
Wearable technology is expanding beyond step counts, but a new sensor or health score is not automatically a medical diagnosis. The useful question is what a specific device can measure, under which conditions, and what evidence supports the interpretation.
This guide separates documented consumer features from medical uses and research ambitions. It also gives you a comparison worksheet for watches, rings and connected eyewear without assuming every product has the same accuracy, battery life or availability.
Three kinds of claims to distinguish
| Claim type | What it means | Evidence to request |
|---|---|---|
| Wellness feature | A score or estimate intended to support everyday habits | Manufacturer’s intended use, limitations and validation method |
| Medical feature | A defined function intended for a specific medical use | Exact model, software version, authorized use and local availability |
| Research or prototype | An approach being investigated | Study design, participant population and whether a consumer product implements it |
These categories should not be blurred in a comparison table. A watch with one authorized feature does not make every metric from that watch clinically validated. A successful laboratory prototype does not establish the performance of a similarly advertised product.
Trend 1: More health information, with specific boundaries
Heart rhythm features
Some wearables include an electrical heart sensor and an ECG app. Apple documents its supported watch ECG as a recording similar to a single-lead ECG, with classifications that include sinus rhythm, atrial fibrillation and inconclusive results. It is not equivalent to a standard 12-lead clinical ECG.
Apple also states that the app cannot detect a heart attack and that a normal classification does not establish overall health. The eligible device, user age, software and feature availability matter. Read the exact Apple ECG documentation rather than assuming all watches provide the same function. Symptoms require appropriate medical attention regardless of a wearable result.
Glucose claims need particular care
A watch showing glucose data from a separate sensor is different from a watch claiming to measure glucose itself without piercing the skin. The FDA’s published safety communication warns against the latter smartwatch and ring claims and explains the risk of incorrect treatment decisions from inaccurate readings. This is a named regulator’s safety finding, not a worldwide approval list.
Check the exact device and its intended use through current official documentation and the relevant regulator where you live. If medical care depends on glucose measurements, discuss an appropriate validated device with your clinician; do not substitute an unsupported wrist or ring estimate. FDA safety communication on smartwatch and ring glucose claims.
Trend 2: Smaller devices and longer-term habit tracking
Rings and other compact wearables can make ongoing tracking less intrusive, but comfort, fit, charging and data interpretation still need evaluation. A score shown every morning should prompt a question about how it is calculated, rather than being treated as an instruction about your health.
Oura explicitly describes its ring as a non-medical product that is not intended to diagnose or treat medical conditions. Its documentation also identifies conditions that can affect certain readings. That illustrates why the product’s own limitations belong beside any accuracy claim. Oura’s medical-conditions guidance.
For everyday use, compare whether the device fits your routine. Can you wear it comfortably overnight? How often must it charge with your chosen features active? Can you export the information you need? These are practical questions you can answer without assuming that an attractive dashboard is clinically authoritative.
Trend 3: Connected eyewear and hands-free interfaces
Eyewear is another area to evaluate by function rather than by category name. Camera capture, audio playback, displays, navigation and voice assistance are different capabilities. Do not assume that a product called smart glasses contains a visual display or works independently of a phone.
Use a simple task to compare devices: reading an instruction while your hands are occupied, for example. Check display readability, required phone connection, input method, battery behavior and accessibility. When cameras or microphones are involved, consider other people’s privacy and the setting in which you plan to use them.
Trend 4: Algorithms that interpret more data
Separate a measured signal from an algorithm’s conclusion. A sensor reading, a derived score and a claim that an illness can be predicted are three different levels of evidence. A product’s marketing may move between them without making the distinction clear.
For a predictive claim, ask what outcome was predicted, what counted as a correct result, how false alarms were measured, and whether the study included people similar to the intended users. An algorithm tested in one study should not be presented as a guarantee that every wearer receives an early warning.
Also check where processing occurs. “On-device” should identify the particular function that runs locally. It does not by itself answer whether synchronization, support logs or another feature sends information to a cloud service. Review permissions and sharing using our personal information privacy guide.
A wearable comparison worksheet
- Purpose: define the task or habit you want help with.
- Exact product: record the model, app version and feature name.
- Evidence: save the intended-use document and any relevant validation study.
- Boundaries: note eligible users, unsupported conditions and availability.
- Daily use: check fit, charging, accessibility and phone dependence.
- Data: inspect access permissions, sharing, export and deletion options.
- Costs: include required subscriptions or replaceable components.
This framework is an editorial comparison method, not a product test. We do not assign a universal battery range or diagnostic accuracy to an entire device category. Those claims require measurements for the exact product and use case.
Glossary
- ECG: a recording of the heart’s electrical activity.
- Intended use: the purpose and conditions for which a feature is documented.
- Validation: assessment against a stated reference under defined conditions.
- False alarm: a warning produced when the target condition is absent.
- Prototype: an experimental implementation that may not represent a supported consumer product.
