A Measurement Is Not Meaning: Validation in Human Technology

Every sensor detects something. That does not mean it detects what its manufacturer says it detects.

The output may be voltage, impedance, reflected light, pressure, acceleration, temperature, conductance, or another physical signal. The instrument can record that signal with extraordinary precision while the biological interpretation attached to it remains wrong.

This is one of the most important distinctions in human technology: measurement is an observation; meaning is a validated relationship.

Begin with the biological question

A valid measurement process begins by defining the phenomenon of interest before selecting the instrument. If the intended construct is recovery, vascular function, sleep, stress, hormonal change, or health itself, the physiology must first be described clearly enough to establish what would count as evidence.

Only then can an investigator ask which signal changes with that physiology, at what location, over what time scale, and under which conditions. Starting with an available sensor and searching afterward for a compelling biological story reverses the scientific sequence. It encourages the device to define the question.

The construct also determines the receptor location. A signal that is detectable at the chest, finger, wrist, skin surface, deep tissue, or interstitial compartment is not automatically equivalent across those sites. Convenience of placement does not establish physiological relevance.

A clean signal can still be the wrong signal

Engineers rightly work to reduce motion artifact, electrical noise, drift, interference, and missing data. These are necessary tasks, but a clean waveform is not proof of construct validity. Removing noise improves the measurement of whatever the sensor is detecting. It does not prove that the detected signal represents the proposed biological process.

The question is not whether the signal is strong. The question is whether it changes accurately with the physiology being measured.

This is where attractive demonstrations often outrun evidence. A graph moves when the person moves. A temperature changes during the night. An optical waveform correlates with another output. A model produces a stable score. None of these observations, by themselves, establish the claimed meaning.

If the relationship fails during controlled physiological change, the signal must be discarded for that purpose. It should not be rescued with an inference supplied by the algorithm above it.

The validation chain

Validation is not a single comparison or a high correlation obtained under ideal conditions. It is a sequence of linked obligations. Each must hold before the next has meaning.

A device may perform well at one layer and fail at another. It may repeat the same number without measuring the intended construct. It may track a laboratory reference in healthy volunteers but fail during illness. It may estimate a variable adequately at the group level while being unreliable for an individual. It may predict an outcome yet provide no useful path for action.

These distinctions should not be hidden inside a global accuracy percentage. They should be reported directly.

Algorithms cannot manufacture validity

Machine learning can identify patterns that a human observer cannot see. It can combine weak signals, improve noise rejection, and generate useful predictions. But it cannot transform an undefined construct into a defined one. Nor can it correct a receptor placed in the wrong biological compartment or a signal that changes for several unrelated reasons.

An algorithm trained on labels will learn the relationship between its inputs and those labels. Its performance therefore inherits the validity of both. If users enter the biological event themselves, the person may be functioning as the true sensor while the device learns to recognize adjacent patterns. If the reference measurement is weak, the model can reproduce that weakness with impressive consistency.

Prediction and explanation must also remain distinct. A system may predict an event without measuring its cause. Calling the prediction a direct measurement overstates what the instrument has established.

Physiology first, product last

In developing the Deep Health® Device, our sequence was deliberate. We spent 25 years conducting physiology studies and clinical trials. Only after establishing the biological relationships did we map them to signals and sensors. We then validated those measurements before creating the product.

This order is slower than beginning with hardware and a marketing claim. It is also the only order that protects the meaning of the output. The body is not required to conform to the instrument. The instrument must prove that it represents the body.

Validation also requires the willingness to lose a favored signal. When a measurement does not change faithfully with the intended physiology, it is not improved by a more imaginative interpretation. It is removed. Scientific discipline is visible not only in what a device retains, but in what its developers are prepared to discard.

What readers should ask

Every human technology should make its chain of meaning inspectable. What exactly was sensed? Where was it sensed? Against which independent physiological reference was it tested? How did performance change across people, conditions, and time? Which variables interfered? What failed? Which conclusion is directly measured, and which is inferred?

These questions do not oppose innovation. They distinguish useful innovation from technological theater. The greater the claimed consequence—for diagnosis, treatment, health, safety, or human decision-making—the greater the obligation to answer them.

A sensor output is data. A validated relationship gives that data scientific meaning. Until that relationship is demonstrated, the most accurate description of the device is simple: it measures a signal.