stroke, which can produce gait patterns that consumer algorithms simply weren’t trained to recognize.

So the next time you think your watch is broken, its likely that the detection system is simply working from assumptions that don’t match your movement pattern.

False Steps Are Real

Wearables not only overlook steps, but they can actually create non-existent ones. Because accelerometers measure all kinds of acceleration, quite a lot of actions completely unrelated to walking may trick the algorithm.

Activities that can cause phantom steps include:

• Driving on bumpy roads

• Typing aggressively on the keyboard

• Washing dishes

• Folding the laundry

• Drumming

• Taking a train or bus

There's an engineering dilemma here: the more sensitive your algorithm is to slow walking, the more prone to errors. The less sensitive it is, the better it works, but undercounts certain activities.

This is the reason for very different results from different brands despite having the same data entered.

Why Lab Accuracy Doesn’t Match Real Life

Fitness trackers tend to be surprisingly accurate within laboratory testing. In tests under controlled conditions, such as on treadmills, a number of wearables manage to reach step counting errors below 10%. But the real world tends to be messier.

Researchers call this “free-living” data because it involves movements made outside of the laboratory in normal life situations. The real-world walking pattern is characterized by irregularities such as rough surfaces, stops, carrying things, changes in speed, unpredictable arm movements and walking intervals. All these factors contribute to making step detection more difficult.

A smartwatch may prove to be exceptionally accurate in a controlled environment, yet still behave inconsistently in daily use.

Do Some Brands Perform Better?

Yes, but it’s rarely that significant. Studies indicate that Apple Watch, Garmin, Fitbit, and Samsung are all decent pedometers during normal walking conditions. Garmins are especially valued for their consistency and reliable tracking when outdoors. The Apple Watch reportedly works very well for altered gait and slower walking. Fitbits use a more sensitive algorithm and that may result in increased step counts.

But there are other elements at play. The speed at which you walk, natural movement of your arms, location of the device on the body, what activity you’re doing, and how the algorithm interprets your movement. The difference between two people using the same watch is often larger than the difference between two brands.

How the Person Wearing the Device Affects Accuracy and What You Can Actually Do to Improve It

A few things can noticeably improve step-count accuracy:

• Walk at a natural pace when possible: Wearables work best at moderate walking speeds. Walking slowly, shuffling and stop-and-go actions are significantly harder to detect with algorithms.

• Wear the device properly: Try putting on your wearable device snugly above your wrist bone to ensure consistent movement along with your arm.

• Set your dominant wrist correctly: Most smartwatches will adjust sensitivity settings based on which hand you use as the dominant one.

• Keep your arms moving naturally: Holding bags, pushing a stroller, or keeping your hands in your pockets might affect the accuracy because most wearable sensors depend greatly on the motion of your arms.

• Keep firmware updated: Most manufacturers continue to refine their algorithms through firmware updates.

• Use trends, not exact numbers: Step counts are better at tracking your habits and long-term patterns on the same device, but are less accurate compared to the number of steps registered by other brands.

• Be cautious with slow or altered gait: Older adults, rehabilitation patients, or those with brain injuries are likely to see undercounting of steps because most algorithms are trained on standard walking patterns.

Final Thoughts

The technology in today’s wearables is truly amazing. A small sensor the size of a grain of rice measures your movement multiple times each second, filters out noise and tries to make sense of the data generated by it. That’s an incredibly difficult engineering problem. But step counting is ultimately an estimation problem.

There are different step counts generated by various brands because they all have different filtering, motion classification, signal interpretation and sensor placement. None of them count your actual steps perfectly. They estimate the probabilities based on your movement patterns. And in reality, human movements aren't always neat and monitored.

So if you have two watches that generate different step counts, don’t panic. They do their best to understand messy movement data.

This technology combines smart sensors, connectivity, and automated emergency response features that detect falls in real time and help reduce the time a person may remain unattended after.

Let's look at what the WHO data says. Every year, 684,000 people die from falls. On top of this, there are 37.3 million cases in which a person ends up sustaining injuries from a fall that require medical attention. And there's also a significant number of people lie on the ground unattended for hours after falling, as is often the case for people living alone.

That’s the gap these wearable IoT systems are trying to reduce. In this article, we'll talk about how fall detection actually works, from the sensors inside the wearable to the alert sent to a caregiver or family member.

Table of Contents

• What You Should Know Before Reading

• What is a Fall Detection System?