Sources and editorial review
Updated 1 October 2026. This guide cites manufacturer limitations and a regulator’s safety communication. It replaces unsupported category-wide clinical labels and illness-prediction guarantees with specific, checkable claims. No independent hardware testing or clinical review is claimed. Send corrections through our contact page.
Written and prepared by Kshitij Gupta.
The Engineering Layer Beneath the Wrist: How Wearables Actually Measure You
Every number a smartwatch, smart ring, continuous biosensor patch, or AR headset displays — heart rate, SpO₂, HRV score, step count — traces back to one of three physical phenomena: photons scattering through capillary tissue, microvolt-scale electrical potentials from cardiac depolarization, or milli-g acceleration frames from a MEMS die. This section documents that signal chain end-to-end: the optical physics, the embedded compute architecture running under a sub-milliwatt budget, the regulatory line separating cleared medical features from wellness estimates, and the encryption posture protecting biometric data in transit and at rest.
Photoplethysmography: The Optical Physics of the Pulse Wave
PPG works by illuminating skin and measuring how much light returns to a photodiode after absorption by blood volume changes. During systole, arterial expansion increases local hemoglobin volume and absorbs more light; the resulting pulsatile (AC) component is typically only 1–3% of the large DC baseline, which is why artifact rejection dominates wearable PPG design. Absorption follows the modified Beer–Lambert relationship A = ε(λ)·c·d + G, where d is an effective pathlength inflated by a differential pathlength factor and G captures geometry-dependent scattering. Because consumer devices cannot control pathlength, they rely on empirical calibration curves rather than first-principles inversion.
Wavelength selection is deliberate. Green light at approximately 530 nm sits near a hemoglobin absorption peak, producing maximal pulsatile contrast from superficial dermal capillaries 0.5–2 mm deep — ideal for heart-rate tracking during motion, where green channels outperform red/IR in pulse SNR. SpO₂ estimation instead exploits the spectral divergence between oxyhemoglobin and deoxyhemoglobin: deoxyhemoglobin absorbs strongly around 660 nm (red), while oxyhemoglobin absorbs relatively more at 940 nm (infrared). The ratio-of-ratios metric R = (AC₆₆₀/DC₆₆₀)/(AC₉₄₀/DC₉₄₀) cancels common-mode intensity variation, and firmware maps R to saturation through a fitted curve of the approximate form SpO₂ ≈ 110 − 25R. Multi-wavelength arrays also enable interleaved dark frames for ambient-light cancellation and venous-versus-arterial component separation.
- Motion artifacts: wrist flexion and tendon shifts modulate the optical path far more than the 1% pulse signal itself. Mitigations include accelerometer-referenced adaptive filtering (NLMS variants), skin-contact force feedback loops that raise LED drive when coupling degrades, mechanical sensor suspension, and multi-photodiode layouts that steer away from corrupted channels.
- Skin tone variance: melanin attenuates shorter wavelengths disproportionately, so green-channel SNR degrades on deeply pigmented skin. Published validation studies have documented PPG-based oximeters overestimating SpO₂ in darker skin tones — the occult hypoxemia problem — pushing vendors toward higher LED current, longer integration windows, and per-tone calibration. This remains an active engineering frontier, not a solved one.
- Peripheral vasoconstriction: cold exposure, sympathetic activation, and vasoactive substances collapse AC amplitude at the extremities first. A chest or upper-arm site retains perfusion where a wrist loses it, which is why cold-weather athletes still see strap-based monitors win accuracy comparisons.
| Optical channel | Primary target | Penetration depth | Strengths / weaknesses |
|---|---|---|---|
| Green ~530 nm | Superficial capillary pulse; heart rate, HRV proxy | ~0.5–2 mm (dermis) | Best pulsatile SNR under motion; most sensitive to melanin attenuation and vasoconstriction |
| Red ~660 nm | SpO₂ ratio numerator (deoxyhemoglobin-sensitive) | ~2–4 mm | Strong spectral discrimination vs HbO₂; higher DC load, more ambient sensitivity than green |
| Infrared ~940 nm | SpO₂ ratio denominator (oxyhemoglobin-weighted); perfusion index | Deeper dermal/subdermal | Least skin-tone bias; weaker pulsatile contrast alone, must be paired with red channel |
Single-Lead ECG, HRV and Electrodermal Activity: The Electrical Channels
Where PPG infers mechanical pulse volume, single-lead ECG measures the electrical depolarization of the heart directly through dry electrodes — typically a watch crown plus back plate, or ring electrodes across two fingers. Signals span microvolts to low millivolts, sampled at 250–512 Hz with 0.5–40 Hz bandpass filtering. R-peak localization accurate to a few milliseconds enables RR-interval statistics far cleaner than any PPG-derived equivalent, which is precisely why AFib screening features combine irregularity indexes (RR-interval variance, RMSSD/SDNN relationships) with morphology features such as absent or unstable P-wave templates.
