A wearable does not observe recovery, stress, sleep stage, or calorie expenditure directly. It records physical signals, applies quality checks, derives features, and then maps those features to an estimate or interpretation. Each layer can add uncertainty.

Accuracy is therefore not one property of a device. The answer depends on the sensor, outcome, activity, population, reference method, firmware, wear conditions, and analysis goal. A heart-rate result cannot validate energy expenditure, and agreement during rest cannot automatically generalize to motion.

At a glance

What to keep in view

  • Ask which physical signal was recorded before evaluating the headline metric.
  • Make signal-quality and missing-data rules visible before feature calculation.
  • Separate direct records, derived features, model estimates, and interpretation.
  • Validate the specific outcome, population, activity, and software version you plan to use.

Layer 1: the sensor and sampling context

Optical, electrical, inertial, temperature, and location sensors observe different phenomena. Sampling schedule, contact, placement, motion, skin or environmental conditions, and device power choices can alter the available waveform or record. The downstream algorithm cannot recover information that was never captured reliably.

Begin an evaluation with a data dictionary: sensor or source, unit, timestamp semantics, sampling behavior, device and software version, and known gaps. If only an app-level aggregate is available, say so rather than describing an unseen raw signal pipeline as fact.

Layers 2 and 3: signal quality and features

Signal-quality logic may reject motion-corrupted or physiologically implausible segments. That can improve the retained data while creating non-random missingness: the hardest activities or wear conditions may be underrepresented. Report both the result and how much data survived quality control.

Features convert accepted records into quantities such as counts, interval summaries, variability measures, intensity zones, or time-window statistics. Window length, filtering, artifact correction, aggregation, and timezone boundaries are part of the definition—not neutral implementation details.

  • Publish inclusion, exclusion, and minimum-coverage rules.
  • Version feature definitions when firmware or processing changes.
  • Do not replace a rejected segment with an apparently measured value.

Layer 4: personal baseline or population model

A personal baseline compares the current feature with a person's sufficiently covered history. It can reduce inappropriate one-size-fits-all comparisons, but the baseline still depends on window length, data quality, routine changes, and update rules. It is context, not a diagnosis.

A population model learns or defines a mapping from inputs to an outcome. Its usefulness depends on the training population, labels, calibration, and external validation. A proprietary score may be convenient, but without a disclosed method the reader should avoid inferring clinical meaning from its precision.

Layer 5: uncertainty and action

Uncertainty belongs in the output. Useful interfaces show coverage, freshness, contributing sources, confidence or limitation language, and the difference between a recorded measurement and a derived estimate. The action should become more conservative as evidence weakens.

Systematic reviews reinforce the need for outcome-specific evaluation. Evidence about HRV measurement, step count, heart rate, or energy expenditure should not be pooled into a claim that a wearable is simply accurate. Check the metric-specific result and the conditions under which it was tested.

  • Recorded: what the source actually stored.
  • Derived: a transparent calculation from available records.
  • Estimated: a model output with validation and applicability limits.
  • Interpreted: decision support that should remain proportional to evidence.

Use boundary

Information, not medical advice

This article explains data and research methods. It does not diagnose a condition, prescribe treatment, establish a universal normal range, or replace qualified professional care. If symptoms or a medical decision concern you, use an appropriate clinical service.