03 · What You Need to Know
How Wearable Algorithms Transform Signals and Influence Research Conclusions
What Does a Wearable Sensor Actually Detect?
A sensor detects a physical quantity through a measurement process. An accelerometer detects acceleration, an optical sensor detects changes in light associated with tissue characteristics, and an electrocardiographic sensor records electrical potential differences associated with cardiac activity.
These signals are not necessarily the final variables that researchers wish to analyze.
For example, a wrist-worn accelerometer does not directly observe whether someone is walking, exercising, or sleeping. It records acceleration patterns that may be consistent with those activities.
Software subsequently processes the signals to estimate events or states.
Researchers should distinguish the directly detected quantity from the derived measurement and the broader construct represented by that measurement.
Distinguish Raw Signals, Algorithmic Estimates, and Research Constructs
Raw sensor signal
The physical measurement acquired by the sensing system, subject to its hardware characteristics and acquisition procedures.
Algorithmic estimate
A value or classification produced by applying processing rules or a statistical model to the sensor data.
Research construct
The phenomenon the researcher intends to investigate, such as physical activity, sleep quality, fatigue, or psychological stress.
For example, a wearable may acquire optical signals, estimate heart rate, and use that estimate alongside other information to calculate a stress-related score.
Even if the heart-rate estimate is sufficiently accurate, the resulting score may not validly represent psychological stress in every circumstance.
Each inferential step requires appropriate justification.
How Do Algorithms Convert Sensor Signals Into Measurements?
Wearable processing pipelines may include signal filtering, artifact detection, segmentation, feature extraction, classification, and aggregation.
Filtering may reduce unwanted signal components. Segmentation divides observations into intervals. Feature extraction identifies characteristics such as movement intensity, periodicity, or variation in cardiac timing.
Algorithms may then apply decision rules or statistical models to classify activities or estimate outcomes.
For example, a step-counting algorithm may identify movement patterns consistent with walking and reject signals interpreted as unrelated arm movements.
The resulting count depends on how those patterns are defined and detected.
Different processing pipelines may therefore produce different estimates from similar sensor signals.
Which Common Wearable Outputs Depend on Algorithms?
| Wearable output |
Underlying information |
Algorithm-dependent interpretation |
| Step count |
Movement signals |
Classification of movement patterns as steps |
| Activity intensity |
Acceleration and possibly other signals |
Estimation of activity categories or intensity |
| Energy expenditure |
Movement, physiological signals, and sometimes personal characteristics |
Model-based estimation of energy use |
| Sleep duration |
Movement, physiological signals, and timing |
Classification of sleep and wake periods |
| Sleep stages |
Multiple indirect physiological and behavioral signals |
Estimation of sleep-stage categories |
| Heart rate |
Optical or electrical signals |
Signal processing and detection of cardiac timing |
| Stress-related score |
Selected physiological and behavioral indicators |
Composite estimation based on a device-specific model |
| Recovery score |
Selected activity, sleep, and physiological indicators |
Manufacturer-specific calculation or classification |
The exact inputs and algorithms vary by device. Researchers should consult relevant technical documentation and validation studies rather than assume that all products calculate similarly named indicators in the same way.
Why Can Two Wearables Produce Different Step Counts?
Two devices may use different sensor placements, sampling rates, filters, detection thresholds, and classification algorithms.
A wrist-worn device may detect arm movements that a hip-worn device does not. Conversely, walking while holding an object or using a mobility aid may reduce the wrist movement available to a step-detection algorithm.
Even when devices are worn at the same location, their algorithms may apply different criteria for identifying a step.
Consequently, two devices may report different counts for the same activity.
The disagreement does not automatically establish which device is correct. Evaluation requires an appropriate reference measure and conditions relevant to the study.
How Do Threshold-Based Algorithms Affect Classification?
Some algorithms classify observations using predefined thresholds.
For example, movement intensity may be categorized according to the magnitude or pattern of acceleration over a specified interval.
Thresholds developed in one population or device configuration may not perform equivalently in another.
A threshold suitable for one age group, movement pattern, or sensor placement may misclassify activity elsewhere.
Researchers should examine the evidence supporting the selected thresholds and determine whether they are appropriate for the intended population.
Thresholds should not be treated as universal physiological boundaries merely because software uses them consistently.
How Do Machine Learning Algorithms Change Sensor Interpretation?
Some wearable algorithms use machine learning models trained to recognize patterns associated with particular activities or physiological states.
For example, a model may learn to distinguish walking, running, and sedentary behavior from labeled sensor recordings.
Its performance depends partly on the training data, labeling procedures, selected features, model architecture, and evaluation methods.
A model trained predominantly on one type of participant or activity may perform less well when applied to substantially different conditions.
Researchers should evaluate external validation and relevant subgroup performance rather than rely exclusively on performance reported during model development.
What Is Algorithmic Measurement Bias?
Algorithmic measurement bias occurs when the measurement process produces systematic errors relative to the intended quantity or reference.
For example, a step-counting algorithm may underestimate movement among participants whose walking patterns differ from those represented adequately in its development data.
Similarly, an optical heart-rate algorithm may perform differently under particular motion, device-fit, or physiological conditions.
Bias may affect the average estimate or differ across participant groups.
Researchers should assess both overall measurement performance and whether errors vary in ways that could distort substantive comparisons.
Why Does Population-Specific Validation Matter?
A wearable algorithm may be validated under controlled conditions involving particular participants and activities.
However, the target research population may differ in age, mobility, health characteristics, activity patterns, or device use.