Heart-rate variability is reported as RMSSD in milliseconds: RMSSD = √( mean over N−1 pairs of (RRᵢ₊₁ − RRᵢ)² ). Healthy adult resting values broadly span roughly 20–100 ms and follow a log-normal distribution, so trend engines operate on ln(RMSSD) rather than the raw value. Two caveats matter for interpretation. First, PPG-based HRV inherits pulse-transit-time jitter and is sensitive to ectopic beats; cross-vendor disagreement can exceed tens of milliseconds on identical nights. Second, HRV is context-dominated — sleep stage, alcohol, illness, and training load all move it more than device error does. Single-session readings are noise-dominated; weekly medians carry signal.
Electrodermal activity (EDA) tracks sympathetic arousal through sweat-gland conductance measured in microsiemens: a slow tonic level plus phasic skin conductance responses concentrated in the 0.05–1.5 Hz band. Wrist-worn EDA is robust to motion by nature of its slow dynamics but confounded by temperature and humidity, and it indexes arousal intensity — not which emotion produced it.
Sensor Fusion and TinyML on a Sub-Milliwatt Budget
A 3-axis MEMS accelerometer sampling at 25–100 Hz supplies the motion reference frame that every other sensor leans on. When a gyroscope is present, an extended Kalman filter fuses them: the prediction step propagates orientation through the gyro rate model, and the update step corrects drift against the accelerometer’s gravity vector whenever linear acceleration is low, with gain K = P·Cᵀ(CPCᵀ + R)⁻¹ balancing trust between sensors. The same fused output drives step cadence, fall-detection state machines, and the artifact-rejection gating applied to PPG before any heart-rate number is computed.
The compute stack is equally constrained. INT8-quantized CNN or DS-CNN kernels run on ARM Cortex-M4/M33-class cores via CMSIS-DSP or TensorFlow Lite Micro inside flash footprints under 100 kB; a single arrhythmia-screening inference over a 30-second PPG window costs on the order of hundreds of microjoules. The always-on power budget works because nothing runs continuously: LED bursts are duty-cycled (for example, a 20 ms multi-wavelength burst each second), radios batch their transmissions, and the MCU sleeps in deep-sleep states between events. Whole-system average draw lands below 1 mW, which on a 200–300 mAh cell yields the multi-day battery life consumers expect.
# Build an always-on fusion firmware for a Cortex-M4 target (Zephyr toolchain)
west build -b nrf52840dk/nrf52840 app -- \
-DEXTRA_CONF_FILE=ppg_fusion.conf \
-DCONFIG_SENSOR=y \
-DCONFIG_LIS2DH_TRIGGER_OWN_THREAD=y
west flash
# ppg_fusion.conf sets: accelerometer ODR 50 Hz, Kalman update period,
# TFLite-Micro INT8 kernel pool, BLE batching interval 60 sPower claims should be verified empirically, not read off datasheets. A USB-based current profiler captures the true duty-cycle waveform:
# Profile average current across PPG sampling bursts vs sleep baseline
ppk2-cli --port /dev/ttyACM0 --vdd 3.0 --samples 1000000 --log ppg_burst.csv
# Separate active-window and sleep averages from the log
awk -F, 'NR>10 {s+=$2; n++} END {printf "avg %.1f uA\n", s/n}' ppg_burst.csv| Subsystem | Typical duty cycle (HR mode) | Contribution to system average |
|---|---|---|
| Green/red/IR LED bursts + photodiode front end | ~20 ms burst every 1 s (2% duty) | Dominant contributor; scales with LED drive current and skin tone compensation |
| Cortex-M4/M33 active compute (filtering + TinyML inference) | Bursts of 5–50 ms per sampling window | Second-largest; INT8 kernels cut energy per inference roughly 3–4× vs float32 |
| BLE radio (batched telemetry) | Short connection events every 30–60 s | Small when batched; grows sharply with streaming raw waveforms |
| Deep-sleep baseline (RTC, RAM retention) | ~98% of wall-clock time | Sets the floor; typically tens of microwatts |
The Regulatory Boundary: Cleared Medical Features vs Wellness Estimates
Regulatory clearance is a claim about intended use, not about general accuracy. A feature cleared through the FDA 510(k) pathway has demonstrated substantial equivalence to predicate devices for a specific indication; CE marking under Class IIa of the EU MDR places it in an analogous risk class. On major platforms, AFib detection and sleep apnea notifications have crossed this line — but as notification or screening features carrying explicit intended-use exclusions (for example, age floors and exclusion of users with known arrhythmias or pre-existing diagnoses), not diagnostic tests. By contrast, VO₂ max estimation — derived from heart-rate-plus-pace first-order models — and composite “readiness” scores remain wellness estimates with no clinical endpoint behind them; published comparisons against laboratory cardiopulmonary exercise testing show typical errors around ±10–15%.



