Science1 publisher2 min readPublished
Consumer wearables hand health trials a step count shaped by a black-box algorithm
Wuyoh Sui's group swapped research-grade thigh sensors for step counts pulled from participants' own phone apps, and he told Nature that the substitution handed his lab algorithms and data handling it cannot inspect.
The Scientist · Science desk
What happened
- Nature has published an interview with Wuyoh Sui, a digital-health and behavioural-change researcher at Western University in Canada, on the ethics of building health studies on consumer smart watches and rings.
- The research-focused inclinometers his group used were expensive and required each participant to shave their thigh and attach the sensor under a waterproof cover.
- His group began collecting step counts from the commercial health apps participants already used, feeding them into Pathverse, a platform built to deploy custom health interventions.
- Sui says the group relied on those apps, and on the accelerometers in the commercial VR headsets it used in separate work, to determine the frequency and intensity of movement, without any independent way to assess how.
- He now ranks data privacy and security above researchers' ability to validate non-transparent algorithms as the more pressing concern.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint When the exposure variable in an activity trial is a vendor's processed output, a null result and a measurement artefact look alike. Reanalysis cannot separate them once the data are in.
- exposure Enrolling a volunteer in a study routed through a commercial app can place their daily habits and physical status in a database that advertisers, insurers and data sellers can reach.
- decision Labs choosing an activity instrument now weigh a research-grade sensor whose preparation deters volunteers against a free consumer feed whose processing they cannot inspect.
Sui gave Nature a concrete version of the problem. If a device reports the wrong step count, a platform built on that feed can hand the participant a flawed recommendation, such as advice to double their daily steps [12]. The number is doing two jobs at once: it is the measurement, and it is the input to the intervention the study delivers [9]. So the vendor's algorithm shapes both the estimate and the treatment being tested. Sui describes these algorithms as black boxes, hard to trace from input to output [4].
The data-handling concern is structural. Smart watches and smartphones added a third-party developer whose commercial interests sit outside the researcher's control [17]. "As scientists, we are sharing the data with the developer, whether we like it or not, yet we have no control over where those data end up or how they are used," he told Nature [15]. "Any entity that collects the data is in a position to broker them," he said [16]. The metrics in question describe a person's daily habits and physical status, and Sui named sale, sharing and theft as the routes by which they reach advertisers, insurers and companies that sell personal information [6].
One item on that list is a methods problem as much as an ethics one. Sui said a device maker could push specific articles or adverts to a participant and potentially influence their behaviour, and that such actions could compromise the research process [18].
The interview does not include an accuracy figure for any device [19]. His group could not check [11], and in my view, for anyone running a trial, that is worse than a bias of known size, which can be corrected in the analysis. Once an exposure variable has come out of a vendor's algorithm and been collected, it stays as it is. Contracts and storage rules can reduce the data-handling risk. They do nothing for the step count. Sui and colleagues put the question in print in a 2023 paper, four years after the group moved off its research-grade sensors [20]: "are we, as scientists, opening up ourselves or our study participants to risks by using consumer wearables?" [13]
What to watch
- Head-to-head accuracy studies against research-grade sensors would put a size on the validity gap this interview only describes.
- A wearable maker revising its terms of service on user databases would change what a trial participant is agreeing to.
- Any contractual or technical route that lets an investigator audit a vendor's movement classification would reopen the instrument choice.