• The Hardware: Sensors That Do the Heavy Lifting

• The Algorithm: From Raw Signals to Classification

  • Threshold-Based Detection

  • Classical Machine Learning

  • Deep Learning

  • The Lab-vs-Real-World Problem

• The IoT Pipeline: How the Alert Gets to You

  • Edge Processing First

  • The Confirmation Window

  • Alert Dispatch

• Consumer Devices: Where Things Stand

• The Challenges Worth Knowing About

• Wrapping Up

What You Should Know Before Reading

Before reading this article, it helps to have a basic understanding of:

• IoT (Internet of Things) devices and how they connect to the internet

• Wearable tech such as smartwatches or fitness bands

• Sensors like accelerometers and gyroscopes

• Basic wireless communication methods such as Bluetooth or GPS

• Mobile apps and notification systems

You don't need advanced technical or programming knowledge to follow the article.

What is a Fall Detection System?

A fall detection system automatically detects a fall and raises the alert to a caregiver, family member, or emergency services, without any action required by the user.

This aspect of “without doing anything” is crucial because once the person has fallen, they may not be in a state of mind to perform any action, even as little as pressing a button.

Modern fall detection systems leverage the advances made in the areas of IoT, machine learning, cloud computing, and communications technologies to build a pipeline for the identification of a fall and raising an alert in less than 30 seconds.

The Hardware: Sensors That Do the Heavy Lifting

The core of all wearable fall detectors is the Inertial Measurement Unit (IMU), which comprises two sensors:

• Accelerometer: detects acceleration on three axes (X, Y, Z). It measures the sudden deceleration when impact occurs. Most commercial devices have an accelerometer with a ±120 m/s² range. Popular in both prototypes and production-level units is the MPU-6050 chip.

• Gyroscope: measures angular velocity (±1200°/s). It detects how your body rotates to determine whether it is a forward, backward, or any other kind of fall. Jumping and sitting rapidly will not be mistaken for a fall in such a case.

Together, these give you a 6-axis picture of continuously measured motion, which is the input that the detection algorithm works with.

More advanced devices are equipped with heart rate sensors, GPS and barometers. Heart rate change upon falling could be useful in determining that an incident occurred. The GPS will enable locating the user easily, while the barometer is able to indicate what floor the person is on.

As for sensor location, it matters a lot more than people expect. The most accurate data is collected by waist-mounted sensors, as this area sees maximum movement during falls. But commercial products usually go for wrist-based ones because people won't care about accuracy if they don't even wear it.

The Algorithm: From Raw Signals to Classification

Raw sensor data is just a stream of numbers. It's the algorithm that answers the one main question: did this person just fall?

It sounds simple but it really isn’t. The physical dynamics involved in making vigorous actions like running, jumping, and sitting down may be transiently similar to those of a fall. An error would either miss the true occurrence or produce too many false alarms, both having real consequences.

Different wearable fall detection systems use different types of algorithms for analyzing the sensor data and determining whether a fall has taken place. These include approaches such as rule-based methods, machine learning and even more advanced deep learning algorithms. This would depend on various factors such as device complexity, processing power, accuracy requirements and battery constraints.

Threshold-Based Detection

This is the oldest approach. You establish certain criteria: if the acceleration is above X g-force, and the person remains motionless for Y seconds afterward, you classify this as a fall. This approach employs a finite state machine (FSM): the sensor moves through certain states (from upright to free fall, then impact and post-fall motionlessness), and identifies a fall whenever all these states happen in sequence.

It's lightweight and power-efficient, but has higher false-positive rates, especially for active users or those with tremors.

Classical Machine Learning

Support Vector Machines (SVMs), Random Forests, and k-Nearest Neighbors (k-NN) are widely used for wearable fall detection systems trained on labeled sensor data containing falls and normal Activities of Daily Living (ADLs). These models identify motion patterns captured through accelerometer and gyroscope to distinguish falls from false alarms.

In a study, SVM achieved 87% accuracy, k-NN reached 84%, and Random Forest achieved 89%.

Deep Learning

The CNN-LSTM hybrid model is capable of capturing both spatial and temporal features directly from the accelerometer and gyroscope sensor data without the need for manual feature engineering. The attention model is another alternative that is able to focus selectively on certain sensors according to the motion pattern.

The Lab-vs-Real-World Problem

The bitter reality is that the lab accuracy numbers don't transfer one-to-one into real-world situations.

During experiments, the falls are taken by young healthy individuals who are well aware of what's happening. But in reality, falls often occur most in older people with co-morbid conditions, in cluttered surroundings, and unexpectedly.