For example, an activity classifier evaluated primarily during ordinary walking may not perform equivalently during rehabilitation exercises or wheelchair use.
Researchers should consider whether validation conditions adequately represent their intended application.
Where existing evidence is insufficient, additional validation may be necessary before using the output as a primary research measure.
Can an Algorithm Be Reliable but Invalid?
Yes. An algorithm may consistently produce similar estimates under repeated conditions while systematically misrepresenting the intended quantity.
For example, a device might consistently overestimate energy expenditure during a particular activity.
Its measurements may be repeatable but biased relative to an appropriate reference.
Likewise, a composite score may be calculated consistently without sufficient evidence that it represents the psychological or clinical construct assigned to it.
Researchers should distinguish repeatability, agreement, analytical validity, and construct interpretation.
What Does the V3 Framework Contribute?
The V3 framework proposed by Goldsack and colleagues distinguishes verification, analytical validation, and clinical validation for biometric monitoring technologies.
Verification concerns whether the technology captures and processes signals according to its technical specifications.
Analytical validation examines whether the technology accurately and reliably measures the intended quantity under specified conditions.
Clinical validation addresses whether the measurement meaningfully supports its intended clinical interpretation.
These distinctions help researchers avoid assuming that technically functioning sensors necessarily produce scientifically valid outcomes.
For nonclinical research, the broader principle remains relevant: technical performance and construct validity are different questions.
Why Is Correlation Insufficient for Algorithm Validation?
Two measurements can correlate strongly while differing systematically.
For example, a wearable may consistently report values higher than an appropriate reference while preserving the relative ordering of participants.
The correlation could be high even though the absolute measurements are unsuitable for the intended use.
Researchers should examine measurement differences, systematic bias, error distributions, and agreement limits where appropriate.
The acceptable magnitude of error depends on the research question and the consequences of inaccurate measurement.
What Happens When Algorithms Use Proprietary Methods?
Some manufacturers do not disclose the complete algorithms used to generate wearable outputs.
This may limit researchers' ability to reproduce calculations, inspect model assumptions, or determine precisely why an estimate changes.
However, proprietary algorithms are not automatically unsuitable for research.
A device may still provide useful evidence when independent validation demonstrates adequate performance for the intended application.
Researchers should distinguish unavailable algorithm details from unavailable validity evidence.
When critical outputs cannot be adequately evaluated, reliance on the proprietary score may be difficult to justify.
Should Researchers Prefer Raw Sensor Data?
Access to raw or minimally processed signals may permit researchers to apply documented algorithms, investigate artifacts, and reprocess observations consistently.
For example, raw acceleration data may support comparison of different activity-classification methods.
However, raw data require suitable processing expertise, computational resources, and quality assurance.
They may also contain sensitive behavioral or physiological information requiring appropriate protection.
Raw-data availability is valuable when it improves transparency or analytical flexibility, but it does not automatically establish that the resulting measurements are valid.
Can Different Algorithms Be Used on the Same Raw Data?
Yes, when the available data and processing requirements permit it.
For example, researchers may apply alternative activity-classification algorithms to the same acceleration recordings.
Differences between the resulting estimates can reveal how strongly conclusions depend on processing choices.
However, agreement among algorithms does not independently establish accuracy. Several methods may share similar assumptions or errors.
Where possible, comparisons should include an appropriate reference measure or other independent validity evidence.
How Can Algorithm Changes Affect Longitudinal Studies?
Software or firmware updates may change the algorithms used to calculate wearable outputs.
For example, an updated activity classifier may identify more movement events than an earlier version.
A longitudinal dataset could then show an apparent increase in activity even if the underlying behavior remained stable.
Researchers should document relevant device models, software versions, and processing methods.
When algorithms change during data collection, comparability should be evaluated before interpreting trends.
Can Researchers Combine Data From Different Wearable Brands?
Potentially, but combining similarly named outputs requires evidence that the measurements are sufficiently comparable for the intended analysis.
For example, two devices may report sleep duration while using different rules for identifying sleep onset and awakening.
Direct pooling may introduce measurement heterogeneity unrelated to participants' actual sleep.
Researchers may consider device-specific analyses, appropriate calibration, harmonization, or sensitivity analyses when justified.
However, statistical adjustment cannot automatically correct fundamentally incompatible measurement definitions.
How Should Algorithm-Dependent Variables Be Reported?
Researchers should identify the device model, relevant hardware and software versions, the specific output analyzed, and available information about the processing algorithm.
Reports should distinguish directly measured signals from algorithmically derived estimates.
Relevant validation evidence, population limitations, and changes in processing should be described.
When raw signals are reprocessed, researchers should document the algorithm version, parameters, filtering, and transformation procedures sufficiently for reproducibility.
Where proprietary processing prevents complete reproduction, that limitation should be disclosed.
When Should an Algorithm-Generated Variable Be Rejected?
An algorithm-generated variable may be unsuitable when its relationship to the intended construct is unclear, validation evidence is inadequate, or systematic errors could materially undermine the study's conclusions.
For example, a proprietary stress score may be inappropriate as the sole measure of psychological stress when its construction and validity for that purpose cannot be established.
The same device may nevertheless provide sufficiently validated heart-rate measurements for a narrower physiological question.
Researchers should consider whether a more directly interpretable indicator or alternative instrument would better support the intended inference.
Watch Out
Do not assume that an algorithm-generated variable is a direct measurement simply because its name resembles the research construct. A wearable's sleep, stress, recovery, or activity score may reflect manufacturer-specific definitions and processing assumptions that require independent validation.
