Data From a Device You Never Touched

Part of Decentralized Trials (DCT) Training: Connected Devices and Wearables.

About 2 minutes. Free, in your language, no account needed.

Provenance starts before the data does

A wearable produces numbers continuously, from a person you cannot see, through a vendor's pipeline, into a dataset.

You did not place the device, observe the measurement, or handle the transmission. So the controls have to sit elsewhere: in setup, in monitoring, and in knowing what the device actually measures.

Know what it measures and what it infers

This distinction matters more than any other. A device may measure heart rate directly and infer sleep stages from movement and heart rate variability.

A measured value and an inferred one are not the same kind of evidence, and an algorithm change by the vendor can shift every inferred value without any change in the participant.

Setup is where most errors are made

Wrong wrist, wrong fit, wrong participant account, wrong time zone, paired to a family member's phone.

Walk through setup with the participant and confirm the first readings look plausible before the visit ends. An error caught at setup costs minutes; the same error caught at analysis costs the whole participant's device data.

Time zones are the quiet killer

A device set to the wrong zone produces data that is internally consistent and wrong against every other record. Dosing times, visit windows and event timings all stop lining up.

Check it at setup and again after any travel.

Gaps are expected, and must be visible

Batteries die. People shower. Phones lose connection. A perfect dataset from a wearable is more suspicious than one with gaps.

What matters is that gaps are recorded as gaps rather than filled or ignored, and that long gaps are investigated while the participant can still remember the period.

Watch for the device that stopped

Nobody is notified by default. A device that stopped transmitting three days before a key assessment is a problem only if somebody looks.

Build the check into your routine rather than waiting for a vendor report.

Never edit device data

If a value is wrong, it is flagged and explained, not corrected. The device record is source data.