During experiments, the falls are induced on young healthy individuals who are well aware of what is happening. But in reality, falls occur among older people with co-morbid conditions, in cluttered surroundings and unexpectedly. 

In one such study conducted on 19 individuals for more than 400 hours, only 10 actual falls were recorded and the performance of the fall detection system dropped significantly compared to controlled laboratory results.

Addressing this gap between laboratory and real-world performance should be a critical goal for future wearable fall detection systems.

The IoT Pipeline: How the Alert Gets to You

Detecting a fall is step one. Here’s what the rest of the pipeline looks like.

Edge Processing First

The detection algorithm runs directly on the microcontroller of the wearable device, not remotely in the cloud. This is key, because the latter approach (in the cloud) is known as “edge computing”, and its importance lies in two things: efficiency (as additional round-trips to the cloud slow down everything), and reliability (fall incidents occur in basements, elevators, and in places without a signal).

Microcontrollers like the ESP32 can execute classifiers and FSM algorithms locally using minimal power.

The Confirmation Window

When the device thinks a fall has occurred, it usually starts a 30 to 60 second countdown instead of immediately calling emergency services. During that time, the wearable vibrates and plays an alert sound so the user can cancel it in case it was a false alarm.

This is very important because otherwise any sudden arm movement, even false alarms, would cause an emergency call.

Alert Dispatch

Once confirmed, the notification, together with GPS coordinates, is sent by:

• NB-IoT:  the ideal protocol for wearables outdoors. It uses an already established LTE network, very minimal battery and excellent wall penetration.

• BLE: links to the user’s phone to forward the alert.

• Wi-Fi: fast and dependable within a home environment with good connectivity.

• LoRaWAN: long-distance and very low battery energy. it's suitable for a large campus or rural setting.

The alert hits a cloud server, from where it's relayed to SMS notifications, push notifications, emergency call-backs to authorities, or a caregiver platform, depending on system configurations.

Consumer Devices: Where Things Stand

Apple watches series 4 and later feature arguably the most advanced fall detection mechanism among consumer wearables. This technology makes use of data collected from an accelerometer and a gyroscope with machine learning algorithms that help detect a fall and automatically call for help if there is no sign of movement from the person wearing it.

Apple regularly works on improving its fall detection system by eliminating any errors.

The Samsung Galaxy watch (Active2+) comes with hard fall detection that utilizes motion sensors for detecting motion based on the accelerometer. Users can adjust when the watch detects falls, such as always, while exercising, or while being active.

Within 60 seconds after detecting a fall, the watch shows an alarm prompt, after which it sends an SOS message to the emergency contact if no response is provided. This feature is off by default and needs to be turned on manually via the Galaxy Wearable app.

Dedicated medical pendants like Philips Lifeline and Medical Guardian continue to be favored in many high-risk elder care situations as they have specialized hardware, round-the-clock monitoring facilities, cellular connectivity features, and GPS technology. Though not as unobtrusive as smart watches, medical alerts are engineered especially for emergency situations.

The Challenges Worth Knowing About

False positives are the main challenge regarding user experience. When alerts fire too often, users turn off the whole mechanism, defeating its purpose.

In order to solve this issue, Apple uses multi-level criteria to trigger an alert: only the combination of impact on the wrist, heart rate change, proper orientation of the body, and post-impact behavior will be able to confirm the occurrence of a fall.

Privacy is also a legitimate concern, especially for camera-based systems. We can address these concerns through emerging approaches like thermal sensors (body heat without identifiable imagery) and Wi-Fi Channel State Information (CSI) analysis, which offers detection without surveillance.

Training for fall detection is done directly using federated learning on the wearable device itself, so there's no need for user motion data to ever be sent out to the cloud. The cloud will only receive updates.

Finally, consistency is easy to ignore, but it matters a lot. A device that isn’t worn detects nothing. That is why smart pressure sensors placed into floors, room monitors based on radar technology, and Wi-Fi CSI analysis techniques are becoming more popular as supplements to wrist wearables.

Wrapping Up

Fall detection sounds like a solved problem at first. All the sensors are ready, the algorithms are functioning effectively in the labs, and there’s plenty of connectivity. But what matters most after a fall is how quickly someone gets help.

With improvements made in wearable AI technology, fall detection is moving gradually from being an emergency-based solution to becoming a constant monitoring tool which silently protects its users. Technology-wise, we're pretty much there. The important thing is to make sure everything works reliably.